69% of executives believe AI agents will reshape their business this year. The agentic AI market will be worth $60B / year by 2031. You already know this and have several AI initiatives at your company.

What should you do, and in what order? More importantly, what should you personally focus on as a leader?

Your fellow CEOs are posting about their weekend vibe-coding prowess on LinkedIn. Leading from the front is great for signaling that AI matters, but it won’t move a 2,000-person organization. Resist the urge to feel like a young solo builder again until you have fixed how your company operates. It’s about delivering an agentic operating model.

Think about where corporate inefficiency lives. It’s not in any single person’s work. It’s in the handovers, the duplicated efforts, the back-and-forths between teams. Getting everyone a chatbot subscription barely scratches the surface. The companies pulling ahead are the ones where employees work more autonomously — using AI to handle tasks that previously required delegating to, or waiting on, someone else. That’s how AI-native startups achieve $1.5M to $6M in revenue per employee, compared to $500K at Fortune 500 companies.

The next step is to fix how work flows through the organization.

Here’s what your CEO roadmap could look like.

Communicate the ambition and urgency

Your employees must act like builders.

The message should be clear: the company is becoming AI-native. That means delivering more value to customers with fewer handovers, flattening hierarchy, and giving individual contributors the tools to do work that previously required a team. Every employee’s job will get more interesting by delegating tedious tasks to agents. Change is urgent: every week, we must see progress.

Two KPIs matter most: % weekly adoption as an input metric, and revenue per employee and speed to market (however your industry measures it) as output metrics. These are the numbers that reflect structural change, not just “initiative X is live”.

78% of C-Suite executives say that getting real value from agentic AI requires a new operating model. If that’s where your industry is headed, the CEO needs to be the one who says it first.

Follow the announcement with early symbolic actions that make AI experimentation easy and visible:

• Basic AI subscriptions for everyone: Google Gemini or Microsoft Copilot.

• Premium subscriptions for anyone who wants them: OpenAI, Anthropic, or Perplexity. Don’t make people justify the upgrade with a business case.

• Specialized tools for specialized teams: design, marketing, and engineering.

• Showcases and knowledge-sharing sessions — in dedicated Slack/Teams channels and at a fortnightly all-hands 45 min panel discussions.

• A hackathon: give people a day to go wild with agents in a sandbox. They can get their hands dirty without risking the business.

These moves are cheap relative to their signaling power. They also flush out your power users (the 20% who will become the engine of your transformation). Today, only 28% of employees know how to use their company’s AI tools. That number needs to move fast.

Build a dedicated AI team

The AI landscape moves so fast that keeping track of what’s possible, let alone what’s practical, is nearly a full-time job. You need a dedicated AI task force, led by the CEO. An operational team with real authority to prioritize, build, and ship.

MIT Sloan Review nails the problem: in most organizations, technology executives focus on infrastructure, and strategic executives focus on markets. Agentic AI doesn’t fit neatly into either box. Only the CEO can bridge that gap.

Team members can retain their departmental reporting lines while dedicating at least 50% of their time to the task force. Think of them as SWAT teams — they parachute into departments, help identify the best use cases, set up the tools, and get projects off the ground.

What kind of people do you need? Not AI researchers. You need people who are good at applying AI, which is a very different skill set:

1. At least 50% dedicated.

2. Tech-savvy enough to read and write Python code and have a working knowledge of APIs, OAuth, and cloud environments. You can’t automate enterprise workflows without understanding how systems talk to each other.

3. Deep enough in the business to spot the high-value workflows.

4. Resourceful and don’t take no for an answer.

5. Strong writers — because communicating clearly with AI is the #1 skill that separates power users from everyone else.

Prioritize and document use cases

Your next deliverable is a list of priority initiatives.

42% of companies abandoned most of their AI initiatives in 2025 — up from 17% the year before — largely because they spread themselves too thin.

Pick workflows that meet three criteria:

• Labor-intensive: They eat up significant employee hours.

• Frequent: They happen several times a week.

• Documentable: You can write a memo to teach an intern.

A $200M/year company can realistically pursue 3 meaningful use cases per quarter. Three done right will generate more value than fifteen done just because every department must show progress.

Prioritization must be decided top-down. Don’t crowdsource ideas through employee surveys. As Bain Capital Ventures found after studying a year of enterprise AI deployments, top-down direction and AI literacy matter far more than having the right tools.

Immediately after prioritization, you must assign team members to document the current process and the ideal end state. Workflow mapping is the most time-consuming part of agentic transformation, so don’t wait to get started.

Deloitte’s Tech Trends 2026 shows that many agentic AI implementations fail not because the technology doesn’t work, but because companies try to automate their existing processes as-is. As a CEO, you must be clear in your expectation that priority workflows will need to be reimagined.

Choose your agentic stack

Now we are getting to the tooling question. The market is crowded and moving fast.

You have to decide between N8N, Google Gemini Studio, Microsoft Power Apps, sidebar agents from SaaS vendors (Notion, Salesforce, Netsuite, ServiceNow), Claude Cowork, Microsoft Cowork, Perplexity Computer, Manus, OpenClaw, Genspark, OpenAI Frontier, and more arriving every month.

First, you need to decide whether to stick with the AI offerings from your existing vendors (e.g., Microsoft, Salesforce, Google) or open up to AI-native agents (e.g., OpenAI, Anthropic, Perplexity). Gartner projects that by the end of 2026, 40% of enterprise apps will embed task-specific AI agents.

To be honest, this decision trips up more companies than it should. The AI offerings of traditional software vendors are not sufficiently useful today. Perhaps they’ll be in 12 months. If you want to move now, you need to build on AI-native agents on top of your systems and databases.

Four questions will help you narrow the field:

What user experience do you need?

• Chat: Quick, low-cost, conversational. A user asks questions and gets answers. Think Google Gemini Flash.

• Flow: Deterministic automation that runs with minimal human intervention. Think invoice processing or data reconciliation. Think N8N.

• Session: A human orchestrating multiple agents through interactive conversations. This is where knowledge work lives — strategy, analysis, content creation, anything requiring judgment and taste. Session-based agents like Claude Cowork are actually coding agents with a nicer interface.

How much customization do you want?

In general, the biggest problem that you’ll encounter with any automation platform is when its functionalities support almost everything that you need (say 80%), but not quite. The remaining 20% is a huge headache, requiring endless workarounds. Assume that you’ll need more customization than you think.

Where does the data live?

Claude Cowork runs on each employee’s desktop. Perplexity Computer runs entirely in Perplexity’s cloud. Some workflows involve sensitive data that must stay on your infrastructure; others don’t. Sorting this out early prevents painful migrations later.

What’s the budget?

Is this $20/employee/month or $200? Is $1 acceptable for each customer interaction, or should it be $0.05? The economics vary dramatically across platforms and use cases.

Answering these questions will narrow down the list of options.

For many types of knowledge work, Claude Cowork is a great and versatile solution for the time being.

But, whatever you choose, design for portability. You don’t want to be locked in. By packaging your company’s automations as skills — reusable text-based instructions that tell an agent how to perform a specific task — you can easily move them from one platform to another. Think of skills as standard operating procedures written for your AI workforce, far easier to maintain than the proprietary visual flows of Microsoft Power Automate or N8N.

Make sure your SaaS apps play nice

To be clear: I’m not suggesting you rip out your CRM, ERP, or HRIS. These systems of record are brutally hard to build, despite claims to the contrary on social media, and migrating off them will drain your team’s energy. Even OpenAI and Anthropic are using Slack, Salesforce, and NetSuite, after all.

But your existing SaaS apps need to be AI-friendly. What does that mean in practice?

• They have extensive, stable, and well-documented APIs.

• They support a healthy ecosystem of third-party integrations.

• They don’t make you wait weeks for API key approvals or new integration permissions.

Your vendors will become obstacles to AI adoption if they have impenetrable documentation, restrictive API policies, and glacial approval processes.

If your AI agents can’t read from and write to a system, that system is a bottleneck, and shutting it down becomes a priority.

Connect the pipes

Credentials, permissions, APIs, CLIs, OAuth, and MCP servers… They enable AI agents to access your company’s data.

Agents need permissions, which fall into 2 main patterns:

• Personal agents inherit the user’s existing permissions. If a sales rep can see their accounts in the CRM, their agent can too.

• Workflow agents receive dedicated API keys scoped to the exact data they need to complete a particular task. Think of a lead generation workflow. Nothing more, nothing less.

Your IT department will push back. New app installations must be approved. New API keys mean new attack surfaces. New service accounts mean new security risks. These concerns are legitimate. Yet, at some point, they will need to be over-ridden. Integration requests for agentic projects must be endorsed at the top. Only the CEO can change the organization's risk posture.

Launch before you’re ready

If you wait until governance, measurement, and scale-up plans are fully baked, you’ll always be a step behind. Most AI use cases will be significantly reworked after launch anyway, based on user feedback, evolving technology, or just the reality of how the workflow actually plays out in production.

Start with basic logging. Use AI coding assistants to spin up a quick dashboard and see what your agents are doing.

It takes a while to figure out the right governance and the right observability and evaluation frameworks for your company’s use cases: probably at least a year of trial and error. You can start working on them early, but don’t make them a prerequisite.

Avoid the usual traps

The same mistakes keep showing up:

Prioritizing by opinion, not data. Someone on the exec team gets excited about a use case, and it jumps to the top of the list without any analysis of how much time or money it would actually save. 56% of CEOs report no significant financial benefit from AI despite doubling their investment. In most cases, they invested in the wrong things.

Staying in experimentation mode too long. McKinsey reports that nearly two-thirds of organizations haven’t begun scaling AI. So, if you are still running pilots, you are not the only one. But it’s time to focus on adoption: what are the use cases that are going to be used by 50% of your employees or your customers?

Hoarding the good tools. Your company is already paying for Google Workspace or Microsoft 365, so authorizing another $200/month for Claude or Perplexity feels redundant. It’s not. The bundled AI features in your existing stack are a starting point. Power users need power tools.

Governing before you’ve proven value. Risk management and compliance frameworks are essential, but they should follow value. If you over-invest in governance before confirming that a use case actually works and matters, you’ll have a beautifully documented process for something nobody uses.

Wrap-up

95% of enterprise AI projects fail. That’s fine. Fail fast, then double down on the 5% that could become 90% of your product or your business.

Meanwhile, stay focused on your priorities as a leader:

• Communicate the ambition

• Build a dedicated AI team

• Prioritize and document use cases

• Choose your agentic stack

• Make sure your SaaS apps play nice

• Connect the pipes

• Launch before you’re ready

Above all, you must instill a sense of urgency.

Using AI, solo founders like Peter Steinberger can make dozens or hundreds of daily GitHub contributions and sell their project to OpenAI for a reportedly large sum.

In Feb 2026, Jack Dorsey’s company, Block, announced a 40% reduction in headcount, attributable to future productivity gains unlocked by AI.

If you are a CEO or CFO already paying for a bunch of Google Gemini or Microsoft Copilot subscriptions, you are probably wondering why you are not seeing these eye-popping returns on investment. Did you select the wrong tools? Are people not using them properly?

What are you going to tell your Board?

Here are some insights from discussions with companies across a wide range of industries.

20% productivity gain is the bare minimum target

The business case for AI adoption is to increase speed and reduce unit costs. 20% productivity improvement is a bare minimum 1-year target for most AI initiatives, calculated on the eligible cost base.

Recent announcements help to illustrate this point. They should be taken with a grain of salt because companies only report their successes, and they calculate their numbers based on narrow cost bases to look good. But your Board will have these benchmarks in mind when they ask you to present the company’s “AI strategy”, so here we go.

Speed increase:

• Klarna: customer resolution time cut by 80%.

• Airbnb: 18-month code migration completed in 6 weeks.

• Duolingo: 4x more lesson content produced with the same headcount.

• Goldman Sachs: 80% less time to generate a first draft of IPO prospectus.

Cost reduction:

• Blackstone: 50% lower real-estate portfolio valuation costs.

• IBM: 40% lower HR costs thanks to the AskHR system.

• Salesforce: customer support workforce reduced from 9,000 to 5,000. AI handles ~50% of customer interactions.

• Lemonade: 68% reduction in insurance claims processing costs.

Improvements are usually delivered workflow by workflow, rather than across the board. Hence, it’s very important to prioritize and sequence AI initiatives top-down so that your teams tackle the frequent, repeatable, labor-intensive workflows first.

How to measure the improvement? If you already have hiring budgets in place for the coming year, you can freeze hiring while you equip your teams with the proper tools. Avoiding layoffs, if possible, will make AI adoption smoother.

Not everyone steps up equally

When you start equipping your team with AI tools, you’ll notice that each person adjusts in their own way.

For example, 20% of your employees will be power users, working 2x faster or more. Perhaps 60% will be casual users, with a 25% higher work output. And at least 20% will be resistant to change.

Your job is to encourage power users, freeing them up from business-as-usual activities so that they can manage the transformation, and asking them to lead others by example. On the other hand, reluctant team members must be made aware that AI adoption is a basic job expectation and that their performance review will reflect that.

When Tobi Lutke (Shopify) tweets about rediscovering the joys of coding and delivering new software in a single weekend, he is setting expectations for how his direct reports should behave.

Put managers to work

Asking junior contributors to work faster isn’t enough to move the needle. First, there aren’t enough of them. Second, this overlooks leakages and inefficiencies between decision-making and execution, as well as in cross-team coordination.

Senior executives and managers must now deliver work outputs that they would otherwise have delegated to their direct reports:

• Marketing directors must manage their own campaigns.

• Sales managers must be given higher personal quotas.

• Engineering managers must start pushing multiple PRs per day.

• Finance managers must create dashboards and reports themselves.

• HR managers must take on more recruiting, onboarding, and operational tasks.

When managers start acting more like individual contributors, they must also demonstrate a healthy disregard for traditional departmental boundaries to minimize handovers. This means:

• Generating a basic graphical design themselves instead of asking the creative team to do it.

• Delivering prototype code to the engineering team instead of a PRD (product requirement document) or a Figma file.

• Querying the company’s systems for business intelligence instead of looking for the first available data analyst.

• Performing the first-pass review of a customer contract instead of always waiting for the legal team to respond.

All of this, while also being accountable for AI adoption within their teams.

Good work requires good tools

It takes a lot of computing power to do good work with AI. Compute power isn’t cheap, even when it’s subsidized by AI companies gunning to gain market share through their Pro and Max subscriptions.

Power users should be equipped with the best commercially available apps for their specific use case. Great commercial apps exist for AI use cases that are generic among many industries and companies, such as:

• Meeting recording and transcription: Fireflies.ai or Granola.

• Lead generation: Apollo, Clay, and/or Lusha.

• Coding: Claude Code and/or OpenAI’s Codex.

• Image/video generation: Weavy.

• Voice transcription/generation: ElevenLabs.

For other employees, basic subscriptions, such as Google Gemini or Microsoft Copilot, may be sufficient for simple tasks, but upgrading to a premium option should be available as these employees become more productive with AI.

Weekly usage should be monitored. If an employee makes little use of a subscription for a month, their manager should discuss whether the subscription is not needed for their job scope or whether the employee is not meeting the company’s expectations for AI adoption.

Always check if the apps have a privacy setting to opt out of helping to “improve the product”. You don’t want your company’s data used for anything.

Meaningful productivity gains require agents

Developer tools are always a step ahead of other AI tools, so coding assistants help us to understand where the industry is headed. Let’s take a look at software development.

In 2023 and 2024, coding assistants like GitHub Copilot and Cursor became widely adopted for research and code-completion tasks, but engineers saw only a 10-20% increase in productivity. Over the past 12 months, the use of coding assistants has evolved into agents that create and execute multi-step action plans to deliver complete software features. Some software engineers now work twice as fast by switching between 6 to 10 agents running in parallel, each managing a different task.

Similarly, for non-technical workers, merely asking ChatGPT, Gemini, or Copilot “how should I do X” questions is no longer best practice. Semi-autonomous agents are the preferred approach to AI-powered workflow automation.

Three big ideas:

• Humans no longer orchestrate multiple tools. Users interact with a single agent acting as their delegate, coordinating tasks among subagents, systems, and databases to complete workflows.

• The agent has access to skills, which are tailored instruction manuals for performing specific tasks, especially those related to the company’s standard operating procedures. These skills are text documents, far easier to maintain than the no-code flows hosted by previous-generation workflow automation systems like Microsoft Power Apps, Zapier, and N8N.

• When predefined skills and system integrations are not enough to fulfill the user’s request, the agent can access a computer environment, develop new code, and execute it to complete the task. In fact, most of these agentic platforms use a coding assistant as the underlying LLM (e.g., Codex 5.3, Opus 4.6, Kimi 2.5) rather than a general-purpose LLM.

Platforms like Manus, Claude Cowork, OpenClaw, Perplexity Computer, and Agentini Workflows enable companies to create and run agentic workflows.

If you are curious, here are some of their differentiating features:

• Manus: the pioneer. Now part of Meta, so it’s unclear where they are headed for enterprise use.

• Claude Cowork: relatively easy to use thanks to Anthropic’s vertically integrated model (their LLM, their app), but runs on each employee’s desktop computer.

• OpenClaw: very tenacious, with a “YOLO” vibe. Great for tech-savvy tinkerers, less for the average enterprise employee.

• Perplexity Computer: the new kid on the block. Very similar to Claude Cowork, but the computer is in the cloud instead (under Perplexity’s control).

• Custom solutions like Agentini: designed to be managed centrally and live in your Slack or Teams channels, it can better integrate with your company’s cloud environment, but requires an initial setup.

Get help from internal or external experts to make the most of these tools. It takes hundreds of hours to become proficient with them, which explains the gap between power users, who rave about them as if they’ve discovered fire, and other users, who continue to add goatees and cowboy hats to their ID photos.

Microsoft Copilot and Google Gemini have some agentic capabilities, but they are currently constrained by resource and security policies that make it impractical for them to coordinate complex tasks outside their respective ecosystems (Microsoft and Google).

Every SaaS vendor under the sun has launched its own agent offering as well (e.g., Salesforce, Slack, ServiceNow, Workday, Notion), with varying levels of maturity outside their own products.

Frontier, OpenAI’s agentic platform, is currently available only to large enterprises.

Be realistic about costs

One of the challenges of delivering ROI is that the cause/effect relationship is not “I’ll give everyone a Microsoft Copilot subcription and they’ll work 20% faster”.

Impacts must be extracted, workflow by workflow, eliminating tasks and removing handovers through a process re-engineering effort involving multiple stakeholders, systems, and databases.

An initial investment is required to map and redefine the process, either in time spent by the internal AI task force or by hiring an expert.

Given that the most important thing is to make the agent deliver relevant output, your team will start by using the most expensive LLM models (e.g., OpenAI’s GPT 5.2 with high reasoning or Claude Opus 4.6). Later, they can start downgrading to cheaper models and compare output quality.

All of this to say, agentic automation requires some resources and time, even though AI makes the process much more efficient than, say, implementing a new ERP or CRM system customization.

Hence, don’t try to do too many things at once. Your company’s AI task force can realistically handle around 3 high-priority use cases at any given time. Make sure that you pick those that are labor-intensive, repetitive, and well-documented.

Practical next steps

Your Board expects an AI productivity narrative, and for most companies, that’s a priority in 2026.

You should:

1. Set productivity improvement targets, ideally through hiring freezes.

2. Prioritize a handful of workflows with the greatest potential. Do this top-down. Don’t ask your employees for “ideas”.

3. Invest in the right project resources and tools to create custom agents.

4. Once the agents are up and running, aggressively question traditional job separations between managers and individual contributors, and between departments, so that each person starts doing the work of several.

The big idea: the only software interface that’s really standardized is the computer screen. Enterprise software vendors have not done a good job at standardizing and documenting their APIs over the last few decades. By contrast, any human can intuitively understand and use an application through a laptop’s screen, keyboard, and mouse. If you train an AI to do that, then it can replace any knowledge worker.

The importance of agentic computer use is acknowledged by other companies as well. Meta recently spent $ 2B to acquire Manus, a Singapore-based startup known for agents that complete long-running tasks on customers’ behalf using cloud-based computers. Anthropic launched a Claude browser extension for Chrome. H Company, a French startup, specializes in computer-use agents for Enterprise use cases.

What if the rest of us could instruct our custom AI agents to use computer browsers as well? Well, it’s early days, but we can already do interesting things, thanks to developer tools like browserbase, browser-use, and firecrawl.

Let’s explore how it all works through practical examples.

A simple use case

It’s not uncommon for an employee to spend 1h / week visiting some internal or external websites just to capture data and put it into an Excel spreadsheet:

• Checking what people are saying on Reddit or Discord about the company’s products.

• Capturing RFPs or data published daily by government agencies.

• Monitoring competitor prices.

• Downloading files from the ERP / financial system because it’s too old to have a proper API.

• Etc.

Obviously, a task is worth automating only if it is repetitive. For this blog post, we’ll use a simple example: opening a real estate website to see what new listings have been added since the day before and extracting the 10 newest listings.

Experience agentic computer use

The fastest way to experience agentic computer use is to use a consumer app: Manus, or the more recent BU app by Browser Use.

Using such an app first is a great way to quickly iterate with various prompts and compare their effectiveness, even if you want to code your own agentic system later.

Manus can control your laptop’s browser or access a cloud-based computer to browse the internet. Let’s select the second approach, since you’ll save time by not having to be at your computer to complete the task.

Here is example of prompt:

Or, you can try your own prompt.

You will notice that the agent often struggles to understand the web page’s structure and is much slower than a human. AI assistants have not yet been well-trained to use web browsers. On occasion, you may see that your connection is blocked by a bot prevention system. That’s another challenge of using agentic browsers.

In this example, however, Manus mostly succeeds in completing the task.

Some websites require the user to log in first. With Manus and similar apps, you can take control of the browser to perform specific actions, then hand it over to the agent.

Incorporate computer use into your agentic workflows

If you plan to automate processes that are more complex or involve accessing sensitive data, your company may need multi-step workflows and likely wants greater control over computers and credentials.

Several cloud providers make it easy to create ephemeral remote computers in the cloud and delete them as soon as the task is completed:

• Browser-use: works great and is simple to use. Browser-use is particularly suited to general-purpose use cases where you want to provide instructions in natural language only, with almost no programming. However, if your needs exceed the limits of the pay-as-you-go plan, you’ll need to pay $500/month for the next subscription tier.

• Browserbase: a powerful alternative that’s highly customizable. Compared with Browser-use, it feels more suited to use cases where having deterministic outcomes is important. It requires more work than Browser-use, but it is simpler to use than the legacy test automation frameworks like Playwright and Selenium. Browserbase works hand-in-hand with Stagehand, its agentic framework, and Director AI, a utility for creating Stagehand scripts.

• Firecrawl: primarily used for text scraping from websites. It does text scraping well and relatively cheaply, but it does not perform as well when agents must take online actions, such as typing text or clicking buttons.

• Vercel’s Agent Browser, Browserless, Hyperbrowser, Crawlee: interesting alternatives to keep on the radar.

With Browser-use

In our experiments, the browser-use SDK succeeded at the task almost every time, using the same prompt as for Manus.

Want to try it with your own set of instructions? See the link to a Google Colab notebook at the end of this post.

With Browserbase

To use the Browserbase platform, you’ll usually need to string together:

• A LLM: Google Gemini 3 Flash with a Google Gemini API key.

• Stagehand: Browserbase’s open-source framework for agentic browser use.

• Browserbase: The infrastructure where browsers are hosted in the cloud.

In our experiments, Stagehand didn’t always perform well when given a single set of instructions, such as the ones we gave to Manus.

However, there are two approaches that seem to work quite well:

Approach #1: Use Director.AI to help create a script, then feed it into your workflow to get repeatable results. That’s the easiest approach and also the most predictable one, if it works.

Approach #2: Create your own agentic loop in which a Google Gemini constantly takes screenshots, decides the next action, uses Stagehand to execute it, and so on until the goal is achieved. This approach is more technically involved but could serve as a fallback if the first approach fails.

Want to try it with your own set of instructions? See the link to the Google Colab notebooks at the end of this post.

Wrap up

The xAI leak really nails it: we’ve been waiting decades for enterprises to standardize their APIs, and it is not happening. Meanwhile, every app already has a universal interface — the screen. Why keep begging for better integrations when AI can just... use software like humans do?

Let’s be honest: these tools are still pretty rough. Agents get confused, they’re slower than you’d expect, and bot blockers are on the rise. But if you’ve been in tech long enough, you know how fast things improve once the big players start throwing money at a problem. Meta didn’t drop $2B on Manus for nothing.

If you want to start playing with this stuff, Browser-use and Browserbase are solid options. Pick something low-stakes and repetitive — like that annoying weekly data pull — and see how it goes. You’ll hit some walls, but you’ll also get a feel for where this is heading.

Once these agents get good enough, a lot of tedious “glue work” between applications goes away. That’s enough to explain why agentic computer use will be a major theme in 2026.

Prompts and code snippets

• Browser-use demo in a Google Colab notebook: link.

• Browserbase demo in a Google Colab notebook:

  • With a step-by-step script generated by Director.AI: link.

  • With “thinking loop”: link.

There are plenty of agentic tools and platforms, ranging from no-code platforms like n8n, Lindy, and Gumloop to the current king of vibe-coding, Claude Code, and of course, the current community favorite, OpenClaw (formerly known as ClawdBot). How to choose?

Below is a set of tasks you can complete at your own pace to become proficient in AI agents. There is no right or wrong answer, but some “solutions” are included at the end of the blog post.

Clarify your goals

It’s easy to get lost or keep switching tools when you are unclear about what you are trying to achieve. The first question is: what workflow are you trying to automate?

The criteria for workflow selection are as follows:

1. The workflow should be digital and document- or message-based. Obviously, AI agents aren’t going to carry your groceries. They also won’t be very effective at tasks that usually require a human-to-human conversation (e.g., sales pitch, check-in with your direct reports).

2. It’s something that you do every week and takes more than 1h/week. If the payoff isn’t there, it does not make sense to invest in automation.

3. The workflow can be documented in a memo you’d send to a team of college-level interns assigned to you for 2 weeks. You should assume that AI agents are pretty knowledgeable and smart, but lack judgment and do not learn through experience.

For this post, let’s take one workflow and implement it using various approaches. Let’s call it “my private banker.”

Introducing “my private banker”

“My private banker” is the workflow we’d like to automate. It generates investment ideas about once a week. She is not a hedge-fund genius, just a rigorous, hardworking advisor who researches the news and markets when you don’t have time to do it yourself on TradingView.

Her workflow consists of the following steps:

1. Calculate the value of your investment portfolio

2. Compare its composition with your target asset allocation.

3. Recommend 1-2 assets to sell to rebalance the asset allocation if required.

4. Look for high-quality, undervalued assets and recommend 1-2 tickers to buy.

Could “my private banker” be an AI agent? Let’s find out.

Prerequisites

First, write down your current portfolio in a Google Sheet. For example:

Generative AI models actually prefer to read JSON format. A JSON list is enclosed between square brackets, with individual items enclosed between curly brackets. If you’d like to quickly convert your Google Sheet data into a JSON list, you can find a utility at the end of this post.

To complete the proposed exercises below, you will need accounts for the following online services:

• Chatbots: I recommend gaining experience with Claude, ChatGPT, Perplexity, Google Gemini, and Manus.

• Financial data: get a free API key from Alpha Vantage.

• No-code platforms: Use a n8n trial account.

• Developer tool accounts: GitHub, Cursor, and Ona (using your GitHub account for login). Also, download Cursor onto your computer.

Using Chatbots

Task #1

Let’s start simple and see if Claude, ChatGPT, or Perplexity can generate investment recommendations for you.

To get the most out of chatbots, you’ll need to gain the following skills:

• Learn from hands-on experience which chatbot is good at what.

• Develop strong written communication skills to give clear instructions to chatbots. (This is sometimes called “prompt engineering.”)

So let’s do that.

✍️ Your first task: Create a detailed prompt instructing the chatbot to create trade recommendations based on your own goals and criteria, and run it in a few chatbots.

You can activate “deep research” when available and, in Perplexity, toggle the “Finance” tool to see if you are getting better results that way.

⏰ Complete the task, then continue reading below. There is no right or wrong answer; just do your best, and then you can see what I’ve come up with.

Results

The verdict? Pretty interesting in my case.

Asset allocation and portfolio rebalancing. The chatbots can retrieve market prices by searching free websites and correctly analyze my test portfolio. Their recommendations differ slightly, but they all make sense in the context of my assets and goals: Claude recommends selling VOO (U.S. index ETF) and buying KWEB (Chinese tech ETF), whereas Perplexity recommends selling AVGO (overvalued) and buying BND (fixed-income ETF).

New investments. All chatbots offer interesting recommendations, though they differ. Claude recommends buying CMCSA, VZ, BTI, and DUK, whereas Perplexity recommends KO, JPM, and UNH. Their recommendations are based on a combination of fundamental indicators (e.g., PE, P/FCF), technical indicators (e.g., RSI), and news/analyst commentary. The Perplexity assistant, who has access to a professional-quality financial dataset, did a better job of collecting indicators than Claude, who performed web searches like “most undervalued stocks in the U.S.” to find the data.

You can see my prompt at the bottom of the post.

Task #2

Let’s push further and see if we can complete a deeper technical analysis.

✍️ Your second task: Aggregating the buy recommendations from all chatbots, create a detailed prompt instructing each chatbot to further evaluate and rank these stocks according to a list of criteria defined by you.

Try it in a few chatbots.

⏰ Complete the task, then continue reading below.

Results

Here, I reached the limits of what chatbots can do with their off-the-shelf web search and data analysis capabilities.

The only chatbot that could finish the task is Perplexity, thanks to its access to a professional-quality financial dataset. High-quality financial data is usually not free or cheap online. If my metrics were based on intraday financial data, none of the chatbots would have succeeded.

You can download my prompt at the bottom of the post.

Conclusion so far: The only way for a chatbot to complete the task is to have access to a specialized financial data-collection tool. Given that other chatbots like Claude and ChatGPT don’t offer this tool, can we equip them with one?

After all, in a business context, it is highly unlikely that your chatbot will have access to your industry’s specific data sources, so you’ll need specialized tools.

Using a generalist agent with specialized tools

Task #3

A generalist agent is an AI chatbot equipped with tools that connect to your personal or work accounts.

The chatbot may run on your laptop (e.g., Claude Desktop, ChatGPT Desktop, ChatGPT Atlas, Perplexity Comet, and the much-hyped OpenClaw) or in the cloud (e.g., Microsoft 365 Copilot, Manus, RelevanceAI). Based on its conversation with you, it decides what actions to take on your files, online and/or through APIs on your behalf. It usually asks for your approval first, though.

Let’s equip a generalist chatbot (e.g., Claude) with a specialized financial data analysis tool powered by the Alpha Vantage online service.

✍️ Your third task: Configure Claude with the Alpha Vantage MCP tool, and then run the second prompt again.

To configure Claude with the Alpha Vantage MCP tool, follow the instructions here. Then test the integration by asking the chatbot: “Can you check if you have access to the Alpha Vantage MCP tool?” Finally, run your prompt.

You can also equip other chatbots with tools. If you try OpenClaw, make sure you research how to install it properly and patch the security vulnerabilities that non-technical users encounter.

⏰ Complete the task, then continue reading below.

Results

Claude can now create a spreadsheet comparing all the stocks against my quantitative evaluation criteria.

My prompt was the same as in task #2, but what changed was equipping Claude with the Alpha Vantage MCP tool.

Conclusion so far: This example shows the power of tools. In an enterprise context, you can imagine how important it is to create tools that connect AI agents to the company’s data to produce relevant outputs.

However, the chatbot approach does not eliminate all manual work. We had to query several chatbots, aggregate their buy recommendations by hand, then go for a second round of analysis, all the while staying in front of our screens to prompt the chatbots and read their answers. Could all of this happen automatically instead?

Using a specialized agent on a no-code platform

A specialist agent specializes in a single workflow, broken down into predefined steps. This is interesting for enterprises because it means the agent’s behavior is more predictable than that of a generalist. Each step is delegated to a dedicated AI assistant that executes specific instructions in natural language.

Task #4

A no-code platform like n8n lets you define workflows visually, inserting AI assistants and tools wherever you need.

To be honest, if you are just starting with n8n, the learning curve is steep. You should definitely work with two browser tabs side by side, n8n, and a chatbot like Claude to help you. Break your work into small pieces, and ask the chatbot to guide you.

✍️ Your task (optional): Use n8n to create a workflow that completes tasks #1 and #2, analyzing the portfolio and coming up with an initial list of trade recommendations, which are subsequently refined through technical analysis.

This task is optional because I am not convinced that no-code platforms are the way to go in 2026, given the tremendous capabilities of AI coding assistants. Every no-code platform comes with its own limitations, whereas code is infinitely customizable and can now be generated and maintained cheaply with the help of AI.

Still, no-code platforms can make sense for small companies.

⏰ Complete the task, then continue reading below.

Results

n8n makes it easy to create a single agent that can run many tools. I created an agent for the first task and added a sub-agent to complete the subsequent financial analysis.

The results are not significantly better than those from previous approaches, partly because the agent continues to perform web searches in a loop and does not always reach a final answer. Further logic customization is possible, but very time-consuming.

You can download an example workflow for this task at the bottom of this post.

Using a specialized agent created with the help of coding assistants

Using a coding assistant on a laptop for the first time can be challenging because it requires installing not only a code editor like Cursor but also many languages and libraries.

Hence, for this tutorial, let’s learn to work in the cloud: the code will be stored on GitHub, your development environment will run on Ona, and you will control that environment from Cursor on your laptop.

Task #5

Let’s get started:

• Go to GitHub and click “New” to create a repository. Type any name, such as “my_private_banker”. Don’t forget to toggle the “Public” button to “Private”. Also, switch on the “Add README” toggle.

• In Ona: create a new project after connecting your GitHub account. Select the repository that you just created. Select a “Small” environment.

• Once the environment is running, click “Open Cursor”. This will open the environment in your laptop’s Cursor app. Rearrange the window tabs (in View) so that you see the file explorer, the code editor, the agents, and the terminal. Select Opus 4.5 as the Agent.

You should be seeing this:

You can use the dialog box to instruct the agent to code your workflow!

✍️ Your next task: Use the Cursor agent to create a Python application that performs an end-to-end portfolio analysis and generates a final report. Build the application by instructing the Cursor agent step by step, tool by tool.

Don’t forget to save your work regularly in GitHub, as the Ona environment is deleted regularly. Refer to the bottom of this post on how to do that.

Tips:

• Ask the assistant to use “Streamlit” for the chat interface.

• Ask the Cursor agent to use LangGraph as the framework for agentic workflow management.

• Put the assistant in plan mode before agent mode.

• Instruct your assistant to search online for how to use LangGraph, Perplexity Search API, and Alpha Vantage’s Python wrapper.

⏰ Complete the task, then continue reading below.

Results

It takes much more time to write step-by-step requirements for the coding assistant than to give a generalist agent an overall goal and let it handle it, as we did in tasks #1 and #2. However, the research improves tremendously in terms of quality, depth, and predictability: hallucinations can be avoided, and financial analysis metrics can be fine-tuned at will.

My conversation with the coding assistant spanned dozens of messages. Sharing it verbatim here would not capture the essence of the conversation. But you can download some of the key prompts at the bottom of this post.

Wrap-up

It’s pretty clear that creating specialized agents with the help of coding assistants is the most enterprise-ready approach.

Its main advantage is that it allows you to evaluate and refine workflows with precision over time, based on user feedback. Agentic workflow automation is a process of continuous improvement.

Having said that, generalist assistants are improving rapidly, and it’s definitely worth trying them as the first port of call before developing specialized workflows.

Prompts and code snippets

Refer to the links below to download some of the work outputs.

Prerequisites

• Utility to convert your Google Sheet data into a JSON list: Google Colab.

• Example of a portfolio in JSON format: link.

Task #1

• Prompt (for a dummy portfolio): link.

Task #2

• Prompt (for a dummy portfolio): link.

Task #3

• Steps to configure the Alpha Vantage tool in Claude: link.

Task #4

• n8n workflow (save it as a JSON file and “Import it as File”to use it in n8n): link.

Task #5

• Instructions on saving your work from Ona to GitHub: link.

• Code generation prompts: link.

We are at an inflection point in AI maturity and adoption, which means that, if you are starting a company this year, it makes sense to build an AI-native company.

Any company, even in a traditional industry, can be AI-native if it is decisive about using generative AI across the organization. Being AI-native is not just about using AI chatbots or vibe-coding apps. It means systematically automating repetitive tasks at the individual and team level, so that AI augments everyone’s productivity.

As far as I can see, there are 5 things that AI-native companies do consistently:

1. Codify knowledge in text format.

2. Become experts at using the best available AI tools.

3. Adopt AI-friendly enterprise applications.

4. Automate workflows with a centralized agentic stack (this is going to be the hot topic in 2026!).

5. Manage the organization to build AI capabilities at every level.

Subscribe now

Codify knowledge in text format

The biggest AI adoption challenge for traditional companies is access to internal knowledge. Large language models do their best work when they can easily ingest company knowledge in text format.

Yet, at many companies:

• Standard operating procedures are poorly documented or executed differently from the way they are documented.

• A lot of knowledge is codified in formats that are hard to access (hard drives, departmental silos) or that LLMs are not good at reading (slides, PDFs, spreadsheets).

At AI-native companies, each department head is accountable for maintaining a knowledge repository that AI assistants can leverage for context.

The knowledge repository must meet the following requirements:

• Format: primarily structured text (e.g., Google Doc, Markdown).

• Metadata: each file’s content must include a header with a title, summary, keywords, reliability score, and creation and modification dates (also called “front matter”).

• Access: each folder can be accessed by authorized AI assistants (e.g., via a service account or an API key for a MCP tool connected to a RAG vector database).

In 2002, Amazon introduced the “Bezos API Mandate” stipulating that each team must expose its data and functionality to other teams through service interfaces and that anyone who does not do this will be fired. In a way, in 2026, as a CEO, your “AI mandate” should require teams to expose their knowledge repositories to one another and to AI agents in text format.

Become experts at using the best available AI tools

In addition to chatbots, there is a growing number of specialized AI-powered apps that significantly improve employee productivity.

AI-native companies are experts at using these tools.

Chatbots

At AI-native companies, all employees have access to at least one company-approved chatbot subscription, and power users can request additional subscriptions.

Not all chatbots are equally great at everything. Power users switch between:

• Claude for tutorials and problem-solving in technical domains.

• Perplexity to speed up web searches.

• Google Gemini for image and video generation, and for book/article reading when combined with NotebookLM.

• ChatGPT for brainstorming and writing, and for day-to-day/lifestyle questions.

• Microsoft Copilot …. when their company’s security policies block other chatbots.

• Occasionally, Manus is used for tasks that require extensive internet browsing and clicking, as well as data collection and analysis.

Power users also have strong written communication skills, enabling them to give clear instructions to chatbots. (This is sometimes called “prompt engineering.”)

Recommended in 2026: Claude, ChatGPT, and Gemini all do a fine job.

Meeting transcription and summarization

Every meeting should be automatically recorded, summarized, translated (if it’s not in English), and converted into knowledge, summaries, and action items, which are fed into the company’s systems.

Google and Microsoft’s native meeting recording tools are strong at recording and transcription, but weaker at integrations.

Specialized apps like Zoom, Otter, and Fireflies, which can be used alongside these tools, offer extensive CRM integrations. IT administrators like these cloud-based solutions because they don’t require installing untested software on employee laptops, and make it easy to store all content centrally, which is important for sales and customer service teams.

If you are a sales leader, you should regularly review your team's call transcripts and provide feedback (possibly with AI assistance).

A point of caution: I find these apps overly eager to create to-dos, and I would not recommend automating task creation without a human-in-the-loop.

Some people find it intrusive when AI meeting recorders join video meetings. For that reason, some companies (such as VC investors) use apps that run directly on their laptops/phones (e.g., Granola, Notion Meeting Notes) or on standalone devices (e.g., Plaud). But this is the exception rather than the rule.

Recommended in 2026: Fireflies, for its ease of use and its convenient API (to customize how meeting data is imported into company systems). Combine it with a systematic process that imports relevant information into your departments’ knowledge repositories.

Creative asset generation (brand, marketing, product)

No doubt, image and video creation are some of the most impressive capabilities of generative AI.

In my view, this is an area where we moved from “anyone can create an agency-quality ad” to “we still need creatives with world-class skills and taste, who are also experts at AI-powered content production”. Not many people have the taste and attention to detail required to move across multiple design tools and deliver stunning, original assets.

AI-native creatives juggle multiple models and tools like Google Flow, Reve, Krea, Leonardo, Kling, HeyGen, Descript, Wan, ElevenLabs, and others.

Rather than sticking to a single ecosystem, it makes sense to use a model aggregator that supports the end-to-end production workflow, such as Weavy (now part of Figma) or Higgsfield.

Recommended in 2026: Weavy for their diverse model choices.

Software development

At the time of writing, the best coding assistants are Claude Code (with Opus 4.5) and Cursor (with several available models). Both are concise and smart. Claude Code is more forgiving if your instructions lack precision.

If your company has a software engineering department or subcontractor, it is essential that they have access to the best tools and know how to use them.

Even if your company is not a tech company, you will have a core group of analysts/engineers who use coding assistants to automate processes.

AI-native companies make advanced use of AI coding assistants:

• They employ experienced software engineers to design system/application architecture and development workflows, and review AI-generated code.

• They generate most of the new code by typing requirements in text.

• They use distinct agent roles to execute the sequential phases of feature development (requirements, architecture, ticket creation, development, testing).

• They launch multiple agents to work in parallel (3-6 per person).

• They customize the assistant configuration with rules, skills, subagents, and MCP tools (including technical documentation tools).

Recommended in 2026: Claude Code + Cursor.

Adopt AI-friendly enterprise applications

There are dozens of startups re-building enterprise SaaS suites with an AI-first approach. Their goal is to become the next Microsoft Office (for documents), Slack (for messaging), Atlassian (for task management), Salesforce (for CRM), SAP / Oracle (for ERP), Workday (for HRIS), or Microsoft (for all of the above).

These offerings are not yet fully mature. It takes a long time to develop fully featured enterprise SaaS offerings, and to have all the functionality required for your business and your countries of operation. That’s the reason why, although open-source software has been around for a long time as an alternative to these software giants, most companies still rely on proprietary enterprise solutions.

In my view, it is much more important to adopt AI-friendly than AI-first applications.

AI-friendly applications are committed to adding AI-powered features over time, but, more crucially, they have a track record of being developer-friendly, allowing enterprise and third-party integrators to easily build custom features on their platforms. They have extensive, consistent, stable, and well-documented APIs, as well as a large existing ecosystem of third-party integrations.

By contrast, SaaS vendors become obstacles to AI adoption by making it hard for enterprises to build custom integrations quickly and iteratively. For example, they make their documentation hard for AI assistants to navigate and have long approval times for new API keys and integrations.

In the sections below, let’s review the main types of enterprise software and the main solutions available to someone building a new company in 2026.

Office documents and collaboration

For email, docs, spreadsheets, and presentations, the Google Workspace ecosystem wins for two main reasons:

• Google Workspace is definitely, unapologetically cloud-first. It offers an integrated experience that AI assistants can easily get their “heads” around, unlike Microsoft’s, which is fragmented by service, platform (local vs. cloud), and environment (work vs. personal). Fragmentation confuses AI assistants.

• The start-up ecosystem’s integration with Google products is unparalleled. Whatever your specialized productivity need, there is likely an integrated AI app or Chrome extension being built to address it. And if there isn’t, vibe coding your own integration is pretty easy.

Recommended in 2026: Google Workspace.

Internal messaging

Slack’s open architecture and extensive ecosystem of integrations make it the winner, for now, compared to Microsoft Teams, for example.

Building an AI-powered Slack integration and getting it approved by your system administrators is surprisingly easy and well-documented.

Recommended in 2026: Slack (or Discord for an affordable alternative).

Task management

Linear has become the go-to task management app for modern tech companies (such as OpenAI), while competitors like Atlassian and Asana have become bloated and slow.

Linear was initially designed for software developers, but it is appreciated by other audiences for its speed, simplicity, keyboard shortcuts, and powerful API. Its MCP server can be integrated easily with any AI assistant that supports it.

Recommended in 2026: Linear.

CRM (Sales and Marketing)

Salespeople dislike Salesforce because it is clunky and expensive to set up, yet even OpenAI and Anthropic reportedly use Salesforce for Enterprise sales. Salesforce’s superpower is its more granular user permissions than HubSpot and other CRM systems.

Having said that, if you are starting a company in 2026, Salesforce is likely overkill. HubSpot is a common choice among smaller companies.

Whatever the choice, the ideal user experience should be that my emails and meetings are automatically used to enrich CRM records, and I can use idle time in the subway or at the airport to converse with my AI assistant, who fills in the gaps and updates opportunity sizes and timeframes on my behalf. To make this happen, you need to create your own agentic workflows. More on this later.

Recommended in 2026: HubSpot + custom agentic workflows.

Customer success/service

Many tech companies are working to reinvent customer service using AI. Up-and-coming players include Sierra, Aisera, and others. Among established companies, Intercom, ServiceNow, Salesforce, and Zendesk have been communicating heavily about their new AI features.

From what I hear, Intercom works really well and is developer-friendly. It also has a strong track record of reliability and security.

Recommended in 2026: Intercom.

FIS and ERP (Finance, Supply Chain)

AI-native companies like Campfire (ERP) and Aleph (Financial planning) are getting good reviews, but established modern applications like NetSuite, Workday, and Rippling remain top choices thanks to their broad functionality. More affordable alternatives like Zoho Books are still popular, with robust APIs.

Recommended in 2026: Zoho Books, NetSuite, or Workday, depending on your needs and budget, plus custom agentic workflows to get really powerful automations. But before signing on, check if they impose any restrictions on API integrations.

HRIS

In this category, too, newer startups like Warp and HiBob are getting good reviews, but Workday and Rippling remain strong contenders. More affordable options exist for smaller companies, but the list varies by country. Providers include Employment Hero and BambooHR.

Recommended in 2026: varies by country, but HiBob and BambooHR are both solid choices.

BI (Business Intelligence, Analytics)

One of the promises of AI applications like Omni and Fabi is that every employee should be able to query company data in natural language and get whatever report they need to make business decisions.

We are not there yet, not because of the limitations of AI technology but because of how fragmented and messy company data often is.

Setting up a centralized warehouse, supported by strong analytical and visualization capabilities, is the necessary first step. From that standpoint, the Google Cloud environment is capable, easy to use, and easy to expand, with BigQuery for the data warehouse, Looker for visualization, and needs-specific services for data import and processing.

Recommended in 2026: Google Cloud.

Automate workflows with a centralized agentic stack

AI-native companies use agentic workflows to automate tedious processes.

Examples of use cases include:

• Parse and pre-approve invoices and expenses before any Finance team member sees them.

• Filter candidate applications before any recruiter looks at them.

• Generate lists of qualified customer leads for the sales department.

• Analyze customer emails and meeting recordings to automatically update the CRM.

• Enable employees to quickly and comprehensively search the company’s knowledge base.

• Send personalized communications to customers and employees.

• Read customer and vendor contracts to highlight problematic clauses.

• And much more.

An agentic workflow is a work process delegated to one or several AI assistants, each equipped with instructions and tools, in which the assistants demonstrate some degree of autonomous decision-making. Tools are capabilities beyond simple chats, such as reading files or data, updating a database, browsing the internet, or sending messages. They can be thought of as plugins or connectors.

Today, enterprise applications tout their agentic capabilities, yet most are not mature enough to support your company’s complex workflows. You end up paying an integration consultant to make Salesforce or ServiceNow agents work.

And yet, thanks to coding assistants, it has never been easier to create custom apps that execute almost any of your company’s workflows.

AI-native companies make it easy for departments to automate their workflows using AI agents:

• Their CEOs prioritize the highest-payoff opportunities.

• They define a centralized architecture for agentic workflows: what tech stack, where to host the agents, and who is responsible for maintaining shared tools.

• They build small teams of AI leads in each department, who are responsible for getting AI automation projects underway before handing them over to regular teams.

• They build evaluation and safety standards to monitor and improve the performance of AI agents.

There are two ways to implement agentic workflows: generalist agents and specialized agents (predefined workflows).

Generalist agents

A generalist agent is an off-the-shelf AI chatbot equipped with connectors to your work accounts.

It may run on your laptop (e.g., Claude Desktop, ChatGPT Desktop, ChatGPT Atlas, Perplexity Comet, OpenClaw - formerly known as ClawdBot) or in the cloud (e.g., Microsoft 365 Copilot, Manus, RelevanceAI). Based on its conversation with you, it can take actions on your behalf with your files, online, and/or via APIs.

Frequent use cases include:

• Reorganizing your files and folders.

• Summarizing PDFs or documents.

• Summarizing your Gmail inbox.

• Helping to prepare your upcoming daily appointments.

• Querying services like Linear, Asana, Notion, Salesforce, or HubSpot via a chat interface.

Generalist agents can speed up many tedious tasks, but generally fall short of reliably managing business-specific, rules-based workflows.

Recommended in 2026: Claude Cowork.

Specialized agentic workflows

Specialized agents will be big in 2026. They are usually run in the cloud, meaning that they are active even when you are not at your computer. Each agent specializes in a single workflow, broken down into predefined steps (“nodes”), each of which is delegated to an AI assistant that can exercise some degree of autonomy in completing the step using tools and/or requesting additional input from the user.

Specialized agents aim to be more deterministic and auditable than generalist agents. For example, at AGENTINI, we deploy custom workflows to support the end-to-end proposal writing/quotation preparation process, among others.

Some companies use a no-code platform like Zapier, Make, N8N, or Microsoft (Copilot Studio, Power Apps) to create basic agentic workflows. But no-code platforms still have significant limitations. They focus on generic use cases that apply to many companies, and they are not designed to handle tasks that take several minutes or several hours (such as parsing PDFs or images). Companies using these tools risk focusing on the workflows that are easiest to automate, rather than those that deliver the biggest payoff for the business.

By contrast, AI-native companies use custom instructions and tools for their AI assistants, while centrally leveraging established frameworks such as Langchain or CrewAI to orchestrate these solutions. LangChain and LangGraph, in particular, are widely used by companies like Uber, LinkedIn, Replit, Elastic, Cisco, GitLab, and Vodafone.

Here is an example of what a LangGraph workflow looks like:

When building an agentic workflow platform, what’s most important is to integrate the company’s customers, employees, knowledge, and systems of record with AI-powered business rules. This usually involves a combination of generic tools and custom-developed ones. This means that your AI team must take ownership of both internally maintained and outsourced data stores and tools that align with your company’s specific workflows.

Manage the organization to build AI capabilities at every level

AI-native companies typically start by mandating a mostly dedicated central AI team, led by the CEO. This does not mean that experimentation is discouraged in other parts of the organization, far from it. But keeping track of the industry’s rapidly evolving trends is too challenging to be left to individual initiative.

Team members can continue to report to their departments, but they must allocate at least 50% of their time to the task force. They prioritize AI initiatives, own and maintain methods, reference architectures, and shared tools, and mobilize like “SWAT teams” to help departments get their AI projects started.

When forming their AI team, companies often assume they need AI experts. The truth is, they need experts in applying AI, not in creating it. Anyone can become an AI expert if they are 1) more than 50% dedicated to the task; 2) tech-savvy (can write and read Python code); 3) experienced enough with the business to quickly understand business processes and prioritize areas where AI is useful; 4) pretty smart and resourceful; and 5) strong written communicators.

Over time, each department must embed AI capabilities within its own org.

Wrap-up

Building an AI-native company in 2026 involves reimagining how work gets done and treating AI capabilities as core infrastructure, much like internet connectivity or cloud computing became table stakes in previous decades.

The five pillars outlined in this post work together. Codifying knowledge in text makes it accessible to AI tools. Expertise in those tools enables employees to work more effectively. AI-friendly enterprise applications provide the foundation for integration. Custom agentic workflows automate repetitive tasks. And building organizational AI capabilities ensures these advantages compound over time.

The companies being founded today have an extraordinary opportunity. They can build AI-native processes from day one, avoiding the painful retrofitting that established companies must undertake. But this window won’t stay open forever.

AI coding assistants have grown impressive reasoning capabilities, particularly Claude Code Opus 4.5. However, they often struggle to generate code that properly integrates connected services and APIs.

Three scenarios are especially common in an Enterprise context:

1. The assistant’s code uses an outdated version of a particular service’s API. This is especially common with AI-powered services undergoing rapid change as they add more features, including Google Cloud, Pinecone, Firecrawl, Browser Use, financial and cryptocurrency services (e.g., Polymarket, ethersjs), and others.

2. The assistant’s code is incompatible with the company’s requirements, which may include a specific version of a development language (e.g., Python, Node, .NET) or a specialized internal service (e.g., internal database, document store).

3. The assistant simply doesn’t know the API in question because it’s too specialized, and hallucinates.

A naive approach would be to ask the coding assistant to search all the docs online before starting its work, but that would be very expensive in terms of token consumption/context window.

The better approach? Ask your company’s AI task force to maintain an internal “read the docs” MCP tool available to every developer.

Creating the tool is straightforward and involves 3 steps:

• Scrape the documentation from each service’s website.

• Break down the content into chunks and store it in a vector database (for RAG, or retrieval-augmented generation).

• Run an MPC tool that retrieves the relevant chunks at each coding assistant’s request.

This blog post summarizes the approach and includes a Google Colab notebook for reference (at the end).

I also mention a third-party service that your team can consider.

Scrape online documentation

Documentation scraping should be done regularly to account for updates, for example, every month. We can use Firecrawl for this.

The Firecrawl Python SDK can be used to generate a set of Markdown documents from a documentation domain, such as https://docs.polymarket.com/.

For a source like Polymarket, this will take 5-10 minutes and generate around 300 documents. You can follow the progress of the task in the Firecrawl console.

The result is a lot of files!

Refer to the Google Colab notebook for an example of how to do this.

Chunk and ingest the docs

For this step, we can use:

• Pinecone as the vector database. We can use a single index for all services, as long as each service is stored under a dedicated namespace folder to make it easier to refresh the information associated with each service.

• OpenAI’s text-embedding-3-large model for embedding generation.

When breaking down Markdown documents into chunks, it is important to track navigation metadata, such as page URLs and section headings, so that each piece of information can be traced back to its original source.

The chunking process results in a big JSON file (in this case, we have around 900 chunks):

The embedding process results in a big list of “vectors” which are subsequently ingested by the vector database:

Run the “read the docs” tool

The “read the docs” tool is basically an MCP server hosted by your internal team.

Here’s an example of a query and a response.

To create the MCP tool, you simply need to give that code to your favorite AI coding assistant and ask it to make a MCP server out of it using the MCP Python SDK. (MCP servers are tools for AI assistants.)

In an Enterprise context, it usually makes sense to host the MCP server at a URL protected by user authentication. The MCP standard requires authentication through OAuth 2.0 and a separate identity provider (such as Auth0) that supports “Dynamic Client Registration”.

Ironically, creating MCP servers with authentication is poorly documented, and coding assistants struggle with that… Stay persistent, and it’ll work out!

Anyway, once the tool is live at your chosen URL, integrating it with your AI coding assistant is very simple. It depends on the assistant you choose. For example, in Claude Code, you create a mcp.json file at the root of your project directory:

{
    "mcpServers": {
        "technical-docs": {
            "type": "http",
            "url": "<URL OF THE MCP SERVER>/mcp"
        }
    }
}

Alternative: third-party service

If your team uses mostly common languages and APIs, they can consider Context7 as an alternative to maintaining the MCP server themselves.

Context7 offers MCP servers with access to various docs curated by themselves or the community, with free and paid plans.

Wrap up

Building a centralized “read the docs” MCP tool transforms how your engineering team works with AI coding assistants. Instead of each developer wrestling with outdated API calls or manually pasting documentation into prompts, the knowledge stays up to date and accessible to everyone.

The investment is modest: a few hours to set up the scraping pipeline, a vector database that doesn’t cost much, and an internal MCP server.

The payoff is immediate—fewer hallucinations, less time spent debugging integration issues, and AI assistants that actually understand how your specific stack works.

Start small. Pick one or two external services that are causing the most friction for your team, scrape their docs, and get the MCP tool running.

This approach works well for other company knowledge as well, but technical docs are the low-hanging fruit.

Code snippets

Subscribe to see the link to the code snippets.

Here is the link to the Google Colab notebook.

Talk with your data

One of the use cases of AI chatbots is to act as a copilot to business users in their data analysis tasks. However, most of the time, this power is used in an ad hoc manner that barely scratches the surface of how companies can integrate generative AI into their data reporting workflows.

Subscribe now

What most people do today

Ask your colleagues how they use AI to “talk with their data,” and the following scenarios prevail:

• Microsoft Office Copilot. When looking at datasheets, users are presented with ideas on how to make the most of Excel’s table formatting and charting features. Those ideas are great, but they primarily focus on learning new ways to use Excel.

• Dump data files into ChatGPT and ask it to analyze them. In this scenario, ChatGPT generally seeks to summarize the data, displaying totals and averages by category and identifying trends (e.g., growth, correlation). This can be helpful in making sense of large datasets, but without an understanding of the business, it rarely produces dashboards with genuine insights.

• Use coding assistants to auto-complete or generate SQL queries. In my experience, this approach usually produces some helpful boilerplate queries, but it ultimately requires someone who can read and write SQL to finalize the analysis.

• Chatbots integrated into data platforms. Many tech companies have already integrated chatbot assistants into their reporting interfaces. For example, you can use a chatbot to query your Google Cloud resource usage, your website traffic in Google Analytics, and blockchain activity statistics in Dune Analytics. However, such chatbots are usually limited to the specific context and data of the platform that offers them.

Most of these use cases are ad hoc. By contrast, within a company, most data analysis tasks are recurring and iterative. In other words, people produce daily, weekly, or monthly dashboards and refine them over time.

What businesses need

To enable business users to generate dashboards and insights from company data, a more holistic approach is needed.

The vision looks like this:

1. All company data is exported to a data warehouse of permissioned read-only database tables. Users have access only to the tables for which they are authorized, and they can generate queries against these tables without ever accessing the original business-critical databases that support the business.

2. All data is described in natural-language “data manifests”. These are Markdown-formatted documents, where each table is described not only in terms of column definitions but also in terms of what each field means to the business.

3. Business users can request reports and dashboards in natural language. These requests are not isolated, ad hoc queries. They are collaborative documents that other business users can reuse or adapt to run their daily, weekly, or monthly analyses, and refine over time.

“Talk with your data”: making it happen

The “talk with your data” app typically uses standard components, but leverages data, data manifests, and report queries that are highly specific to your business.

It relies on:

• A chatbot-type interface where users can converse with an AI agent, but also see data reports and charts.

• A collaborative document folder where data manifests and report queries are stored in natural language.

• A main AI agent (the orchestrator), which plans how to respond to user queries and calls relevant tools.

• A data analyst AI assistant, which can be called by the main AI agent to execute specific tasks as part of the overall plan, and has access to a standalone computer (a coding sandbox) where it can try out various queries to explore the data and generate requested reports.

• A data presentation AI assistant, which can be called specifically to format reports in a user-friendly way according to the company’s standards. For example, it formats numbers, shades table cells green or red depending on whether they contain positive or negative growth rates, and so on.

Here is a visual representation of these components working together.

Let's review each component individually.

User interface

For internal company users, Streamlit is a great framework for implementing a chatbot while also displaying data tables and charts.

Here’s an illustration:

Data manifests and report queries

Data manifests and report queries are typically formatted in Markdown, a type of simple text format that existing documents can be easily converted into.

A data manifest describes the data and how it can be used. For example:

A report query describes the requested table in natural language, using examples and tables to make the description as concrete as possible:

The essential insight here is that data manifests and report queries are not ad hoc. They are maintained and refined over time to cover the business’s entire suite of daily, weekly, and monthly data reports and dashboards.

Main AI agent (the orchestrator)

The orchestrator agent’s primary task is to devise a plan to address the user’s query and review the work of the underlying assistants (data analyst assistant and data presentation assistant) to produce the exact report the user is requesting.

The orchestrator agent’s prompt looks like this:

# Overall context

You are a data analyst supporting the sales operations team at COMPANY.

You are responding to natural language queries written by the user, requesting to analyze the data in a certain way in order to generate reports requested by management.

Using available AI agents and tools, your end goal is to generate a **text response** that includes the following four sections:

-   Your plain text message to the user, which summarizes the status and results of your work, enclosed between the <plain_text> and </plain_text> tags.

-   The text content of the CSV file containing the report requested by the user, in comma-delimited format, enclosed between the <csv_report> and </csv_report> tags.

-   The full Python code that generates the CSV report answering the user's query and saves it under the `./report/report.csv` path. You will enclose this code between the <python_code_for_report_generation> and </python_code_for_report_generation> tags.

-   A Python code snippet that can be used to visualize the content of the CSV file as a user-friendly table in Streamlit, assuming that the CSV file has already been loaded into a Python Dataframe called `report_df`. You will enclose this code between the <python_code_for_streamlit_visualization> and </python_code_for_streamlit_visualization> tags.

# Responsibilities

Your workflow typically includes the following actions:

1. CREATE_PLAN: Understand the user’s query, clarify any ambiguous points, and create an execution plan.

2. EXECUTE_STEP: Execute a planned step yourself, such as updating some parts of the response in order to better answer the user’s query. At each step, you must maintain an up-to-date state and plan.

3. DELEGATE_STEP: Delegate a planned step to one of the available agents, such as REPORT_CODE_GENERATOR or STREAMLIT_CODE_GENERATOR. At each step, you must maintain an up-to-date state and plan.

4. REVISE_PLAN: Adapt your plan based on intermediate results, and go back to EXECUTE_STEP or DELEGATE_STEP if needed.

5. FINALIZE: When you are satisfied with the <plain_text>, <csv_report>, <python_code_for_report_generation>, and <python_code_for_streamlit_visualization> contained in the state, you can set the next_action to FINALIZE.

The EXECUTE_STEP and DELEGATE_STEP steps may be omitted or repeated several times, depending on the user's query and the output of previous steps.

…

Data analyst AI assistant

The data analyst assistant is an AI assistant that has access to a coding environment for experimentation and report generation.

E2B is a great platform for coding environments. You can pre-define Docker images that already include environment variables and the specific Python libraries available to the data analyst assistant, so that no time is wasted configuring the environment.

The data analyst assistant’s prompt looks like this:

# Overall context

Using available AI agents and tools, your end goal is to generate a **text response** that includes the following four sections:

-   Your plain text message to the user, which summarizes the status and results of your work, enclosed between the <plain_text> and </plain_text> tags.
-   The full Python code that generates the CSV report answering the user's query and saves it under the `./report/report.csv` path. You will enclose this code between the <python_code_for_report_generation> and </python_code_for_report_generation> tags.

# Workflow

## Workflow overview

You must follow the workflow steps described below:

1. **Explore**: using the `execute_python` tool as many times as needed, you execute POSTGRESQL queries to explore available data and their shape, determine what data you need for the report, and trial-run the code that generates a CSV report answering the user's query, until you are confident that you have generated Python code that produces the required CSV file and saves it under the `./report/report.csv` path.
2. Finalize: when you are satisfied that the Python code produces the required CSV file, you return your final response, including the <plain_text> content and the <python_code_for_report_generation> code.

## Using the `execute_python` tool

The `execute_python` tool executes your Python code within a single code cell in a Jupyter notebook, within a sandbox environment, and returns the output.

The sandbox environment has access to the POSTGRESQL tables described in the Data Manifest. To query the data tables, you must use the psycopg2 library, which is pre-installed in the sandbox environment. A connection string is provided in the DB_CONNECTION_STRING environment variable, which means that you can create a database connection with read-only permissions using the following code snippet:

```python
import os
import psycopg2
db_conn = psycopg2.connect(os.environ.get("DB_CONNECTION_STRING")
db_cursor = db_conn.cursor()
```

Typically, your code will query the POSTGRESQL tables to load the relevant data into Python DataFrames, and then manipulate these DataFrames to produce the required report.

The following libraries have been already installed into the environment using pip: pandas, numpy, openpyxl, pytz, psycopg2 and streamlit.

Data presentation AI assistant

If you are using Streamlit as the user interface, the data presentation AI assistant is just an assistant that’s focused on making your reports and charts look good in Streamlit.

Sometimes, the data presentation AI assistant won’t even be called by the main AI agent if the data is simple enough, because the main AI agent can generate basic visualizations directly.

The data presentation assistant’s prompt looks like this:

# Overall context

An AI assistant has generated a Python DataFrame called report_df, which answers the user’s query. The content of report_df is provided below in CSV format, between the <report_df_csv> XML tags.

Your task is to generate a Python code snippet that can be used to visualize the content of the report_df DataFrame as a user-friendly table in Streamlit.

The formatting instructions are provided below between <formatting_instructions> XML tags.

You will enclose your code response between the <python_code_for_streamlit_visualization> and </python_code_for_streamlit_visualization> tags.

Your code will be executed by the user within a Streamlit app using the `exec()` method to display the report as a DataFrame.

# Rules

Your Python code will be executed after the following lines of code, which therefore do not need to be included in your response:

```python
import streamlit as st
import pandas as pd
import numpy as np
import datetime
import json
st.set_page_config(page_title="Report", page_icon="📊", layout="wide")
st.title("Your report")
report_df = pd.read_csv("./report/report.csv")

```

At a minimum, your code should display the `report_df` DataFrame in the Streamlit app, so for example your response could be:

<python_code_for_streamlit_visualization>

```python
st.dataframe(report_df)
```

</python_code_for_streamlit_visualization>

To make the table more user-friendly, you should format large numbers with comma-separated thousands and zero decimals, and format percentages with the percent sign and zero decimals, so for example your response could look like:

<python_code_for_streamlit_visualization>

```python
styled_df = report_df.style.format(
    {
        "example_column_with_large_numbers": lambda x: f"${x:,.0f}",
        "example_column_with_percentages": lambda x: f"{x*100:.0f}%",
    }
)
st.dataframe(styled_df)
```

</python_code_for_streamlit_visualization>

However, the user may request specific formatting, in which case you will follow the user’s instructions.

Putting it all together

Here’s a video showing how the main AI agent receives a query and leverages available tools to generate a report for the user.

Note: the prompts shown in this blog post have been shortened and simplified for readability, compared with how they’ve been used on real-world projects.

Part of their promise is that anyone can dream up an app and launch it on their own, whether they are students, product managers, office workers, or C-level execs.

In reality, some of the most powerful desktop-based agents, such as Claude Code and Codex, require the use of the computer's Terminal window and can intimidate users without a technical background.

This post has two goals:

1. Help you get started with Claude Code if you are a non-technical user.

2. Boost your motivation to learn by offering a solution to Spotify’s most frustrating feature: the shuffle mode that’s not random.

Subscribe now

Problem to solve: Spotify’s shuffle mode

Spotify’s shuffle mode does not feel random at all. If you have 1000+ tracks in a playlist, you have probably experienced the frustration of hearing some tracks play dramatically more often than others. Spotify denies it, and the internet doesn’t believe them.

So, here is your motivation for learning to use Claude Code: we are going to build an app that takes all your liked tracks, shuffles them, and stores them into a new playlist. This way, you can listen to the newly created playlist from beginning to end, and that means you'll be listening to your liked tracks in a truly random order.

How to use Claude Code

In this tutorial, we choose to provide precise instructions to Claude Code instead of simply writing, “create an app that does this and that.”

Both approaches are possible, but high-level vibe coding runs the risk of taking the coding assistant into dead ends. By using precise instructions, we can ensure that the app’s architecture is logical and that the code remains stable as the app expands.

As a user of coding agents, your mindset should be that of a product manager rather than a business user: you don’t need to write code, but you need to have a clear understanding of the workflows, data structures, and third-party integrations that your project requires, and you need to be able to communicate them clearly to coworkers (in this case, the agent).

Prerequisites

In this post, we assume that you want to be able to use Claude Code on your laptop, and that you are using an Apple computer (running macOS).

Developing on your laptop is different from using a web-based coding platform like Lovable, Bolt.new, or v0, which takes care of all the underlying setup for you.

Even though coding assistants have made coding much easier, an initial configuration is required to enable your laptop to support software development if it’s your first time doing it. You only need to do this once.

You also need to register for third-party services such as Spotify and Anthropic’s Claude. Let’s go through each prerequisite.

Spotify developer account

In your browser, go to spotify.com and log in with your regular account.

Then, go to https://developer.spotify.com/, click on your user ID in the top right corner, and go to the dashboards (after accepting some terms and conditions). Create an app, using http://127.0.0.1:8080/callback as the redirect URI.

After creating the app, save the Client ID and Client Secret somewhere for future reference. In the “User Management” tab, add yourself as a development user (with your full name and email address associated with your Spotify account).

Code editor (IDE): Cursor

Install Cursor by downloading it from this link.

XCode command line tools

This app is a utility required for coding on Apple devices. Open the Terminal application on your laptop and execute the following command:

xcode-select --install

Package manager: Homebrew

Install Homebrew by following the instructions at this link: open the Terminal application on your laptop and execute the following command (on a single line):

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Python

Restart the Terminal. Execute the following commands to install Python, a popular programming language:

brew install xz
brew install readline
brew install pyenv
brew install pyenv-virtualenv

echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.zshrc

echo '[[ -d $PYENV_ROOT/bin ]] && export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.zshrc

echo 'eval "$(pyenv init - zsh)"' >> ~/.zshrc

Restart the Terminal. Then:

pyenv install 3.11
pyenv global 3.11

JavaScript Node (NodeJS)

Restart the Terminal. Execute the following commands to install JavaScript Node, another popular language that’s required to run Claude Code. (See official instructions at this link).

Execute this command on a single line:

curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash

Restart the Terminal. Then:

nvm install 22
node -v
nvm current
npm -v

Claude Code

Go to claude.ai and create an account if you don’t already have one. To use Claude Code, you need to pay for a Pro subscription ($20/month).

Restart the Terminal. Execute the following command to install Claude Code. (See official instructions here).

npm install -g @anthropic-ai/claude-code 

Project folder

You need to create a project folder on your laptop. For example, using the Mac’s Finder app, create a folder called ‘spotify-shuffler’ in the Documents folder.

Then:

• Open Cursor and open the folder that you just created.

• Open Terminal and type cd ~/Documents/spotify-shuffler to move to the project folder.

In the Terminal, create a Python virtual environment, using Python version 3.11, which will isolate your work from the rest of your laptop.

pyenv virtualenv 3.11 spotify-shuffler-1

pyenv local spotify-shuffler-1

Install the libraries that you’ll need for the app: although you could ask Claude Code to install dependencies for you, it’s generally safer and more effective to decide which ones you’ll use.

Using Cursor, create a requirements.txt file with the following contents:

python-dotenv
requests
flask

Then, install the dependencies:

pip install -r requirements.txt

Now, let’s start Claude Code:

nvm use 22
claude
/init

Select basic settings (such as dark mode, standard settings) and log in with your Claude Pro subscription. Press Enter until Claude Code is running.

Arrange the two windows side by side. Now, your screen looks like this:

In Cursor:

• The .python-version file contains the name of the virtual environment.

• The CLAUDE.md file contains documentation used by Claude Code to track the project's workflow.

• The requirements.txt file contains the dependencies that we have defined.

Let’s start coding

Now, we are ready to create code by giving instructions to Claude in the Terminal window.

A key working principle of this post is that Claude Code must be guided precisely if we want to control the quality of the output. This is the difference between “vibe-coding” for fun and using Claude Code to create quality software.

Each of the steps below describes a specific task that leverages Claude Code to get work done. Usually, after each instruction, Claude Code thinks for a while and asks you to validate code changes. Review these changes and click 'Yes' if Claude’s messages appear to make sense.

Inform Claude about what we’ve already done

Instruct Claude:

The name of the local virtual environment (created with pyenv virtual env) is `spotify-shuffler-1`. It is activated with pyenv local. The following dependencies have already been included in the requirements.txt file and installed: python-dotenv, requests, flask. Update CLAUDE.md accordingly.

Save the Spotify credentials in the project folder

You need to execute this task yourself. Using Cursor, create a .env file and input the Spotify Client ID and Client Secret in this format:

SPOTIFY_CLIENT_ID=xxx...
SPOTIFY_CLIENT_SECRET=xxx...

Then, save the file:

Create a web app to log into Spotify and store credentials

Instruct Claude:

Create a Flask app called `login.py`, which opens a local web page at http://127.0.0.1:8080 and allows the user to log in to their Spotify account. It then obtains access tokens and refresh tokens and saves them to a local file (spotify_tokens.json).

Ensure that you use the official Spotify API documentation at https://developer.spotify.com/documentation/web-api, specifically the "Authorization Code Flow" for authorization.

Note that the redirect URI declared to Spotify is http://127.0.0.1:8080/callback, as localhost is no longer accepted by Spotify.

Note that the Spotify app credentials are already stored in the .env file, which can be loaded with python-dotenv. The relevant environment variables are SPOTIFY_CLIENT_ID and SPOTIFY_CLIENT_SECRET.

The login flow must allow the app to use the following scopes:
Users
* user-read-email
* user-read-private
* user-personalized
Playlists
* playlist-read-private
* playlist-read-collaborative
* playlist-modify-private
* playlist-modify-public
Library
* user-library-modify
* user-library-read

Comment your code to make it understandable by a junior analyst.
Print logs so that I can see the script's progress.
Once you are done, update CLAUDE.md accordingly.

When Claude is done, open a new Terminal tab, and execute the app:

cd ~/Documents/spotify-shuffler
pyenv local
python login.py

In your browser, visit http://127.0.0.1:8080 and log in with your Spotify credentials.

Once you are done, you can enter Ctrl-C in the Terminal tab to stop the server.

Create a script that collects your Liked tracks

Instruct Claude:

Now, we will create a series of Python backend scripts, which aim to read my Liked tracks, shuffle them, and save the resulting track list to a new Playlist.

For now, let's focus on the first script, which will be called read-tracks.py.

This script should simply read my Liked tracks and save them into a local file called `my-liked-tracks.json`. It uses the requests library to make API calls.

The `my-liked-tracks.json` file will contain a list of track objects.

Ensure that you use the official Spotify API documentation at https://developer.spotify.com/documentation/web-api, where Liked tracks are called Saved tracks (see the relevant API docs here: https://developer.spotify.com/documentation/web-api/reference/get-users-saved-tracks).

Note that I have several thousand Liked tracks. Given that Spotify returns 50 tracks at most by API call, and that the rate limits are unknown but implemented for 30-second windows, you must space out API calls (say, every 10 seconds), and manage the situation where the API call may return a 429 error so that you can pause and continue.

Remember that the Spotify access tokens are saved in spotify_tokens.json.

Comment your code to make it understandable by a junior analyst.

Print logs so that I can see the script's progress every time an API call is made or an error is detected and handled.

Once you are done, update CLAUDE.md accordingly.

In the other Terminal window, execute the script:

python read-tracks.py

The output will likely look like this:

Using Cursor, you should be able to see the list of tracks in the my-liked-tracks.json file.

Create a script that shuffles tracks and creates a new playlist

Instruct Claude:

Now, let's create another Python backend script, called shuffle-tracks.py.

This script:
* Creates a new private playlist, called shuffle-xxxx where xxxx are 4 random digits (refer to the API doc here: https://developer.spotify.com/documentation/web-api/reference/create-playlist)
* Reads my Liked tracks which are already saved locally under my-liked-tracks.json as a list of track objects. Check the structure of that file by reading the first 100 lines.
* Reorders these tracks randomly to create a new list of track URIs (the track URI is the uri field of the track object).
* Adds these reordered track URIs to the newly created playlist (refer to the API doc here: https://developer.spotify.com/documentation/web-api/reference/add-tracks-to-playlist)

Note that a maximum of 100 tracks can be added to a Playlist in one request, but I have several thousand Liked tracks. Spotify implements rate limits for the API. The rate limits are unknown but are implemented in 30-second windows. Therefore, you must space out API calls (e.g., every 10 seconds) and manage situations where the API call may return a 429 error, allowing you to pause and continue.

Remember that the Spotify access tokens are saved in spotify_tokens.json.

Comment your code to make it understandable by a junior analyst.
Print logs so that I can see the script's progress every time an API call is made or an error is detected and handled.
Once you are done, update CLAUDE.md accordingly.

In the other Terminal window, execute the script:

python shuffle-tracks.py

The output will likely look like this:

Congratulations! You can find your newly created playlist in your Spotify app:

Create a script that aggregates all previous scripts

Instruct Claude:

Python backend script, called main.py.

This script will:
1. Check if the Spotify access tokens saved in spotify_tokens.json are valid, and refresh them if they need to be refreshed, or instruct the user to run `python login.py` if the user must log in again.
2. Pause to print out what the script is about to do, and require the user to confirm with Yes/No (Y/N).
3. Execute the two following scripts in sequence: read-tracks.py and shuffle-tracks.py
3. Print a final message to inform the user that the script is completed.

Comment your code to make it understandable by a junior analyst.
Print logs so that I can see the script's progress.
Once you are done, update CLAUDE.md accordingly.

Next time that you need to run the app, you can execute the following commands:

cd ~/Documents/spotify-shuffler
pyenv local
python main.py

You are done! Unless you wish to use Claude Code to customize the app further. For example, you could create other scripts that shuffle existing playlists, instead of your Liked tracks.

Takeaway messages

• Coding agents like Claude Code work best with precise instructions rather than vague "vibe-coding" requests to produce quality, stable software. Think like a product manager who clearly communicates workflows and requirements.

• The initial technical setup (Terminal, Python, Node.js) is painful, but it’s a one-time investment that unlocks powerful development capabilities.

• Real-world problems make excellent learning motivators. To learn how to use coding assistants, work on personal or professional mini-projects that solve everyday frustrations.

How do they compare with best practices? Are they being adopted by 2%, 20% or 80% of the organization?

The GenAI Maturity Scorecard

The GenAI Maturity Scorecard is a structured report card that measures the progress made by the organization as a whole, as well as each of its departments, in terms of GenAI adoption.

More importantly, the Scorecard can be updated over time, typically on a quarterly basis. This enables the CEO to access a bird's-eye view of the actual progress that has been made.

Typically, the Scorecard is structured around 4 areas:

• Governance and organization. The dimensions associated with this area do not vary much by department. They include:

  • Governance

  • Organization

  • Infrastructure

  • Risk management

• Productivity tools. The dimensions associated with this area do not vary much by department either. They include:

  • Individual productivity tools

  • Team productivity tools

• Business operations. This area focuses on the application of GenAI for internal process optimization and automation. Its dimensions vary by department. For example, for the software engineering department, the dimensions may include:

  • Requirements and planning

  • Code generation and DevOps

  • Documentation and knowledge transfer

  • Testing & QA

• Customer experience. This area encompasses initiatives that impact customer onboarding, product delivery, and customer support. Its dimensions vary by business. For example, dimensions may include:

  • Customer onboarding

  • User interface

  • Customer success

  • Customer support

The Scorecard is not purely technical, since it does not focus solely on whether a specific GenAI use case is implemented, such as software development agents or customer chatbots. It is heavily dependent on the level of adoption of use cases across the organization.

For example, if Claude Code is used for agentic software development (Level 4), but only 10% of the team uses it, while 90% of the team merely uses code completion in VS Code (Level 1), then the weighted average score of the team is 4x0.1 + 1x0.9 = 1.3.

The following diagram illustrates this approach.

Obviously, the Scorecard is not the only way to look at GenAI initiatives. It can be complemented by:

• Project management dashboards, to track the progress of each initiative.

• Business dashboards, to track the impact of each initiative in terms of additional sales leads, increased productivity, higher customer satisfaction, or any other metric.

Defining scorecard dimensions and levels

Scorecard dimensions and levels must be customized according to the company's business and key processes. The following paragraphs provide a starting point.

Governance and organization

This area measures the level of adoption of GenAI best practices in terms of governance, organization, infrastructure, and risk management. Below is a possible starting point.

Productivity tools

This area measures the level of adoption of commercially available productivity tools, usually provided to the team as per-user subscriptions.

Such tools include personal chatbots, meeting transcription and summarization tools, content generation tools for the marketing & sales team, and coding assistants for the software engineering team. Below are some examples.

Here's a possible starting point for level definition.

Business operations

This area focuses on the application of GenAI for internal process optimization and automation, using either commercially available software integrated into day-to-day workflows, no-code platforms, or custom software. Its dimensions vary by department. For example, for the software engineering department, the dimensions may include:

Specific dimensions vary for the marketing team, the sales team, the finance team, the HR team, and so on.

Customer experience

This area encompasses initiatives that impact customer onboarding, product delivery, and customer support. Its dimensions vary widely by industry and company.

For example:

• Level 1 may involve the deployment of a chatbot for customer support (many technology companies, telecoms, and airlines already have this).

• Level 2 could entail the deployment of a chat copilot to help users in their actual use of the company's product (e.g., Microsoft Office Copilot).

• Level 3 often enables users to easily deploy automations within the product (e.g., Microsoft Azure AI Studio).

• Level 4 typically involves redesigning products from the ground up with an AI-first perspective (e.g., how Lovable reinvents the front-end development and deployment process centered on the chat interface, in contrast with how app platforms like Google Cloud App Engine or Digital Ocean start with repositories or Docker images).

Measuring scorecard levels

Once the areas and dimensions are defined, establishing levels and scoring team members usually requires the involvement of experts who have seen GenAI implementations at other companies (particularly tech companies) and can probe the organization through interviews to assess the actual level of GenAI adoption.

The investigation can involve:

• Workshops

• 1:1 interviews

• Online team surveys

Regardless of how expert the evaluators are, the first scorecards will incorporate some level of subjective judgment. What's more important than the actual scores is their evolution over time. Quarterly scorecard updates are essential in tracking the organization's progress.

Takeaway messages

• Companies need a systematic way to assess whether their AI initiatives are moving the needle beyond isolated pilot projects. The GenAI Maturity Scorecard offers a structured framework for measuring adoption.

• The scorecard's value lies in measuring what percentage of teams use AI tools, not just whether the tools exist. Importantly, the scorecard should be updated quarterly to reveal adoption trends.

• 4 assessment areas can be considered: Governance & Organization (company-wide practices), Productivity Tools (individual and team subscriptions), Business Operations (department-specific process automation), and Customer Experience (external-facing AI applications).

• Scorecard dimensions and levels must be tailored to your company's industry, processes, and strategic priorities.

Share

AI has been used as a decision-making tool for decades, resulting in breakthrough use cases such as spam filtering, fraud detection, and, more recently, self-driving cars and semi-autonomous robots.

Generative AI, which gained mainstream popularity in 2022 with the release of ChatGPT and Stable Diffusion, is a new type of AI that encompasses models capable of mimicking human behavior by creating content (text, images, audio, and video) in response to user prompts.

GenAI is made possible by several concurrent technological advances: novel training algorithms based on the Transformer architecture (2017), the broad availability of powerful chips, and the digital availability of vast human knowledge on the internet. These advances underpin foundation models, which can be further trained, fine-tuned, orchestrated, and prompted to achieve specific objectives.

Using GenAI, businesses can automate an increasing number of human activities, particularly those that do not require substantial interaction with the physical world. Companies can leverage GenAI to deliver better value to their customers at a lower cost. If they don't, they risk being displaced by AI-first competitors.

Most of the business value from GenAI initiatives is likely to come from software engineering and IT, customer operations/customer service, sales and marketing (along the entire customer acquisition and retention funnel), and product development operations (especially product documentation). Additional opportunities can be found in supply chain, procurement, risk management, and compliance.

That said, the technology is evolving rapidly, and many applications are not mature yet. GenAI transformation is a journey rather than a blueprint.

For businesses, especially mid-sized companies with limited resources, this means finding a balance between launching short-term experiments and building long-term internal capabilities that deliver business value.

In this blog post, we discuss a possible roadmap for GenAI adoption and highlight some key challenges to watch out for.

Roadmap overview

The GenAI adoption roadmap of a company consists of 3 steps, which may overlap with each other:

• Individual productivity: This step involves employees adopting commercial GenAI apps to complete office tasks more efficiently. The tasks and tools are not company-specific: they can include image and video generation for marketing campaigns, coding, and online research.

• Internal process optimization and automation: This step involves using GenAI tools to streamline internal administrative processes. GenAI can be used for data clean-up and transformation, decision support, and automated execution. It is important to always simplify processes before automating them.

• Product and process innovation: This last step introduces transformative changes to a company’s products, services, and operating model. It requires rethinking the business from a blank slate, like an AI-first competitor would do. It may involve proprietary data or models leveraging the company’s unique expertise, but this is optional.

Many experiments and projects are typically launched as part of the above steps. Not all of them work out. Sometimes this is because recurring costs are too high, but more often, it is because the value proposition does not drive sufficient user adoption.

The company should learn from every initiative and build long-term knowledge and capabilities that can be leveraged across projects. Such capabilities include: chatbots, AI assistants, AI agents, observability and evaluations, APIs to data and third-party services, cloud infrastructure, and team competencies.

This does not mean, however, that every prototype must be built on top of production capabilities. Projects often begin with an ad hoc stack, subsequently transitioning to a more robust infrastructure as they move into production, as illustrated in the diagram below.

Let us review each step of the roadmap in more detail.

Individual productivity

This first step requires:

• Training staff about AI technology, opportunities, and tools.

• Trialing many commercial GenAI tools to settle on a core set of app subscriptions that best match the company’s requirements.

The range of tools to be evaluated includes chatbots, meeting assistants, coding assistants, text generators, image and video generators, voice and music generators, web scraping tools, and other tools as needed.

Internal process optimization and automation

This step requires:

• Initial brainstorming. As outlined in a previous post, the question "How would we use cheap interns?" can be a good starting point.

• Top-down tool prototyping and deployment. The reason why this usually starts top-down is that teams may lack both the bandwidth and the competencies to deliver meaningful projects organically.

• Killing projects that are not used every day. Many projects won’t garner sufficient user adoption even if they look good on paper. Better to terminate them and move on.

It often makes sense to prototype internal process automation tools using no-code or low-code platforms, but teams must not shy away from writing code, given that vibe coding assistants make coding easier than ever. Zapier + Cursor + Digital Ocean is often a great choice, but for more options, read this previous post.

Product and process innovation

This step requires:

• Use case brainstorming and prioritization

• Prototyping. Often, this is done by a small team of AI-savvy employees.

• Handover to production teams. For projects to work at scale, they must be absorbed into the relevant product or function.

The diagram below illustrates a possible framework for brainstorming.

Production deployments present numerous challenges that are often underestimated by influencers who post viral use cases on social media platforms.

Challenges include:

• Use case selection: When use cases show low adoption, is it a user education problem, or is the business value just not big enough for now? When to kill projects?

• Data quality: How much time is it worth spending on upfront data clean-up?

• Cost control: How to launch apps at a limited scale before costs and infrastructure requirements start to escalate?

• Team buy-in: What's the right balance between top-down push and organic innovation?

• External dependencies: What's the right balance between in-house development and reliance on immature third-party service providers?

• Long-tail risks: How to manage customer and reputation risks (e.g., hallucinations, tone-deaf outputs, jailbreak, leakage of private information)?

Long-tail risks, in particular, require both an observability and evaluation framework, as well as significant testing (all of which can be assisted by AI), which can add significantly to the project timeline if they aren’t anticipated.

Takeaway messages

GenAI transformation requires a structured approach that balances experimentation with sustainable capability building. Companies should start with individual productivity gains, progress through internal automation, and gradually move toward product innovation. Success depends on learning from failed experiments, building reusable infrastructure, and maintaining realistic expectations about adoption timelines. The journey is iterative—not every prototype will succeed, but each attempt should contribute to organizational AI literacy and technical foundations that enable future breakthroughs.

Subscribe now

Share

There are dozens of new AI-powered no-code platforms and thousands of YouTube tutorials that promise AI bliss. However, the industry is still immature, which means that if you have 10 workflows to automate, you may need to subscribe to 5 different SaaS platforms, each with their own supported functionalities, and learn how to use each of them.

Additionally, in a recent blog post, I suggested that the rapid development of AI coding tools means that it may not make sense to limit yourself to no-code tools, even if you are a novice coder. Given that AI can help to write code for you, you should consider low-code or even full-code, too.

This blog post assumes that:

• You are a novice coder.

• You have many disparate business workflows to automate that require customizations not currently supported by no-code platforms.

After reading this post:

• You will be equipped with a general-purpose approach to automate your business workflows with generative AI.

• You will have activated a small number of tools that are sufficiently flexible to support a wide range of workflows.

• You will have created a custom AI agent using a framework that can be easily extended to many other use cases.

• You will have experienced "vibe coding", perhaps for the first time!

General-purpose approach to automate business workflows

Business automation starts with a founder's mindset. In his book about Elon Musk, Walter Isaacson describes Musk's 5-step algorithm for optimizing business productivity. The bottom line: hands-on entrepreneurs must experience and optimize processes manually before they start automating them. Automation comes last.

This rule applies to AI automation. A business process is only worth automating after you've experienced it enough to be able to teach its optimized version to an intern. 

The approach to automate business workflows starts from the bottom up and is a simple 4-step process:

1. Identify a repetitive workflow that you'd like to automate.

2. Perform the workflow manually enough times to ensure you are clear about the inputs, outputs, and exceptions. You must be clear enough about the process that you should be able to teach it to an intern.

3. Automate the workflow using the toolkit described below in this post.

4. Move on to the following workflow.

The actual workflows depend on your specific business needs. You can refer to this previous post for more guidance on identifying them: AI as the Intern.

Here are some common examples:

• Go-to-market:

  • Generation of content ideas and production of social media content.

  • Newsletter generation.

  • Marketing lead generation.

  • Monitoring of competitor and influencer presence on social media.

  • Data clean-up and population for new marketing leads.

  • CRM logging and summarization of interactions between customers and the sales team.

• Operations:

  • Data extraction from documents in non-standardized formats.

  • Standard operating process enforcement: verify that all criteria are met before advancing a dossier, application, or document to the next step.

  • Conversion of meeting recordings into summaries and assignment of next steps/tasks to team members.

• Administrative:

  • Employee request approvals and tracking.

  • Preparation of employee performance reviews.

  • Business expense claims.

  • Supplier invoice processing.

Recommended toolkit

If your workflows are not supported out of the box by no-code platforms, you must become comfortable with the idea that some coding will be required. 

You don't need to train to work as a software developer, but you need to learn enough to guide and utilize AI code generation tools, such as Cursor, Replit, Deepnote, Windsurf, and GitHub Copilot, to implement your projects. If code intimidates you, you need to overcome that situation. This tutorial will help you.

Here is the recommended toolkit to get started with workflow automation:

We will use the following paid tools:

• Airtable. AI agents must be able to access a database to read inputs and save outputs. Airtable is a great starting point because it offers an excellent user interface and an easy-to-use API. Visit Airtable to create an account and subscribe to the most affordable plan ("Team"). Create a new base ("My First AI Agent") and visit the Builder Hub to create an API token ("aiforbiznews") with read and write access to that base. Copy the API token; you will need it later. Also, copy the Base ID, that's the string that starts with "app" in the base's URL (for example, appnoVp4rupoSPJFE).

• Zapier. Zapier is a no-code service that many entrepreneurs use for automation due to its extensive number of third-party integrations. We will utilize Zapier to seamlessly integrate our AI agents with external services; however, the AI agents themselves will not run within Zapier. Visit Zapier to create an account.

• Replit. Replit is where we will generate and run all our custom code. It is a web-based coding app that offers excellent code generation and completion functionalities. Visit Replit to create an account and subscribe to the cheapest plan ("Core"). To use Replit well, you need to learn the basics of Python. If you have never coded before, start with the 5-hour Coursera course "Programming for Everybody (Getting Started with Python)".

• OpenAI. This API service offered by ChatGPT's creator provides access to the latest AI models. Visit OpenAI to create an account, purchase a few U.S. dollars of credits, and obtain an API token at https://platform.openai.com/api-keys (copy it, as you will need it later).

Let's create a custom AI agent

To familiarize ourselves with the recommended toolkit, let's create a simple address book management agent.

It works like this: whenever you receive an email from a contact that includes a signature, you can forward the email to the AI agent, which will read its content to create or update your contact database. The AI agent should also be able to process business card scans, which are also sent to it via email.

This type of AI agent is a good illustration of business automation, as many company processes require receiving information from external sources and reformatting it to update company databases. In a world where AI is available, you'd rather just send the information to your AI agent and let it deal with it.

Manual workflow

Before automating the workflow, we must implement it manually in Airtable and document the process.

Create two tables in your base:

• The inputs table is where you store incoming data, either in the form of an email body or an email attachment. The inputs table stores the status of each input as well: an input will be retrieved and processed by the AI agent if its status is “Todo”, after which the AI agent will change its status to “Done”.

• The contacts table is where you store your address book, with fields such as Full Name, Job Title, Company, Email, Phone, etc.

Play around with Airtable to create tables that look like this:

Then, document your manual workflow. The workflow looks like this:

1. Store the email body and attachment (business card) in Airtable.

2. Read the email body and attachment to identify any contact information, and organize it into the relevant fields such as Full Name, Job Title, etc.

3. Verify existing contacts against the Full Name of each identified contact.

4. If a contact already exists, update it.

5. If a contact does not exist, create a new one.

Store email contents into Airtable's “inputs” table

This automation step can be implemented in Zapier using a 2-step workflow that retrieves emails from an email address assigned to you by Zapier and saves their content into Airtable. Don't forget to publish the "Zap".

Refer to the screenshots belo

“Vibe code” the AI agent in Replit

Vibe coding means generating code by giving instructions to an AI coding agent in natural language. It is generally recommended to give instructions bit by bit, so that you can check the code at every step in case the AI hallucinates.

Follow the instructions below.

Start by typing the following instruction in your Replit home page: "Create a Python script that runs continuously with pauses of 15 seconds between runs."

Replit creates an app that looks like this:

Then, you can proceed with the following step-by-step instructions and answer the questions asked by the Replit agent. Along the way, the Replit agent will ask you to provide the API keys. Here are the step-by-step prompts.

First, execute_task should read an Airtable base and collect all records with Status = "Todo". The base_id and table_name will be stored by the ContinuousScript constructor as instance variables. Save the records in the records variable.

For each record in records, parse the content of the "Email Body" field, which is a text field. Leverage OpenAI's API with the model o4-mini and ask the AI assistant to extract all contact information contained in the email body. The result should be saved as a JSON string that contains a list of objects, where each object contains the following keys: Full Name, Job Title, Company, Email, Phone, and Address. If any value is not provided, the value should be an empty string. Save the resulting JSON string into a variable called contact_details.

For each contact in contact_details, check if the Full Name already exists in the "contacts" table of the same Airtable base. If it exists, update the record with the non-empty values of the contact detail. If it does not exist, create a new Airtable record.

In execute_task, the Airtable record may also contain attachments in the "Email Attachment" field. If that is the case, use OpenAI to analyze any attachment that looks like a business card, and extract all contact information from the business card. Return the result as a JSON array containing objects with these exact keys: "Full Name", "Job Title", "Company", "Email", "Phone", "Address". Like the information from the email_body, it should then be synced to the "contacts" Airtable table.

For each record of the "inputs" Airtable table that has been processed, finally update their status to "Done" in the Status field.

This is working, now deploy the app.

In Replit, you should end up with something like the screenshot below. For your reference, the full code is published at this link.

If Replit returns deployment errors, use the Assistant to debug them. For example, in this case, the Assistant suggests changing the deployment type to "background worker".

At this point, your AI Agent is live! If there are bugs, you may need to fix them manually or request that the Replit Agent address them.

Here is the address book generated by the AI agent:

Takeaway messages

You are now equipped with a general-purpose approach and tech stack to automate your business workflows, using Zapier, Replit, Airtable, and OpenAI.

Using vibe coding, you can take it a step further and automate a wide variety of custom workflows, even if they are not supported out of the box by no-code platforms.

Subscribe now

Share

Given this level of user adoption, enterprises have started deploying internal chatbots to converse with employees too. Internal chatbots are easier to implement than external ones, due to higher tolerance for errors and tone-deaf comments.

Among internal use cases, employee training is one of the most promising areas where AI-powered chatbots are already good enough for large-scale use, not to replace remote training sessions, but to complement them.

Let’s examine how to use GenAI for employee training and how to get started. A demo with code is provided in the post.

Online training: what’s already working and what can be improved?

Most mid-sized and large companies already offer online employee training programs, either created internally or obtained from MOOC vendors.

Such online trainings are usually delivered by humans, either in real-time or from pre-recorded sessions. There is, in my view, no need to change that, and you should continue to look for trainers who know their subject, have a vast repository of “war stories” to share with the trainees, and bring a combination of energy, gravitas, and plain-spokenness.

A great way to see how humans can excel at training delivery is to watch Andrew Ng’s Machine Learning Specialization online course.

In short, humans can already deliver great training at scale. One of their main flaws, which is that they don’t speak every language on Earth, can easily be addressed by using GenAI-powered dubbing or subtitle platforms like ElevenLabs and Descript.

However, there are two areas where online training programs show clear improvement areas, where GenAI can make a difference:

1. Simulations (role plays). Most training programs include quizzes to test the trainee’s knowledge, but these tests lack interactivity. Ideally, you’d want the trainee to experience realistic role-plays and receive fine-grained feedback during sales training or a course on managing team performance.

2. Q&A (community forums). Online programs usually invite trainees to join an online forum to ask questions. Even with proper staffing, it takes hours for a trainee to get an answer, interrupting their flow.

The AI-powered employee training assistant

The AI-powered employee training assistant includes three complementary components:

1. Training delivery, which can still be handled by a talented human.

2. Role-play program, including exercises immediately after the training but also at regular intervals (e.g., 1, 3, 12 months). What is important is not just the role-play but also the personalized feedback.

3. Online community, which can be hosted on any discussion platform (e.g., Slack) and can be staffed by a Q&A assistant that answers questions immediately using training materials, a FAQ repository, and past discussions. In addition, a training feedback assistant should listen to trainee discussions to understand and summarize feedback about what aspects of the training can be enhanced.

The diagram below outlines these components.

Role-play assistant

The role-play assistant is a chatbot that participates in role-play sessions to complement the employee training session.

For example, the role-play assistant could play the role of a buyer so that a salesperson can practice their sales conversation skills.

To generate responses to the salesperson’s messages, the assistant can leverage general instructions provided to it, as well as a knowledge base of past buyer messages.

The following diagram describes the assistant’s architecture.

The chatbot can be text-based or voice-based.

To implement a voice-based assistant, you can use OpenAI, Google Gemini, or Assembly AI APIs to transcribe the salesperson’s voice into text. Then, you can use a voice generator like ElevenLabs to convert the assistant’s response into a spoken message.

Here is an example of a prompt to guide the assistant.

You are participating in fictitious role-play that is used to train new salespersons at company TOPAGENCY.
You, the assistant, are playing the role of Roberta, the BUYER, who is the prospective customer in a sales conversation.
The user is playing the role of the SALESPERSON.
Throughout the conversation, you answer the SALESPERSON's questions and ask questions of your own to role-play the scenario below.
Here are the scenario guidelines.

1. About the BUYER

The BUYER is the marketing director of a cryptocurrency project called NEWBLOCKCHAIN.
The BUYER is looking to engage a graphic design agency to work on social media visuals, web banners, physical posters and video animations to promote NEWBLOCKCHAIN.
The BUYER's key decision creteria are: experience in the cryptocurrency industry, track-record of coming up with original ideas, speed of delivery, and budget.
The BUYER's budget is between 500 USD and 1000 USD per month, but they are not keen to disclose it initially to the SALESPERSON in order to get the best deal.
Regarding speed of delivery, the BUYER expects 2 business day turnaround on images and 72h turnaround on videos.
The BUYER is not keen to work with agencies that have been in operation for less than two years, and the BUYER thinks that a company size of at least 25 people is required to deliver against turnaround timelines.

2. Goals of the conversation

From the BUYER's point of view, the most important goal is to avoid wasting time with unqualified vendors, given that the BUYER receives lots of messages from prospective vendors.
The BUYER wants to determine quickly if TOPAGENCY is likely to meet the selection criteria, particularly in terms of cryptocurrency industry experience and track-record.
As the BUYER receives lots of proposals, they don't feel like prying the information out of the SALESPERSON.
Rather, they expect the SALESPERSON's pitch to be good enough that it makes it worthwhile to keep the conversation going.
Once the BUYER has established whether TOPAGENCY is a serious vendor, they can decide to either continue or terminate the conversation.
If the conversation continues, the BUYER's secondary goal is to provide enough information to TOPAGENCY so that they can prepare a proposal and price quotation.

3. Other guidelines

The conversation is happening by direct message over telegram.
We expect the BUYERS to send no more than 30 messages, given that they are busy.
The BUYER's messages tend to be short and sometimes include typos, given that the BUYER is busy.
To obtain examples of past BUYER messages, the assistant may call the file_search function, which returns examples of conversations between other buyers (who may have different needs than this BUYER) and other TOPAGENCY salespersons.

You can find the full demo and code at this link for details:

Visit the notebook.

Feedback assistant

The feedback assistant listens to real or simulated conversations completed by the trainees practicing the skills taught by the training program. Based on these transcripts, general instructions, and the training materials and FAQs, the assistant generates personalized feedback for the trainee.

The following diagram describes the assistant’s architecture.

Here is an example of a prompt to guide the assistant.

As an expert in business to business sales, please evaluate the following conversation between a BUYER and a SALESPERSON, and provide feedback to the SALESPERSON about what they did well and what they can improve.
The conversation is provided between the <conversation> tags below.
Be specific in your feedback, quoting specific examples of what they did well and what they can improve, and giving examples of what the SALESPERSON could have said.

You are provided some context about the role-play and the salesperson, between the <context> tags below.

<context>
The SALESPERSON works for TOPAGENCY, a graphic design agency that employs graphic designers to create images and video animations.
TOPAGENCY has around 40 employees around the world, working remotely. It was created 3 years ago, in 2022.
TOPAGENCY has limited experience working with customers in the cryptocurrency industry, but they have successfully established long-term working relationships with technology companies like Slack, Atlassian and Stripe.
From TOPAGENCY's point of view, the most important goal of the conversation is to capture information from the BUYER in order to prepare a proposal that has a high chance of success.
The SALESPERSON must understand what the BUYER is looking for, how many designs monthly they will require, and what is their approximate budget.
They must also make the BUYER agree that the next step is to discuss a proposal from TOPAGENCY.
</context>

The desired SALESPERSON behaviors are listed below, between the <desired_salesperson_behaviors> tags.

<desired_salesperson_behaviors>
* Build rapport authentically: Start with brief small talk before transitioning to business, but find genuine connection points rather than forced conversation.
* Listen more than you speak: Aim for an 80/20 ratio (them speaking 80%).
* Focus on problems before solutions: Uncover pain points and business impact before discussing your product.
* Ask strategic questions: "What challenges are you facing with your current solution?" "What would success look like for this project?"
* Demonstrate industry knowledge: Show you understand their sector's unique challenges without being generic.
* Present value, not features: Connect your solution's capabilities directly to their specific business outcomes.
* Address objections with curiosity: "That's interesting, can you tell me more about your concern?"
</desired_salesperson_behaviors>

Now, here is the conversation:

<conversation>
{role_play_conversation_example_str}
</conversation>

You can find the full demo and code at this link for details:

Visit the notebook (same notebook as above).

Trainee Q&A assistant

The trainee Q&A assistant is connected to the platform used for online communication between trainees and trainers (e.g., Slack). It listens to trainee questions and concerns and answers promptly based on training materials as well as an ever-expanding FAQ list (frequently asked questions).

The following diagram describes the assistant’s architecture.

Training feedback assistant

The training loop needs a mechanism to collect and process trainee feedback. The training feedback assistant listens to all online communications between trainees and trainers. At regular intervals (e.g., weekly), it generates detailed feedback for the trainers about improving the training, including materials, delivery, and new FAQs.

The following diagram describes the assistant’s architecture.

Takeaway messages

AI-powered employee training represents a significant untapped opportunity.

Integrating GenAI assistants for role-play, feedback, Q&A, and training improvement can enhance training effectiveness while preserving the human element.

The key is to view AI not as a replacement for human trainers, but as a tool to augment their capabilities. This leads to more engaging, personalized, and efficient learning experiences while reducing training costs.

Subscribe now

Share

You’ve either dreamed about or started experimenting with ways to save your team’s time using these no-code platforms.

Is it sensible to build on these platforms? How do you choose the right one for you and your organization? What are the pitfalls?

In this post, we will answer these questions.

A brief history of no-code

Where is this no-code trend coming from?

No code workflow automation is not new. It started with spreadsheets, especially Microsoft Excel, which was launched in 1985. Excel is one of the most widely adopted business software on the planet, with around 1B monthly users.

The rise of APIs in the 2010s led to an explosion of no and low-code workflow automation platforms with better connectivity to business applications. Over time, the space consolidated around a few leading competitors in each market segment, such as Airtable, Zapier, and Make for SMEs (small and mid-sized enterprises), and Microsoft, Salesforce, and ServiceNow for large enterprises.

In the late 2010s, the no-code space for front-end interfaces grew as well and matured with a few leading companies: Shopify, Wix, Squarespace, and Webflow in the SME segment, and Microsoft, Salesforce, OutSystems, and SAP in the large enterprise segment.

Today, no-code is a $30B/year business growing at 25%/year.

Microsoft and Salesforce lead the market with multi-billion dollar workflow automation businesses. ServiceNow makes around $3 billion annually, while OutSystems makes around $500-800 million. Shopify, Wix, and Squarespace are in the $1-2 billion range. Zapier, Dataiku, Make, and Airtable are estimated to make $300-400 million each.

However, a new generation of AI-first no-code platforms is emerging, aiming to offer unique functionalities to business users.

For workflow automation, you’ve heard of n8n, Pipedream, Gumloop, Lindy, and Retool. For front-end/app development, the most talked-about startups include Bolt.new, Lovable, Softr, UXPilot. (There is some overlap between these categories).

With this context in mind, let’s discuss what you need to know before choosing a no-code AI automation platform.

1. No-code is getting better, but you should definitely consider code

No-code platforms are leverarging AI to dramatically improve user experience.

Specifically:

1. GenAI can significantly speed up app creation. You write what you want, a chatbot builds it, and then move some boxes to refine your creation.

2. GenAI enhances app capabilities by introducing innovative functionalities such as AI agents understanding natural language and deciding actions autonomously.

At the moment, #1 is hit or miss, but #2 works well.

However, at the same time, coding is becoming easier. Applications like Cursor, Github Copilot, Windsurf, Cline, and Replit have introduced agents that generate code iteratively based on natural language instructions.

This raises the question: shouldn’t you just code instead of using a no-code platform?

If coding isn’t too intimidating for you, now is a great time to start. Consider coding instead of a no-code platform subscription.

2. If you are going for no or low-code, prefer low-code

A platform that supports code editing offers far more customization possibilities than one focused on visual app design, given that code editing is easier than ever. If you are going to choose no or low-code, prefer low-code.

If you only use no-code platforms, it’s likely that a single platform won’t meet all your needs and you’ll need to switch to implement various AI automations.

Among workflow automation platforms, Pipedream supports custom code well. n8n and Lindy have launched it in beta and need to enhance the feature.

In front-end development, Bubble, Replit, Lovable, Bolt.new and Replit are pretty good at supporting customization.

Two caveats:

• Ironically, low-code platforms currently struggle with AI-powered autocomplete in their code interfaces. To generate code, you may need to open Cursor or another app, chat with the agent, and then copy and paste the code into the low-code app. Hopefully, the user experience will improve soon.

• While code generation is easy, hosting a live app can be technically challenging. Platforms like Replit, Lovable, and Bolt.new are effective at bridging the gap between code and app deployment.

3. Make sure you understand how app usage drives subscription costs

No and low-code subscription costs can escalate quickly due to pricing models based on user count and the number of monthly actions.

For example, Softr charges front-end apps per end-user. This makes subscription costs more suited to internal company applications than external-facing ones.

As for n8n (and many other workflow automation apps), they charge by number of executions, with self-service pricing tiers limited to 10,000 monthly tasks.

If you’re building an app that automates a lot of low-value interactions, you need to consider costs carefully. For example, an app that manages hundreds or thousands of social media influencers, or an app used by your company’s vendors to submit their bills for automated processing.

In a scenario with many users and/or monthly tasks, selecting Bubble for the front-end and Replit for the back-end may be the right approach, given their pricing.

However, Bubble and Replit don’t offer the security and reliability guarantees that larger enterprises expect. If your requirements are stringent, be prepared to pay per user.

4. Compare multiple platforms to pinpoint their strengths and limitations

Each no/low-code platform represents a functionality trade-off along the following dimensions:

To understand the differentiating features of each platform, you need to ask the following questions:

• Is the platform primarily for SMEs/startups or large enterprises?

• Is the platform more focused on UI creation (front-end) or on workflow automation (back-end)?

• Tech skills required: What level of coding skills do you need to use the platform properly (no / beginner / intermediate)?

• Code customization: Can you insert or edit code to customize the automation?

• AI workflows: Does the platform facilitate the deployment of AI assistants to execute workflows and call external services in a no/low-code way?

• App creation chatbot: Is it possible to describe your requirements in natural language to get the platform to build your app?

• Pricing model: Is the platform priced by project, by user, or by task?

Let’s compare a few of the most talked-about startups:

5. Experiment before committing to a platform

It often takes weeks to realize that a key functionality of your automation isn’t supported by the no or low-code platform. While platforms tout hundreds of integrations, the devil is in the implementation details.

You need to get your hands dirty and try implementing your use case on several of them to understand the advantages and limitations of each platform.

Don’t rush to migrate your workflows when you see a new and shiny platform. The new no and low-code platforms that we are seeing now are very impressive, but chances are high that their teams will be acqui-hired by a Microsoft or Salesforce in the coming months and put to work to improve legacy products, while they stop investing in the new products.

Takeaway messages

No and low-code platforms are absolutely fantastic for prototyping ideas. If you are going to use them to automate your business workflows, there are 5 things to keep in mind before choosing a platform:

1. No-code is getting better, but you should definitely consider code.

2. If you are going for no or low-code, prefer low-code.

3. Make sure you understand how app usage drives subscription costs.

4. Compare multiple platforms to pinpoint their strengths and limitations.

5. Experiment before committing to a platform.

Subscribe now

Share

With AI-powered CRM, salespersons should be able to dump information into the CRM’s lap and let the CRM organize, format and store it. Similarly, when they need information about a company or contact, they should be able to ask a single question and get all the relevant information displayed in a neat dashboard. No more navigating drop-down menus and filling multiple form boxes.

By improving CRM workflows, companies can save hours of weekly data entry tasks.

Let’s review this approach in more detail. A demo notebook is provided at the end of the post.

The new CRM paradigm

Most CRMs already support email BCC. This means that salespersons can BCC a dedicated email address when they contact a customer, enabling the email’s content to be logged into the CRM.

However, not much more is done with the email. Salespersons must complete dozens of CRM form fields with information about the contact, the company, and the potential deal: contact name, address fields, industry sector, web links, deal details, next steps, etc.

In an ideal world, the CRM should just be able to listen into all conversations across channels (e.g., video conference, email, DM), or receive conversation dumps from the salesperson, and pre-populate all data.

The CRM should be able to:

• Automatically create contacts, companies, and interactions.

• Parse email signatures.

• Scrape social media platforms and the customer’s websites for information.

• Understand conversations to identify potential deals and next steps.

• Automatically create reminders for the sales team.

• Query financial systems to understand invoices and late payments.

• Summarize all the above data in response to salespersons’ requests.

• Proactively identify new sales opportunities when new product offerings are launched.

The target architecture looks like this:

Migrate to a new CRM?

Many AI startups have identified the need for this new paradigm and are working to develop brand new CRM platforms powered by AI.

Does it make sense to migrate to these new CRM platforms?

In the long run, it will likely make sense to migrate to platforms created with an AI-first approach, because these platforms will be able to release new features faster, at a lower cost, than CRM platforms that are bogged down by legacy infrastructure and technical debt.

However, in the short term, migrations may not make sense. They require heavy manual work that negates AI’s business benefits for a year or more and expose the company to a dependency risk over fledging vendors.

Thus, in the short term, it makes more sense to build GenAI-powered interfaces on top of existing platforms and databases, communicating with them through APIs. After delivering the business benefits, then consider a migration.

The diagram below illustrates how the GenAI-powered interface can stand on top of existing systems:

How to demo this approach

You can easily create a demo of this next-generation CRM.

The demo will:

1. Create a Telegram bot that can be added to Telegram discussion groups.

2. Listen to conversations and store them.

3. At regular intervals, summarize and organize conversations to extract relevant data and store it into a CRM database.

4. Answer questions about leads and deals.

Create a Telegram bot

To create a Telegram bot, you need to have a Telegram account and converse with the @BotFather, starting with the command /newbot. You can find many online tutorials about this. Write down the bot’s API token, you will need it later.

To make sure that the bot can listen to conversations, go to Bot Settings, Group Privacy, and Turn off.

Listen to conversations

Once the bot is created, you can invite it into Telegram groups with other people. Then you must make sure that the bot listens to messages and store them into a database. There are many ways to do this. You can ask an AI chatbot to code the bot for you or use a no-code platform such as Pipedream, N8N or Make.

To get started, I recommend using:

• Pipedream as the low-code platform, with a “New Message Updates (Instant) from Telegram” trigger, followed by a “Insert Row in Supabase” step.

• Supabase as the database to store conversations.

Once the Telegram bot listener is wired up, conversations will start appearing in the aiforbiznews-crm-conversations database.

Create a dummy conversation for development purposes

To get started, you can ask an AI chatbot to generate a fictional conversation for you. Here’s a prompt:

Create a fictional sales conversation between two characters.

Richard is the customer from company "CRYPTOPROJECT". They are interested in purchasing AI consulting services to make their technical documentation easier to access and understand.

Aaron is the key account manager, selling AI consulting services.

The conversation takes place on a messaging app like Telegram. It is a follow-up from a face-to-face meeting that Richard and Aaron had at a conference. Richard mentioned their needs briefly to Aaron, who now wants to better understand the requirements.

The conversation should include around 15 messages. It should include details about Richard's company, his address and contact information, as well as their needs, and a discussion about a potential proposal and budget.

Please provide the conversation as a CSV file. The CSV should be comma separated, with quotation marks around each text string. The columns should be message_sender as the name of the sender, and message as the content of the message. Column values should be quotation marks around each text string. Do not include anything else in your response.

You can see an extract of the conversation below:

Conversation history for chat_id C-123456789:

Message from Aaron Henderson (chat_username: aaron ) at 2025-03-01 00:00:05+00:00: Hi Richard, it's Aaron from AIEXPERT. Great meeting you at the BlockTech Conference last week! I wanted to follow up on our conversation about helping CRYPTOPROJECT with your technical documentation challenges. Do you have some time to discuss your needs in more detail?

Message from Richard Rory (chat_username: richard ) at 2025-03-01 00:00:10+00:00: Hey Aaron, thanks for reaching out. Yes, I've been meaning to follow up with you as well. We definitely need help making our technical documentation more accessible. When would be a good time to chat?

Message from Aaron Henderson (chat_username: aaron ) at 2025-03-01 00:00:15+00:00: I'm available tomorrow afternoon or Thursday morning. Alternatively, we could also discuss via messages here if that's more convenient for you right now?

Message from Richard Rory (chat_username: richard ) at 2025-03-01 00:00:20+00:00: Messages work fine for now. Let me give you a bit more context about our situation. We've built a complex blockchain platform with various APIs and SDKs, but developers are struggling to understand and implement our documentation. We need to make it more intuitive and searchable.

Message from Aaron Henderson (chat_username: aaron ) at 2025-03-01 00:00:25+00:00: Thanks for the additional context. That's certainly something we can help with. Could you share a bit more about your current documentation structure? Is it primarily API references, implementation guides, or a mix of both? Also, what formats are you currently using?

Message from Richard Rory (chat_username: richard ) at 2025-03-01 00:00:30+00:00: It's mostly a mix of API references and implementation guides. Currently hosted on GitHub with markdown files and some auto-generated API docs. The main issues are fragmentation, inconsistent terminology, and lack of practical examples. Also, finding specific information is chal

...

Extract data from conversations

The AI agent must be able to access external services like the CRM database to extract, organize and store data from conversations. This is called “tool calling” or “function calling” depending on the LLM platform (see OpenAI docs here).

To get started, you must equip the AI agent with the following functions:

• For the CRM’s contact database:

  • create_contact: create a new contact.

  • get_contact_by_id: get contact details.

  • query_contacts: retrieve existing contacts by name and company.

  • get_contact_by_chat_username: get contact details from Telegram username.

  • update_contact: update the information of a contact.

• For the CRM’s company database:

  • get_industry_sectors: get the list of industry sectors.

  • create_company: create a new company.

  • get_company_by_id: get company details.

  • get_company_by_chat_id: get company details from Telegram group ID.

  • update_company: update the information of a company.

• For the CRM’s interactions database:

  • get_interaction_types: get the list of interaction types.

  • create_interaction: log a customer interaction.

  • get_interaction_by_id: get interaction details.

  • get_interactions_by_company_id: get interactions associated with company.

• For the CRM’s deal database:

  • get_deal_stages: get the list of deal stages.

  • create_deal: create a new deal opportunity.

  • get_deal_by_id: get deal opportunity details.

  • get_deals_by_company_id: get deals associated with company.

  • update_deal: update the information of a deal.

The following prompts illustrates how you can instruct the AI agent to extract data from recent conversations and store structured data into the various CRM databases.

First, we have the system prompt:

You are a CRM Assistant for the company AIEXPERT. Your primary role is to:

1. Record information from conversations into the CRM's database tables:
   - Contacts: Store personal and professional details about external individuals (each contact has a unique contact_id). Internal salespeople are not contacts.
   - Companies: Maintain information about organizations (each company has a unique company_id)
   - Interactions: Track all communications and interactions with external individuals (each interaction has a unique interaction_id). An interaction is the summary of a conversation, not a single message. The summary should put particular emphasis on decisions agreed and next steps.
   - Deals: Manage business opportunities and sales pipeline (each deal has a unique deal_id)

2. Retrieve information from these tables in response to user queries

3. Ignore requests that are unrelated to CRM functions

Your focus is on maintaining accurate and up-to-date CRM data. When processing conversations, prioritize extracting relevant business information and relationships. When responding to queries, provide clear, data-driven answers based on the CRM database.

Then, we have the actualy conversation processing prompt:

A customer conversation, captured from a Telegram chat, is provided below between the <customer_conversation> tags.
Use the conversation to create or update the CRM database, including the contacts, companies, interactions and deals tables.
Remember that an interaction is the summary of a conversation, not a single message.
When you are done, respond with a summary of the actions you took, and the data you created or updated in the CRM database.

<customer_conversation>
{customer_conversation}
</customer_conversation>

Answer questions about leads and deals

The AI agent will be able to answer questions about leads and deals simply by calling the same tools as described above, with a simple prompt like “For the company associated with chat_id {chat_id}, who are the main contacts and what are the deals under way?."

Demo in action

You can find the full demo and code at this link:

Visit the notebook.

In the demo, you can see that the conversation analysis step uses a dozen function calls, resulting in the following response:

Here is a summary of the actions taken and data created in the CRM database based on the provided Telegram conversation:

1. Company Created:
   - Name: CRYPTOPROJECT
   - Industry Sector: Technology
   - Associated with chat_id: C-123456789

2. Contacts Created:
   - Richard Rory: 
     - Email: richard.nakamoto@cryptoproject.io
     - Phone: (415) 555-8976
     - Address: 555 Blockchain Avenue, Suite 300, San Francisco, CA 94107, USA
     - Associated with CRYPTOPROJECT
     - chat_username: richard
   - Sarah Vitalik:
     - Job Title: CTO
     - Email: sarah.vitalik@cryptoproject.io
     - Associated with CRYPTOPROJECT

3. Interaction Logged:
   - Type: Telegram
   - Contact: Richard Rory
   - Company: CRYPTOPROJECT
   - Summary: 
     - Aaron Henderson (AIEXPERT) and Richard discussed CRYPTOPROJECT’s technical documentation challenges. Details of current issues and needs were reviewed. Aaron proposed an AI-powered solution and will deliver a detailed proposal and case studies. Richard provided budget ($75,000–$100,000), contact details, and requested CTO be CC'd. Next steps include sending a tailored proposal and scheduling a review call (Monday, 2 PM Pacific).
   - Key agreements: Proposal due by Friday, review call scheduled, priorities set for phased implementation, and security/maintenance needs noted.

Furthermore, if we ask the AI agent to summarize the contacts and deals associated with a particular company, we get the following output:

The main contacts for the company "CRYPTOPROJECT" are:

1. Richard Rory
   - Email: richard.nakamoto@cryptoproject.io
   - Phone: (415) 555-8976
   - Address: 555 Blockchain Avenue, Suite 300, San Francisco, CA 94107, USA

2. Sarah Vitalik
   - Job Title: CTO
   - Email: sarah.vitalik@cryptoproject.io

Current deal under way:
- Deal Name: AI-based Technical Documentation Revamp for CRYPTOPROJECT
- Stage: Prospect
- Value: $100,000 USD

If you need more details about any of the contacts or the deal, let me know.

Takeaway messages

AI-powered CRM interfaces offer significant advantages over traditional systems. They can understand natural language and process unstructured information, eliminating the need for extensive drop-down menus and data entry forms.

These interfaces can be prototyped fairly quickly, allowing businesses to iterate their approach and get feedback from salespersons.

To get started, the AI-powered workflows should be deployed “on top” of existing databases and CRMs and communicate with them through APIs, so that no data migration is needed.

Subscribe now

Share

By deploying AI capabilities throughout the sales cycle, businesses can increase sales productivity by 50%.

This increase can be achieved by increasing funnel throughput (more qualified leads), increasing conversion rates and retention rates (personalized automation), and reducing time spent on low value activities. What’s more, by using GenAI for solution development, it’s often possible to start demonstrating impact within 3 months.

Read this post, and you’ll know exactly how to get started.

Identifying the sales pain points

The first step is to ask marketing and sales teams about improvement areas, as they usually have a good sense of where the company is leaving money on the table.

Complaints that are often mentioned include:

• Lead generation activities struggle to break through in an increasingly noisy marketplace. Whether it’s for outbound marketing campaigns, lead magnets (gated content) or account-based marketing activities, an ever-growing amount of work is needed just to maintain the company’s share of voice and lead quality.

• Lead qualification is slow and not customer-friendly. Usually, lead generation activities generate large contact lists that are handed over to sales development representatives (SDRs) for manual qualification. While there has been tremendous progress in the use of lead data aggregation platforms like LinkedIn Sales Navigator and Apollo, lead qualification invariably goes through a stage of sending follow-up communications to contacts. This stage is time consuming for salespersons and viewed as unpleasant spam by recipients.

• Salespersons spend too much time inputting data into CRM tools during the sales cycle. Key account managers spend 5-10 hours/week updating platforms like Salesforce and bringing stakeholders up to speed, or hire sales assistants to do the work.

• Product onboarding is painful for customers, hurting retention. Customers must often re-input data that they have already provided. At worst, they must undertake significant manual work to migrate their data into a new system.

• Sales/product feedback loops are imprecise and opaque. Customer feedback is essential for product managers to iterate on new features, but it is conveyed sporadically by salespersons and often biased by whatever large deal they are focused on at the moment.

GenAI quick wins

Companies can address many of these pain points by implementing AI-powered sales automation systems, and they can do it quickly by using AI for platform development.

The following diagram summarizes potential quick-wins, each of which can be prototyped in just a few weeks.

Creative GenAI tools for content generation

This AI use case is well known, thanks to the tremendous adoption of ChatGPT (400 M users and counting). Often, a small number of young, passionate prompt engineering experts emerge organically in each organization.

The main implementation challenge is to ensure that the vast majority of sales and marketing employees are trained to properly use chatbots to brainstorm new content ideas, produce high-quality copy and images, convert content into multiple formats (e.g., animated banners, video explainers, Instagram and Tik Tok clips), and even monitor key opinion leaders on social media.

AI-powered data population

A specialized AI agent can perform web searches, scrape and parse web pages on the contact company’s website, and query third-party APIs (e.g., Apollo, Proxycurl, SimilarWeb). It concatenates all that information to extract and format relevant pieces of information such as company turnover, number of employees, fund raising history, web traffic and number of published job postings.

Personalized lead follow-up

In a recent post, I talked about how to use an internal knowledge base of customer success stories to generate personalized follow-up emails based on the recipient’s industry sector and company department.

This concept can be further enriched by referring to the contact’s activity on LinkedIn or Twitter/X to personalize messages, for example.

New CRM paradigm: from structured to unstructured data entry

With AI-powered CRM, salespersons can dump information into the CRM’s lap and let it organize, format and store it. Similarly, when they need information about a company or contact, they can ask a single question and get all the relevant information displayed in a neat dashboard. No more navigating drop-down menus and filling multiple form boxes.

Companies can save hours of weekly data entry tasks by improving CRM workflows. Ideally, the AI-powered workflows should be deployed “on top” of existing databases and CRMs and communicate with them through APIs, so that no data migration is needed.

Sales co-pilot

The sales co-pilot is an AI agent that continuously thinks about what’s the best next action that the salesperson can take based on the state of their pipeline. It generates a first draft for the salesperson’s review, such as:

• An email to a customer contact.

• An outreach to other potential decision makers at the contact’s company.

• A question to an internal stakeholder, or an internal approval request.

• A progress report for the salesperson’s manager.

When deploying co-pilot agents, it’s important to keep things simple at first, to avoid swamping the sales team with distracting signals. A potential approach is to focus on creating, say, only 3 relevant draft emails per day, and measure adoption.

AI-powered customer data migration agent

An AI-powered customer data migration agent aims to reduce the amount of manual work required by the customer to fully migrate to the product that you have just sold to them.

Its specific implementation depends on the context, but at a high level, the agent’s role can be broken down into two main steps:

• Planning: Based on what it knows about the target system’s data structure, and by reviewing the customer’s existing systems and data, the agent can design a migration workflow and even generate code snippets to map data from one format to another format.

• Actual migration: The agent can run migration scripts and call on very low-cost AI assistants to perform data mappings that require language interpretation. For example, mapping a “Technology & Software” industry segment to either “Technology hardware” or “Technology software” based on other company information.

AI-powered sales/product customer feedback loop

The goal of the customer feedback agent is to aggregate customer comments from all sales interactions and generate feature requests for the product manager, together with a quantitative assessment of the number of accounts with explicit or latent demand for each feature.

The agent’s workflow is fairly straightforward, closer to that of an AI assistant than an AI agent:

• Step 1: Read and summarize all customer insights from sales conversations (either verbatim or interpreted by the salesperson) and generate a first list of feature requests.

• Step 2: Based on the list of feature requests generated in Step 1, check all available sales conversations to score whether each particular account is likely to be converted or retained if this feature is added.

Takeaway messages

AI-powered sales automation can transform traditional sales processes into more efficient operations that leverage the best of what humans and AI agents have to offer. Quick-win solutions can be prototyped and implemented within weeks, as long as the company is prepared to remove them quickly if the organization’s feedback is poor. Targeted improvements, such as the ones described in this post, should be preferred over full-fledged re-engineering of sales workflows.

In future posts, we will discuss and demo most of the AI agents discussed today.

Subscribe now

Share

The publication is worth a read to get a sense of what businesses have been working on for the past couple of years.

There is a broader point: any B2B company should have a knowledge base listing customer use cases and impacts, and should use it to personalize all outbound sales communications.

I was recently invited to attend a company’s presentation about AI solutions for sales. The company’s follow-up email was customized with my name, but the rest of the message was generic, leading to the obvious question: shouldn’t they have used AI to personalize the email?

We can use Google Cloud’s blog post to try this out.

Let’s consider the following scenario: You are selling AI solutions and you’ve got a list of email contacts from a webinar. Each contact is associated with an industry sector (e.g., financial services, healthcare) and a function (e.g., sales, product, customer service).

You can use the list of customer case studies to generate personalized emails and tell each prospect about the specific ways that GenAI can help their business.

Read below for details.

The “GenAI for you” outreach assistant

The “GenAI for you” outreach assistant is able to take a list of email contacts, access a list of customer use cases and business impacts from a knowledge base, and send one customized email to each contact.

Each email looks like this:

Dear {first name},

Thank you for joining our AI masterclass last week.

I hope that the session provided you with valuable insights. I just wanted to follow up and highlight some of the ways that leading companies have applied GenAI in your industry or your function.

Here are a handful of relevant case studies:

{list of relevant case studies}

I’d love to continue the conversation and explore how we can help you to identify the right opportunities and create value with AI. If you’re open to a quick 30-minute conversation, feel free to share a time that works best.

Let’s look at the steps needed for a simple implementation of this approach.

Inputs

Email contact list

Typically, a registration form for gated content (e.g., webinar, report) generates the email contact list, which looks like this:

Knowledge base

The knowledge base is a list of GenAI use cases and business impacts.

For the purpose of our demo, we can use Google Cloud’s list of 601 GenAI use cases, which looks like this:

Customer Agents

* Continental is using Google's data and AI technologies to develop automotive solutions that are safe, efficient, and user-focused. One of the initial outcomes of this partnership is the integration of Google Cloud's conversational AI technologies into Continental's Smart Cockpit HPC, an in-vehicle speech-command solution.

* General Motors’ OnStar has been augmented with new AI features, including a virtual assistant powered by Google Cloud’s conversational AI technologies that are better able to recognize the speaker’s intent.

* Mercedes-Benz is providing conversational search and navigation in the new CLA series cars using Google Cloud’s industry-tuned Automotive AI Agent.

…etc….

We can get the knowledge base online from Google Cloud’s post. For the purpose of the demo, we simply get it from a knowledge_base.md file.

Email generation

The list of relevant case studies can be generated using the following prompt for each prospect:

You are an sales assistant tasked with generating between 3 and 5 AI case studies for your customer. 

You are provided a knowledge base consisting of a list of real-world AI use cases, located between the <knowledge_base> tags below.
You know the industry and function of the customer:
* Industry: {industry}
* Work fuction: {function}

Select the case studies that are most relevant to your customer and exhibit the most striking business value, prioritizing the use cases that include quantitative impact.

You must respond with a list of case studies in bullet points in markdown format, with each case study as a bullet point, and the case study details as sub bullet points. The case study details must include, first, a **summary** sub-bullet point describing the concept of the use case and the name of the company, and second, a **business value** sub-bullet point describing the impacts achieved. Do not include any other text in the response.

When formatting Markdown, the bullet points must start with a star, and the sub-bullet points must be indented with four spaces followed by a star.

Here is the knowledge base:
<knowledge_base>
{knowledge_base}
</knowledge_base>

You can find the full demo and code at this link:

Visit this notebook.

In case the knowledge base is much larger, you may consider filtering it using RAG first, before the email generation step.

Takeaway messages

Sales outreach personalization is achievable at scale using GenAI tools and relevant knowledge bases. Of course, the quality of outputs depends heavily on the knowledge base and prompt engineering. By tracking email opens, clicks and responses, businesses can measure the performance of their agents and refine their approach.

Subscribe now

Share

Such sources may include:

• Chats or contact details from third-party messaging tools, such as WhatsApp, Telegram, LinkedIn, or Slack.

• Visuals from webinars or unrecorded video conferences with customers, suppliers, or competitors.

• Web pages that actively detect and ban web scrapers, such as Twitter/X and Amazon.

With the emergence of multimodal AI tools, it is now possible to ask GenAI services like Google’s Gemini API to analyze videos and extract information according to specific business requirements.

Video recordings are especially useful in situations where the information is spread across multiple pages, slides or interfaces that can’t be easily accessed by a standard web scraper.

Of course, it’s possible to use screnshots instead of screen recordings, but taking successive screenshots requires more manual actions than simply recording the screen.

A straightforward workflow is as follows:

1. Equip your team members with screen recording tools and automatically upload the recordings to a central repository.

2. Associate each recording with an appropriate data extraction prompt, depending on its source / purpose, or specific employee instructions.

3. Use a GenAI API to process each video and save the desired information in a specified format.

Step 1 - Screen recording tool and central repository

Various tools and services exist to achieve this.

Both Windows and Mac operating systems have built-in screen recording tools. For Mac users, Cleanshot is a well-designed app. More advanced services include OBS and Loom.

Once the recording is created, you can also use any number of tools to upload it automatically to a central repository.

I use Cleanshot for screen recordings, Dropbox to save the recordings, Zapier to detect new recordings, and Airtable as the central repository to save the recordings as well as their data extraction prompts.

Step 2 - Associate each recording with a relevant prompt

Depending on its source and purpose, each recording must be associated with a prompt that describes what information must be extracted and how it should be organized.

You should structure the prompt as follows:

• Context: Who is the AI assistant and what is their job?

• Inputs: What does the input contain and how to access it?

• Output: What exactly should be the format of the output, with an example?

• Final remarks: Use this section to tweak the behavior of the AI assistant when you notice undesired outputs.

For the purpose of demonstrating our approach, let’s use the recording of multiple pages from the Polymarket website. Polymarket is a well-known prediction market. Actually, its website can be scraped easily, but it’s a more appropriate choice for a demo than, say, the recording of a webinar that actually contains private information.

We use the following prompt:

You are an AI assistant tasked with extracting information from a video recording, which is provided to you. 

The video recording contains web pages from a website that allows users to bet on future events. Some pages show an overview of multiple bets available in a specific category, and the current odds of each outcome. Other pages show more detail about a specific bet, its possible outcomes and their odds.
When the bet is just a yes/no question, the odds are displayed in the adjoining text "x% chance" where x is the percentage chance of Yes. You are interested in collecting the list of all available bets, the category that they belong to, the list of possible outcomes and their current odds.

Make sure that your response includes every single bet shown in the video, whether it is shown on an overview page or on a detail page. If you can only see partial data regarding a bet's outcomes and current odds, just include the data that you can. If a bet appears on multiple video frames, you should only include it once.

You must format your response as a list of bets in JSON format. Each bet should be a JSON object in the following format:
<example_object>
{
    "name": "Example Bet",
    "category": "Politics",
    "outcomes": [
        {
            "name": "Outcome 1",
            "probability": 0.5
        },
        {
            "name": "Outcome 2",
            "probability": 0.3
        },
        {
            "name": "Outcome 3",
            "probability": 0.2
        }
        # Etc.
    ]
}
</example_object>

Ensure that your output is a list of objects in the example_format, with nothing else.

Step 3 - Process each video

In the demo, we use Google Cloud’s Gemini 1.5 Pro model (gemini-1.5-pro) to analyze video recordings.

While running the demo, I noticed that Step 2 doesn’t always produce clean data. A data clean-up step must be added, using a LLM too.

Here is an example of data clean-up prompt:

You are an AI assistant tasked with cleaning up the output of an tool that extracts information from a video recording. 

You are given an input between the <input> tags below, which is a JSON list of bets. Each bet is an object containing the name of the bet (name), the category of the bet (category), and alist of possible outcomes (outcomes).
Each possible outcome is an object containing the name of the outcome (name) and the current odds of the outcome as a number between 0 and 1(probability).

Please perform the following clean up tasks:
* If a single bet is included multiple times, merge the data in the way you see fit.
* If a bet does not have any outcomes, remove it from the list.

Ensure that your output is a list of objects in the same JSON format as the input, with nothing else.

Here is the input:
<input>
...
</input>

In our demo, the final output looks like this:

[
    {
        "name": "Next Prime Minister of Canada after the election",
        "category": "Canadian Election",
        "outcomes": [
            {
                "name": "Mark Carney",
                "probability": 0.78
            },
            {
                "name": "Pierre Poilievre",
                "probability": 0.23
            }
        ]
    },
    {
        "name": "Romania Presidential Election Winner",
        "category": "Politics",
        "outcomes": [
            {
                "name": "Crin Antonescu",
                "probability": 0.33
            },
            {
                "name": "Nicusor Dan",
                "probability": 0.3
            },
            {
                "name": "George Simion",
                "probability": 0.26
            },
            {
                "name": "Victor Ponta",
                "probability": 0.1
            },
            {
                "name": "Elena Lasconi",
                "probability": 0.02
            },
            {
                "name": "Calin Georgescu",
                "probability": 0.01
            }
        ]
    },
...
]

You can find the full demo and code at this link:

Visit this notebook.

Takeaway messages

Multimodal AI tools like Gemini 1.5 Pro now make it possible to reliably extract structured data from screen recordings. This opens up new possibilities for businesses to automate data collection from sources that were previously difficult or impossible to scrape.

Using proper workflows and tools, teams can create valuable datasets for business intelligence and decision-making.

Note: Before capturing data from screen recordings, check whether the terms and conditions of the information source explicitly prohibit this practice.

Subscribe now

Share

If you are an entrepreneur or business leader, you have likely found yourself in brainstorming sessions where the goal was to find ways to use GenAI in your business.

You were overwhelmed with ideas that sound exciting and could reinvent every aspect of the company. Sprinkle a few buzzwords like agentic, operator, multimodal, and AGI, and you’ll end up no closer to a decision than you were at the beginning of the meeting.

Granted, every company needs a few AI moonshot projects that reinvent the business. But before going for the Olympics, it needs to develop “AI fitness.” In today’s environment, this means launching practical wins within 3-month production cycles.

AI as the intern

Instead of asking what AI can do, I’ve found it helpful to ask the team: if you had a large number of free, reasonably educated interns, what would you ask them to do?

When we visualize AI agents as interns, our mindset shifts from concepts to tasks. We imagine how we’d give detailed instructions to the interns and what specific tools we’d give them access to.

We have more realistic intuitions about scenarios where the interns can save us time (or save our customers’ time), but we can also leverage our experience of the many stories when interns have taken us more time than if we just did things ourselves.

AI agents, like interns, are particularly effective at:

• Doing extensive research and summarizing findings in a predefined format.

• Cleaning data or transferring it from one platform to another.

• Executive repetitive processes, once their supervisor has performed the work enough times to know how to break them down into individual tasks.

• Monitoring the industry or the business and escalating promising or unusual developments to the supervisor.

By contrast, AI agents are not (yet) very good at:

• Getting things done when goals or processes are unclear.

• Coordinating people (either customers or internal teams).

• Interacting in real-time with high-value customers or stakeholders.

• Hustling and taking extreme ownership for business outcomes.

Also, you don’t want an intern or an AI agent to keep a lot of information in their own ad-hoc notes and memory. You’d much rather have them populate the systems and databases already used by the business, so that they augment your existing processes instead of creating new silos.

Evaluation framework for GenAI ideas

When you ask your team to generate GenAI business ideas, request that they answer the following questions for each idea:

• What is the job title of the intern?

  • E.g., social media intern, lead qualification intern, accounts payable intern.

• What information sources must the intern have access to?

  • E.g., social media, contact form responses, invoice database.

• What systems and databases should the intern fill in?

  • E.g., content calendar, CRM, accounts payable database.

• What is the specific sequence of tasks that the intern should be responsible for? For example, for a sales lead qualification intern:

  • Clean up the contact form data.

  • Query existing systems/databases and populate the lead record with all the identifiers that this record is associated with. If unclear, send an email for clarification.

  • Query external systems to check or complete information provided by the lead (e.g., job title, seniority, financials).

  • If the lead has high potential, notify an account manager immediately.

  • If the lead has unclear potential, schedule a 15-minute call with an SDR.

  • On a daily basis, review pending leads and remind salespeople of the 3 contacts they need to follow up with.

• What performance metrics should be used to evaluate if the intern is doing a good job?

  • E.g., number of tasks pending for human intervention, percentage of leads misclassified.

How to implement

GenAI should initially be implemented with minimal changes to internal processes and databases. It should stand on top of legacy platforms before deciding if legacy platforms must be migrated.

Here is a framework:

Takeaway messages

When evaluating GenAI contributions to the business, get your team to think about how they’d use an intern rather than how they’d use an AGI, to avoid falling into science fiction scenarios. Aim to augment and accelerate your existing business processes before reinventing how your customers and employees interact with your company.

This does not mean that businesses should disregard moonshot ideas. But they need to develop AI fitness before going for the Olympics.

Subscribe now

Share

Introduced by Anthropic in late 2024, Model Context Protocol (MCP) is a standardized syntax to inform AI agents / AI assistants on how to use external tools and services, like databases and APIs.

MCP has quickly become the talk of the town, with the CEOs of OpenAI, Microsoft and Google rushing to announce their support.

If you’re an entrepreneur or leader embedding AI capabilities into your products and systems, here are 4 main things to know about MCP.

1. Assistant tools are crucial to integrate AI into business applications.

2. Until today, creators of AI business agents had to write dozens of code lines to instruct the agent on using each tool. Given that many agents use the same tools (e.g., Google Calendar, Airtable), there must be a better way.

3. MCP enables a new, standardized syntax for telling an AI agent how to use an external tool or service. It accelerates tool integration for AI agent creators, as they can reuse code published by others to obtain tool specifications .

4. While MCP can speed up software development, it doesn’t fundamentally change what agents can do. If your company is developing agents, let the devs decide if and when to use MCP. But if you are a vendor wanting AI agents to use your service, ensure that there are publicly available MCP resources to facilitate the integration of your offerings.

Let’s review each of these insights in more detail.

Assistant tools are essential to integrate AI into business applications

Let’s consider an AI assistant for a sales development representative (SDR).

This AI assistant handles the following tasks:

• Collect customer form submissions from various sources (e.g., information requests, events) by reading internal databases and spreadsheets.

• Populate these contacts with additional information (e.g., existing or new customer, industry segment, number of employees) from internal databases (e.g., Google Contacts, CRM) and external services (e.g., Apollo, CoreSignal, Crunchbase).

• Prioritize these contacts based on business criteria.

• Email high-priority contacts to propose a follow-up call with an SDR.

• Check the SDRs’ calendar availabilities to finalize the appointment.

• After the call, follow up with the SDR to collect feedback and route the potential lead to the appropriate channel (e.g., self-service, account manager).

Each task requires the assistant to send queries to internal or external services, called “tools” in the AI assistant context.

Most tools involve the AI assistant calling an API with credentials and parameters and interpreting the response.

Without these tools, AI assistants would be ineffective for businesses.

If you are developing AI assistants for business, you must absolutely integrate tools.

Without MCP, tool integration takes a bit of time

Let’s consider the SDR AI assistant described above. To schedule meetings, it needs to call the Google Calendar API.

To achieve this, most LLM models (e.g., OpenAI, Anthropic, Gemini, Cohere) require the assistant developer to define 2 tools, “list_events” and “create_event”, and to write a “tools” object for the LLM to understand their functionalities and arguments.

The “tools” object looks like this (for an Anthropic LLM).

In addition to declaring the tools, the developer must code each of the functions that translate tool calls from the AI assistant into API calls using Google Calendar’s specifications (e.g., authentication, parameters).

Honestly, when using AI coding assistants like Cursor, most of the above code can be generated by starting sentences and pressing “tab” multiple times. Most tools can be integrated in minutes. But if you’re doing this work for the first time, your tool definitions may not be perfect.

MCP enables a standardized syntax to tell an AI agent how to use tools

With MCP, the developer of the AI assistant (also called the MCP client) can leverage pre-defined services (also called MCP servers) created by SaaS product vendors or the community.

Instead of declaring and coding the tools, the developer can instruct the AI assistant to call the MCP server, which will:

• Provide the list of available tools and their declaration.

• Translate the AI assistant’s tool calls into appropriate API calls.

The list of MCP servers is growing every day, so MCP is a great resource for AI assistant developers who want to integrate many tools quickly.

For a demo of MCP, let’s use a free MCP server provided by Gumloop. First, create a free Gumloop account and an integration with your Google Calendar account. Then, under “Credentials,” create a guMCP API Key.

In our demo, you can see that the MCP server provides the full declaration of Google Calendar tools. This is comparable to the tools declarations required without MCP, except you did not have to write them down yourself:

The full demo is shown in this notebook.

While MCP can speed up software development, it does not fundamentally change what agents can do

From the standpoint of the AI assistant developer, MCP is mainly a time saver, as you can see from the above demo. If someone has already created a MCP server for the tool that you want to integrate, that’s great news! You don’t need to declare the tool’s specifications yourself.

However, MCP is still new, and not all tools have robust MCP servers.

Also, to be honest, the hardest part about tool definition is not the declaration of the tool’s functionalities and arguments; it’s the authentication. For example, if you want the users of your AI-powered product to connect to their own Google Calendar, you need to develop a workflow so that they can log in with their Google email, and then you need to store their Google credentials securely.

Authentication workflows can be complex, and MCP does not make them easier.

So, whether you declare the tool yourself, or use an existing MCP server, is mostly a tech team decision. You can let the devs decide if and when to use MCP.

In many scenarios, you may be better off coding the tools yourself, or using the MCP server to query the tools declaration but handling tool calls yourself so that you can manage authentication appropriately.

However, if you are a vendor and want AI agent developers to use your service as a tool, it is definitely good practice to publish an MCP server yourself, or to work with MCP server operators like Gumloop. This will ensure that other developers can easily integrate your offerings in their AI-powered products.

Exciting times!

Subscribe now

Share

The adoption of audio and video recording transcription services like Otter, Fireflies and Fathom is growing strongly, particularly among salespeople and business development executives.

However, it’s not clear what level of privacy these services provide.

If you are a business leader, it may make sense to limit the number of cloud infrastructure companies that have access to your most sensitive data. For example, if you already trust Google Cloud with some of your data, it would be logical to use them for meeting transcription and summarization as well.

While it’s not as convenient as using a specialized service, it’s fairly easy to set up. So, let’s review how you can use Google Cloud to transcribe meeting recordings and save the transcripts and summaries in your own environment.

The meeting summarization assistant

The meeting summarization agent runs in the background. It monitors meeting recordings wherever they are stored (e.g., a Google Drive, a Dropbox folder, an Airtable base, or another storage location).

Whenever a new recording file has been added to the location, the agent performs the following tasks:

• Uploads the recording (.mp3) file to Google Cloud Storage.

• Transcribes the meeting verbatim using Google Cloud’s Speech-to-Text API v1 with diarization, which recognizes different speakers.

• Creates a summary of decisions and next steps with Gemini.

• Stores the meeting summary, while deleting the recording and transcript for privacy reasons.

Simple implementation steps

Here are the steps to follow to get the agent up and running quickly and easily.

Prerequisite: set up Google Cloud

First, you need to create a Google Cloud project and a service account with the appropriate permissions:

• Visit Google Cloud to set up a project and billing.

• Visit Google Storage to create a storage bucket. You can choose any name and any region. The default settings are fine. Write down the unique name of the bucket, which will be accessible via the https://storage.cloud.google.com/{bucket_name} URL.

• Visit Google Speech-to-Text to select the project and enable the 'Cloud Speech-to-Text API'.

• In the sidebar on the left, under IAM & Admin, select Service Accounts and click on Create Service Account. Choose any name, select Cloud Speech Administrator role as well as Storage Object User, and click Done. If you are reusing an existing service account, go to IAM instead to add this role to the service account.

• Then, under the newly created service account, go to the Keys tab and click on “Add Key” to create a JSON file containing the service account credentials. Download the JSON file to your computer for future use.

• Finally, create a Google Gemini API key by visiting the AI Studio.

Step 1: Upload to cloud storage

As a first step, the agent simply uploada any new recording to Google Cloud Storage. (Scroll to the bottom of this post for the link to a notebook that shows how to do this).

Step 2: Transcribe

In this step, the agent calls Google Cloud’s Speech-to-Text API v1 with diarization. Note that meeting transcription can take 15 minutes or longer.

Step 3: Summarize

In this step, the agent calls Google Cloud’s Gemini API with an affordable model such as Gemini 2.0 Flash, to request a list of decisions/agreements and a list of next steps.

Here is a summarization prompt that you can use with your favorite LLM API.

prompt = f"""Summarize the following text provided between the <text> tags, by generating a Markdown summary in the format provided between the <example> tags.
The Markdown summary consists of two bullet points, 'summary' and 'next steps', each with less than 10 sub-bullet points.
Under summary, please make sure to list the main agreements and decisions reached.add()
Under next steps, please make sure the list the agreed actions and next steps.
Bullet points must start with a star character.
Sub bullet points must be indented with 4 spaces followed by a * character.
Please do not include any other text in your response, other than the list of bullet points.
<text>
{transcript}
</text>
Here is an illustrative example of the output:
<example>
* Summary
    * The meeting participants agreed to pursue business relationships
* Next steps
     * The meeting participants agreed to meet again in two weeks.
     * They will revert back with names of potential team members within 1 week.
</example>
"""

Visit this notebook for the full demo in Python.

Takeaway messages

Having an AI meeting note taker at your beck and call is a fantastic time saver. However, if you feel unsure about trusting a 1-year-old startup with your most sensitive data, you can easily build that agent capability in-house using Google Cloud.

Subscribe now

Share