AI Can Run Your Company. It Just Can't Decide Anything.

A company run entirely by AI agents ran for a year. Here's exactly where it broke, and what that tells founders about what to automate and what to keep.

STORY TIME!

In early 2025, KPMG and the University of Amsterdam started an experiment with a single question behind it: can a company function with no human staff at all, run entirely by AI agents?

They didn't give the agents a business. They let the agents decide what business to build. The agents chose a webshop selling personalised AI-generated art. It's been running for over a year. The CEO is an agent called Avery Jameson, and the plan runs five years.

Fun, right?

Now, the limitations showed up early, and they showed up in a very specific place. When agents were put into broad, complex roles, things like financial leadership or organisational strategy, they drifted. They ignored instructions. Some shut down unexpectedly.

The overall finding is that AI is well suited to task automation and administrative support, while human judgement remains essential for strategic decisions, oversight and ethical safeguards.

Read that sentence again, because it isn't a story about what AI can't do.

It's a map on how to leverage it.

The framework underneath it

I've been using Anand and Wu's framework from Harvard Business Review, and it explains the KPMG result almost perfectly.

They argue that whether a task suits AI depends on two things, and neither of them is how clever the model is.

First, the cost of an error. Low impact, like a clumsy sentence in a draft. Or high impact, like legal liability, reputational damage, a client relationship going sideways.

Second, the type of knowledge the task needs. Explicit knowledge is structured or unstructured information that can be captured and processed. Tacit knowledge is empathy, ethical reasoning, intuition, contextual judgement, the things built through human experience.

This forms two axes and four quadrants:

No Regrets (low error cost, explicit knowledge). AI handles it. Full stop.

Quality Control (high error cost, explicit knowledge). AI produces, a human verifies.

Creative Catalyst (low error cost, tacit knowledge). AI generates options, you pick and refine.

Human-First (high error cost, tacit knowledge). You lead, AI assists. It is not the decision-maker here.

Now look at where the Amsterdam agents fell apart. Financial leadership. Organisational strategy. High cost of error, tacit knowledge, top-right quadrant. Exactly where the framework says they would.

Nobody should be surprised. And I'd argue nobody should be discouraged either, because the other three quadrants are a genuinely enormous amount of your week.

So what does this mean for your company, practically?

Let me put your actual work into the grid so you can use it:

Automate it. Meeting notes and summaries. First-draft documentation. Data cleaning and formatting. Status roll-ups. Research gathering. Scheduling logic. Translating your standard into a written version of itself (yes, that one, we'll come back to it). Low cost if it's slightly off, entirely explicit. Stop doing these by hand and stop paying anyone else to do them by hand either.

Hand over with a check. Client proposals. Financial modelling. Anything contractual. Technical work that touches production. AI drafts it, a named human verifies it. Note the word NAMED, because "someone will catch it" is not a verification step, and we already know what happens when review capacity doesn't keep up.

Use as a thinking partner. Positioning ideas. Campaign angles. Naming things. Structuring an argument you're stuck on. Twenty ways to say something before you pick the one that sounds like you. Low stakes, tacit, and honestly this is where I get some of my best momentum.

Keep it. Who to hire and who to let go. Whether to take the client whose money is great and whose behaviour isn't. How to tell somebody their work isn't good enough without breaking them. What your company is actually for. Which of two good options you're going to live with for three years.

That last quadrant is YOUR JOB.

Where founders get it backwards

Let's talk about a pattern:

Most founders delegate by volume. What takes the most time? Give that away. What's annoying? Automate that.

But the framework doesn't sort by volume. It sorts by cost of error and type of knowledge. Those are completely different questions, and sorting by the wrong one is how you end up with the arrangement almost every founder I meet is currently running:

You gave away the tasks and you kept every decision. :)

Which means the explicit, low-risk work left your desk, and the tacit high-stakes work stayed, PLUS all the approvals attached to the work you handed over. So your desk didn't empty. It concentrated. Everything left on it is now the hardest category of work there is, and there's more of it than before. You failed to sort properly.

The futuristic part with the numbers

Gartner predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from effectively 0% in 2024, and that 33% of enterprise software applications will include agentic AI by then, up from less than 1%. In customer service specifically, they expect agentic AI to resolve 80% of common issues without human intervention by 2029, cutting operational costs by around 30%.

So the decision layer is moving. Not all of it though. Only the explicit, repeatable, bounded-error part of it.

But hold that next to this: Gartner also expects over 40% of agentic AI projects to be cancelled by the end of 2027, on cost, unclear value, or inadequate risk controls. And MIT's GenAI Divide study found that 95% of enterprise generative AI pilots delivered no measurable impact on the P&L at all.

Those two sets of numbers are not in conflict. They're the same finding from two angles.

The capability is real and it's accelerating. The implementations are failing because companies are dropping AI onto processes nobody ever designed properly in the first place. MIT's 5% who got value did it by embedding into high-value workflows with real context, not by bolting a tool onto a mess and hoping.

Which brings me to the thing I actually want to say.

You cannot automate a decision that was never defined.

If "good" only exists in your head, no agent can apply it. If your approval thresholds are vibes, no system can enforce them. If nobody has ever written down who owns what, there is nothing for an agent to plug into.

The companies that win the next three years won't be the ones with the best tools. Everyone will have the same tools. They'll be the ones whose decisions were already legible enough to hand to anything, human or otherwise.

Your actual job, in about three years

I think this is where it lands.

Right now your job is making the decisions. Increasingly, your job becomes designing the system that makes the decisions, and personally holding only the ones in that top-right quadrant.

That means knowing, explicitly, which decisions in your company are high cost of error and which aren't. Which rely on judgement and which rely on information. Who owns each one. What the standard is, written down, outside your head.

That work is not AI work. It's structural work, and it's the prerequisite for all of it.

Two things I've built that help with exactly this. The CEO Decision Filter is free and it's the fastest way to sort what's actually yours from what isn't. The Delegation Architecture is the full system underneath it, the decision filter, the four levels, the handover template, the ownership map, so nothing you hand over, to a person or an agent, comes back.

Both are on my website. And if you'd rather work through it properly, on your actual company, I work one-to-one with people like you on exactly this. I'm taking three founding clients right now and I'd rather tell you it's not the right fit than sell you something that isn't. So, feel free to reach out.

Drink your water. Take care,

Natasha