For the past several months, I’ve been running an experiment. Instead of spending all my time trying to predict what artificial intelligence might eventually do to business, I decided to start building around it. What happens if you stop treating AI as a chatbot and start giving it actual work? What happens if AI researches a problem, develops a plan, creates something, checks its work, learns from what happened, and then moves on to the next assignment?
And perhaps most importantly: What happens to the role of the human when machines can do more and more of the work? I don’t have all the answers. In fact, this experiment has probably created more questions than it has answered. But I’ve learned enough to believe we’re at the beginning of a much larger change in how companies operate.
An AI-Native Company Isn’t a Company That Uses AI
Almost every company will soon be able to say it uses AI. That doesn’t make it AI-native. Most businesses today were designed around people. People sit in departments. People operate software. People move information between systems. People attend meetings to coordinate other people. Managers oversee those people. Then we started dropping AI into that structure. We gave employees copilots. We added chatbots. We added AI features to existing software. That’s useful, but it doesn’t really change the company.
I’ve become interested in a different question: If capable AI workers already existed when you designed the company, how would you build the company differently? That question leads you in some interesting directions. Maybe some work doesn’t belong inside traditional software screens anymore. Maybe some tasks don’t need to be organized into jobs. Maybe an AI system can perform a piece of work, another AI system can evaluate it, and a person only needs to become involved when something unusual happens. Once you start thinking this way, you’re no longer just adding AI to a company. You’re beginning to rethink the company itself.
The Models Are More Capable Than I Expected
One of my biggest lessons has been discovering how much modern AI can actually do when you move beyond simple questions and answers. Most people still experience AI through a chat window. Ask a question. Get an answer. Ask another question. That’s useful, but it’s a very limited view of the technology.
Give AI a well-defined objective, enough context, the right tools, and clear boundaries, and something very different begins to happen. It can research, compare alternatives, create documents, analyze data, write software, inspect work, find errors, propose corrections, and sometimes carry a surprisingly complicated assignment from beginning to end.
That doesn’t mean it can do everything. It certainly doesn’t mean everything it produces is correct. But it has changed the question I’m asking. I used to think primarily about what questions AI could answer. Now I’m much more interested in what work AI can complete. That’s a very different way to think about this technology.
The AI Is Often Smarter Than the System Around It
This has probably been my most frustrating lesson. Sometimes the AI understands exactly what needs to happen, and then the system surrounding it gets in the way. A file doesn’t transfer correctly. State isn’t preserved. A long-running process stops. A tool isn’t available. Something that worked yesterday doesn’t work today. Context gets lost between steps. The intelligence may be capable of continuing, but the wrapper around the intelligence can’t reliably support the work.
That has led me to an important conclusion: Better AI models alone won’t give us dependable AI companies. We also need better systems around the models.
Think about a very talented employee working in a company where the computer crashes constantly, the filing system loses documents, nobody remembers what happened yesterday, and every few hours someone changes the employee’s tools. The employee’s intelligence isn’t the problem. The operating environment is. I think we’re seeing a version of that with AI today. The intelligence has advanced incredibly quickly. The systems that allow that intelligence to work reliably over long periods still have catching up to do.
Autonomy Is Harder Than Intelligence
Another lesson surprised me. Getting an AI to understand what should happen can sometimes be easier than deciding what it should be allowed to do. Imagine an AI system discovers a problem and knows exactly how to fix it. Should it fix it? Maybe. But what if the fix involves spending money, changing a production system, contacting a customer, accessing confidential information, changing someone’s permissions, or publishing something publicly?
Those are very different actions from researching a topic or creating a draft. This means AI-native companies need to think seriously about authority. The question isn’t simply, “Can the AI do this?” The better question is, “Should the AI be allowed to do this without a human?” Those are not the same thing.
AI Needs Management Too
For years we’ve talked about AI replacing parts of management. My experiments have made me think about the other side of that equation: AI itself needs to be managed.
An AI worker benefits from many of the same things a human worker does. It needs a clear assignment and an understanding of what success looks like. It needs boundaries and access to the information required to do its job. It needs to know what it can decide, when it needs to ask for help, and how its work will be evaluated.
The methods will be different because machines aren’t people, but the basic organizational problem remains. Someone—or something—has to coordinate the work. This makes me wonder whether one of the important skills of the next decade will be the ability to design organizations containing both human and machine workers.
AI Needs the Freedom to Fail Safely
There’s an uncomfortable problem with AI autonomy. If a person has to approve every step an AI takes, you’ve created a very expensive automation system. The machine may do the work, but the person is still sitting there supervising it. On the other hand, simply turning an AI loose isn’t a reasonable answer either.
The solution I’ve become interested in is safe freedom. Give AI room where it can work, experiment, make mistakes, discover problems, correct them, and try again—without those mistakes causing serious real-world consequences.
This is particularly interesting in areas such as software development. An AI can try something, test it, discover that it failed, change its approach, test it again, and have another system inspect the result. None of that necessarily needs to affect a real customer or production system. Only when the work reaches an important boundary does a person need to become involved. I suspect this idea will become important far beyond software.
Creating Work Isn’t Enough
There’s another trap in AI adoption. AI can produce enormous amounts of stuff: documents, research, code, analysis, plans, presentations, and recommendations. But producing more stuff isn’t necessarily valuable. Someone has to determine whether it’s good.
That’s why I’ve become increasingly interested in independent evaluation. One AI system might create something, while another process challenges it. Did it follow the instructions? Does it actually work? Did it make assumptions it shouldn’t have made? Did it cross a boundary? Can the result be reproduced? If the work fails, it goes back for correction.
That’s starting to look less like a chatbot and more like an organization. There are workers. There are reviewers. There are standards. There are controls. And there are exceptions that eventually reach a human. That is much more interesting to me than simply generating more content faster.
Human Attention May Be the Scarce Resource
This may be the most important lesson I’ve learned so far. AI work is getting cheaper. AI work is getting faster. And the amount of AI work we can generate is becoming enormous. But my time isn’t expanding. There are still 24 hours in my day.
That creates a new bottleneck: human attention.
Imagine I create ten AI workers and each produces eight hours of work every day. That sounds incredible. But what if I need an hour to review each worker’s output? I’ve created eight additional hours of management work for myself. That isn’t leverage. It’s a new job.
So I’ve started thinking differently about AI productivity. The question isn’t simply, “How much work can the machine produce?” A better question is, “How much useful machine work can be produced without requiring human attention to increase at the same rate?” That’s a much tougher problem, and solving it may be one of the keys to building truly AI-native organizations.
AI Could Change the Economics of Small Companies
This is where the experiment starts becoming really interesting. Historically, a company that wanted more capability usually needed more people. Want more research? Hire someone. Need software? Hire developers. Need marketing? Build a marketing department. Need analysis? Hire analysts. Need more management? Hire managers. Every new capability added cost and organizational complexity.
AI may begin separating capability from headcount. A very small company could potentially have access to research, analysis, software development, marketing, operations support, and other capabilities that once required dozens of people.
I don’t think this means the future is companies without humans. I think something more interesting may happen: very small groups of humans may be able to control very large amounts of capability. That could change entrepreneurship dramatically. A person with an idea may eventually need much less capital and a much smaller organization to test it. A ten-person company might be able to compete in areas that once required 100 people. And a single capable person might be able to experiment with businesses that previously would have been impossible to start alone.
Software May Eventually Sell the Result
There’s another idea I’ve been thinking about. Traditional software sells us tools. Microsoft Excel doesn’t prepare your financial analysis. It gives you a tool for preparing one. A CRM doesn’t make the sale. It gives salespeople tools for managing the process. Accounting software doesn’t run the company. People operate the software.
But what happens when AI can operate the software? The customer may eventually care less about the tool. They may simply want the result. Instead of buying software that helps accomplish a task, businesses could increasingly buy completed work.
That starts to blur some familiar lines. Is it software? Is it a service? Is it consulting? Is it an employee? The answer may sometimes be a little of all four. I think we’re going to see a lot of experimentation around this.
We’re Still Very Early
After all this experimentation, you’d think I’d have a clearer picture of where everything is going. In some ways, I have less certainty than when I started. That’s because I’ve now seen both sides.
I’ve seen AI do work that would have seemed almost impossible just a few years ago. I’ve also watched simple things fail for reasons that had very little to do with intelligence. Reliability isn’t solved. Long-running autonomy isn’t solved. AI management isn’t solved. The economics aren’t fully understood. The human-machine organizational model certainly isn’t solved.
And while we’re trying to figure all of this out, the underlying models keep improving. That makes prediction difficult, but it makes experimentation incredibly valuable.
I’d Rather Experiment Than Predict
I don’t know exactly what an AI-native company will look like five years from now. I don’t think anyone does. That’s part of what makes this moment so interesting.
We can spend the next several years debating what AI will eventually do, or we can start testing some of these ideas now. That’s the path I’ve chosen. Build something. Give AI real work. See what happens. Find the weak points. Improve the system. Measure what actually works. Throw away what doesn’t. And repeat.
I expect I’ll be wrong about plenty of things along the way. That’s fine, because the purpose of the experiment isn’t to prove that my predictions about AI are correct. It’s to learn.
And the biggest lesson I’ve learned so far may be the simplest:
We’re not just getting better software. We’re learning how to build a different kind of company.