AI automation is bad for business. That’s why I’m building my company around AI.

I’m the sole owner of Product.ai. I bootstrapped it, we’re profitable, and I don’t answer to investors.

This gave me a unique opportunity to see what our team could accomplish with AI without adopting the strategy that the market rewards most right now, which is “invest in AI to cut jobs.”

And, for the record, we’re not anti-automation at all. We use AI agents all the time and run them 24/7 to work on goals the team sets.

It’s simple. Most people today are afraid that AI will take their jobs. More than half of Americans told a recent poll they’re worried they or someone in their household will lose a job to AI.1

And, with the constant drumbeat of layoff announcements blamed on AI, it’s a reasonable fear. In May 2026, AI led all reasons for job cuts for the third month in a row, with 38,579 announced cuts.2

Some people are taking things even further. In a 2026 survey of 2,400 workers, 29% admitted they were actively sabotaging their company’s AI rollout, and among the youngest workers it climbed to 44%.3

I believe the automate-and-cut worldview is responsible for this resistance, because it puts employers and workers on different teams with different goals. Of course, people aren’t eager to train the technology that will ultimately replace them.

Stop asking what you can automate. Ask what you’re actually trying to do.

Task-first AI automation doesn’t work very well

The strange thing about today’s rush to replace people with automation is that it’s not delivering major productivity gains.

A study of 5,172 customer-support agents using a generative AI assistant found an average gain of about 15% an hour, and most of that went to the least experienced workers while the strongest performers barely moved.4

And these modest gains are not enough to compensate for large-scale layoffs.

The classic example of this happened in early 2024, when Klarna replaced most of its customer-service team with AI. By May 2025 it was hiring people again, with its CEO conceding that over-prioritizing cost had produced ‘lower quality’ service customers didn’t want.5

Klarna is not an isolated example.

In 2026, Gartner found that AI-driven layoffs didn’t deliver return on investment (ROI). Layoff rates were nearly identical at firms reporting high returns and those reporting little or none.6 So the companies choosing layoffs are effectively trading long-term capabilities for short-term market gains.

If reorganizing around a technology instead of using it to cut headcount sounds like wishful thinking, I suggest taking a closer look at recent history.

In the early 1960s, when the computer was the new thing that was going to take everyone’s jobs, Licklider and Engelbart argued its real value would be to amplify what people could do, not to do it for them.7

Licklider was a psychologist who would go on to steer the government research behind the internet, and Engelbart was the engineer who invented the computer mouse. Today, sixty years later, it’s clear that they were right.

Work looks nothing like it did in the 1960s. Whole office jobs are gone, like the stenographer, the typing pool, the switchboard operator, and the clerk who added figures by hand.

But new kinds of work took their place, from software developers to UX designers to data analysts. These careers all require directing machines instead of being replaced by them.

I think AI will take this path, too.

Its real value would be to amplify what people could do, not to do it for them.

What happens if you amplify instead of cut

At my company, I decided to use AI to amplify how people work instead of automating individual tasks and jobs.

Over the past year, we built an AI-native team, and we’re all doing new jobs now. And we’re all still here, working on the same problems, but making more progress.

How did we get to this place?

We did not look at tasks and jobs and think about how they could be sped up or eliminated. Instead, we reimagined our organization from the ground up and thought about how we could structure our systems and processes around what AI and people can do best.

This led us to build a connected, AI-powered collaboration platform that captures everyone’s ideas and work and makes a real-time snapshot of our collective knowledge easily accessible.

When initially designing the collaboration platform, I had three main requirements.

First, it had to run on top of tools we already used. My goal was not to build a new system that would eat up our developers’ time and pull us away from our actual product.

Second, it had to work as well for our non-technical people as for our engineers. A software developer and a public relations lead both had to be able to use it for their own kind of work.

Third, it had to encompass all the collective knowledge in our organization in a way that would not build a culture of surveillance or require us to continually store and sync a flood of half-baked drafts and experimental deliverables.

The surprisingly simple platform architecture

I built the platform around three tools that many teams already have:

1. GitHub, a version-control system made for code. It can store and track any information that can be written as text.

2. Claude Code, an AI agent designed to write software. In practice, it can work on almost any project made of information or content.

3. Claude Cowork, an agentic tool for the people on our team who don’t write code.

Is there a database?

The short answer is no, at least not the kind most software runs on. Our knowledge isn’t stored in a database that you query with SQL, with tables and joins to maintain. It’s written as plain, markdown-formatted text in GitHub repositories, where both people and AI agents can read it directly.

Our GitHub repos include current information on every aspect of our operations, including our people, recruiting activities, product development and testing, research, marketing, content, and more. And while not everyone can access everything, it gives our team an easy and convenient way to know what’s happening in their domain and in the company overall.

We do build a search index on top of that text so the AI can find things quickly across thousands of documents. The index itself runs on SQLite, which is a small, standard database stored as a single file, but it isn’t where our knowledge is kept. The index is disposable, which means it’s generated from the text and can be rebuilt whenever we need it. This keeps the text files as the single source of truth and the index as a convenience.

At a high level, this gives our people and our agents the context they need to operate, in one form both can read.

Separating public and private spaces

Not every piece of work is ready to share. People need somewhere to think out loud, test rough ideas, and store early drafts before any of it goes to the team.

So I designed the platform with two privacy zones. The first is a public space, where published work is available to everyone who needs it.

The second is a private workspace where each person keeps their own files, notes, and goals. Nothing moves from a private workspace into the shared space on its own. It’s published only when the person decides to publish it.

Governance

The rules for the platform are all written in plain markdown, in the same repositories as everything else, so any team member or AI agent can read them directly. The governance documents cover how the platform should be used, who can see which information, and where the line between public and private is drawn.

The technical details

Internally, we call the platform Cortex, because it functions as our shared brain. We think of it as six parts that work together:

1. The substrate: This is the body of markdown files in our repositories that include our goals, strategy, people, operations, shipped work, project details, and governance documents. It’s the company’s memory, written as text.

2. Governance enforcement: We built custom software that checks each new piece of work against our operating rules and flags anything that breaks one. These are the rules for how the company runs, and they cover where each kind of information belongs, what stays private and what gets shared, and how a piece of work connects to the goal it serves.

3. The work model: This is how we define our goals and track our progress. For every project, we document outcomes detailing what success looks like, and we track progress against them as GitHub Issues. Finished work becomes a shipped-work record, and each person decides what to record as shipped and when. We evaluate performance based on how effectively shipped work achieves outcomes, not how busy someone looks or how much AI they used.

4. The interface: This is how people interact with the platform day to day. For most of us, that’s Claude Code or Claude Cowork, plus a shared library of skills. A skill is a named command that documents the steps and context for a particular job, from building a prototype to drawing a branded diagram. The shared skills are available to everyone, and people can also keep their own private skills in their personal settings.

5. The private workspace: This is the personal workspace I described earlier, where each person keeps their active areas of focus, their projects, and their working notes. It also has a step we call memorializing. At the end of a work session you save its context, so the next day the AI picks up where you left off instead of starting cold.

6. Surfaces: These are the dashboards, generated from the markdown. They show your outcomes, your progress in each area of focus, and the latest product updates at a glance. A surface is only ever a view of what’s already in the text. It’s generated from the substrate and from people’s private workspaces, and it’s never a source of truth of its own.

How employees interact with the platform

Our people work with the platform in different ways, depending on what they do and how they like to work.

Engineers tend to work in Claude Code, in the terminal where they can easily access and run code files. People in non-technical roles like marketing or recruiting more often work in Claude Cowork, which behaves more like a traditional chatbot than a command line. Either way, they’re working with the same shared brain and the same library of skills.

In practice, this means someone in public relations can ask the platform about the product roadmap and get an accurate, helpful answer, and an engineer can ask it to draft a customer note, and neither has to learn the other’s tools to do it.

We shifted from a conventional operating model involving lots of Slack, email, and meetings (both virtual and in person) to working almost entirely through the platform. Now everyone has access to advanced AI models, and our daily work looks very different. Some examples include:

No more status meetings

We used to hold weekly meetings so everyone could share an update, and I could get a better idea of what everyone was working on. It was a hassle, but necessary for smooth operation.

With our new AI-enabled work model, these meetings disappeared. Now people share their working sessions with the collaboration platform and AI has access to all of this context.

So, when I need to know what someone’s working on, I just ask AI instead of interrupting them to ask for a report.

Prototype-first development

How we built products also changed dramatically. We used to produce written documents and specs, which all had to be approved by multiple stakeholders before development could begin. This process could add days or weeks to every new release.

Now we start by building a rough prototype that stakeholders can see and interact with. We can quickly review multiple prototypes by simply talking out loud in a session that our platform records.

Most ideas die in their first hour of life, quickly and cheaply, without anyone spending hours on writing or reading documentation.

Mini-CEOs with agent teams

We’re also beginning to give teams of AI agents tasks that they keep working on until they reach the goals we define.

In this case, the human team member’s role is to define outcomes for the agent team, evaluate what it produces, and redirect it in constructive ways.

Because agents can work independently for hours, we can easily supervise one or more agent teams and still have time to do the human aspects of our jobs, like meeting with colleagues and customers.

What a “normal day” at our company looks like

There’s no single normal day, but a few things are true of most of them.

I usually start by catching up on what the team and our agents accomplished while I was away. Our agents run around the clock, so a lot happens overnight. Perhaps a research task finishes, a draft gets further along, or a prototype gets built. Instead of calling a status meeting and interrupting several people to find out, I read a short summary the platform puts together. Or I just ask it what changed and who’s working on what.

From there, most of my day splits between two kinds of work. Some of it is hands-on. For example, I might sit in a session, talk an idea through out loud, build a rough prototype, and attack a decision from a few angles before I commit to it. The rest is supervising. I define an outcome for a team of agents, hand it off, and come back later to judge what they produced and steer them somewhere better.

At the end of the day, the work doesn’t vanish into someone’s head or an unsearchable chat log. People save their sessions to the shared brain, so tomorrow the context is there for anyone, person or agent, who needs it.

We’re all still here, working on the same problems, but making more progress.

In my experience, the biggest risk of AI-native work is comfort work.

We define comfort work as work that feels productive, but doesn’t actually advance our organizational goals. And, because AI can produce multiple iterations of everything so fast, it can be tempting to over-polish everything.

Examples include refactoring code that already works fine, creating ten variants of a landing page that isn’t approved yet, documenting features that nobody is using, and building a deck for every meeting.

All of this activity can make you look and feel busy, but it takes your company further away from its goals. And nobody’s immune. When facing a big decision, I used to commission a thorough written plan like a deck or a phased roadmap to show myself and the team I was taking it seriously. With AI, I can generate multiple options, simulate them, and attack them adversarially in an afternoon and end up with a better decision.

I also think the proliferation of comfort work is why so many AI projects are failing. Gartner expects more than 40% of agentic-AI projects to be scrapped by 2027.8

Tips for building your own AI collaboration platform

If you’re considering this path, this is what I’d recommend.

Build on tools you already have. We didn’t write an AI collaboration platform from scratch. We assembled it from tools our team and plenty of others were already using. The less custom software you have to maintain, the more time you spend on the product you sell.

Make it work for non-engineers from day one. If only your technical people can use the system, a significant percentage of your company can’t. Choose interfaces a recruiter or a writer can use as comfortably as a developer.

Keep your knowledge as plain text. Text is searchable, permanent, and readable by both people and AI. It also doesn’t trap what you know inside a custom product or a database you’ll have to migrate out of someday.

Reward outcomes, not AI usage. The goal should be progress toward your mission, not token consumption for its own sake.

Start from your mission. Before you automate anything, make sure you’re clear on what you’re trying to achieve, then design the work backward from there.

Choose to amplify, not automate

Some very smart pessimists think AI will inevitably be implemented in ways that are bad for individuals and their employers.

Daron Acemoglu, the MIT economist and 2024 Nobel laureate, has spent years showing that most automation is mediocre, displacing people without producing much. He estimates only about a quarter of AI-exposed tasks can be automated profitably at all.910

But most of today’s studies are looking at automation from the bottom up, which automates tasks piecemeal instead of looking at the organization and its mission from the top down.

I think the pessimists are wrong on this one, underestimating both human creativity and the potential of AI, in the right hands, to open new worlds of opportunity.

If this sounds attractive to you, I encourage you to try it at your own organization. Get a blank piece of paper. Pretend you’re starting your business today and redesign it around AI capabilities and what you know your people can do.

Then enlist your team and explain that you’re interested in reimagining the business, not cutting jobs (especially if this is true). And reward them for outcomes, not simply using AI (e.g., tokenmaxxing11).

We’ve been operating this way at my company for the past year, and I’d choose it again. Stop asking what you can automate. Ask what you’re actually trying to do, and what is now possible with AI.

p.s., If you’re working on this too, send me a note. I’d like to hear what you’re doing.

1. Max Rego, “Poll: Half of Americans Concerned over AI’s Jobs Impact,” The Hill, June 10, 2026, thehill.com/policy/technology/5918589-ai-job-concerns-poll.
2. “The Challenger Report: May 2026,” Challenger, Gray & Christmas, June 2026, challengergray.com/wp-content/uploads/2026/06/Challenger-Report-May-2026.pdf.
3. “AI Adoption in the Enterprise” (second annual survey, n=2,400), Writer and Workplace Intelligence, April 2026, writer.com/blog/enterprise-ai-adoption-survey-results-press-release.
4. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” Quarterly Journal of Economics 140, no. 2 (May 2025): 889–942, academic.oup.com/qje/article/140/2/889/7990658.
5. Kristen Doerer, “Klarna Changes Its AI Tune and Again Recruits Humans for Customer Service,” CX Dive, May 9, 2025, customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/747586.
6. “Gartner Says Autonomous Business and AI Layoffs May Create Budget Room, but Do Not Deliver Returns,” Gartner, May 5, 2026, gartner.com/en/newsroom/press-releases/2026-05-05-gartner-says-autonomous-business-and-artificial-intelligence-layoffs-may-create-budget-room-but-do-not-deliver-returns.
7. J. C. R. Licklider, “Man-Computer Symbiosis,” IRE Transactions on Human Factors in Electronics HFE-1 (1960), groups.csail.mit.edu/medg/people/psz/Licklider.html; Douglas C. Engelbart, “Augmenting Human Intellect: A Conceptual Framework” (SRI, 1962), dougengelbart.org/content/view/138.
8. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” Gartner, June 25, 2025, gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027.
9. Daron Acemoglu and Pascual Restrepo, “The Wrong Kind of AI? Artificial Intelligence and the Future of Labor Demand,” Cambridge Journal of Regions, Economy and Society 13, no. 1 (2019): 25–35, economics.mit.edu/sites/default/files/publications/The%20Wrong%20Kind%20of%20AI%20-%20Artificial%20Intelligence%20and.pdf.
10. Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487 (2024), economics.mit.edu/sites/default/files/2024-04/The%20Simple%20Macroeconomics%20of%20AI.pdf.
11. “Tokenmaxxing as a Weird New Trend,” The Pragmatic Engineer, blog.pragmaticengineer.com/the-pulse-tokenmaxxing-as-a-weird-new-trend.