Field notes

Built by AI stopped in April. The agents didn't.

Nine jobs. In every one, the agent drafts, proposes, builds and checks. A person sends, pays, and talks to people.

An audit, a rule I had to break, and what happened in my company after this newsletter went quiet.

Kind

Founder report

Status

Working

Published

19 October 2026

Revised

27 October 2026

Reading time

About 4 min

Six months ago I stopped writing here, in the middle of the experiment. This is what happened.

A week after my last post, an audit was due and I had to take it alone. Our consultant, who had prepared me for every audit we ever had, fell seriously ill. I needed help, fast.

I had one rule since the start of this experiment: AI does not touch my real company documents. I broke it. I gave an agent access to the folder that holds our ISO documentation and asked it to become my consultant. Before it started, I wrote down what it was allowed to do: rework any document it liked, but put nothing back without my review. What it was responsible for: our documents and processes in order for the audit, and a briefing the way my consultant would have given it. And who answered for it: me.

It worked so well that I was delighted. It was also the end of the experiment I had been writing about here. The company I was building on that Mac Mini never launched. I stopped working on it the moment I started bringing agents into my real company, one job at a time.

What moved, and what did not

The finance work moved first, after our part-time helper found a full-time job and the simple accounting tasks were left undone. The agent started with the routine: find each expense, link it to its bill, ask me when a bill is missing. Then it built a financial picture of the company: what was missing from the books, a profit and loss from the data it had, then cash flow, then scenarios. Then it looked at unpaid incoming invoices and suggested payments, but did not make them. It read our contractors’ reports and drafted the outgoing invoices, but did not send them. I review, and I send.

Update, 27 October 2026: how I gave the finance agent the keys to our books, and the lock I built between them, is in The banks are blocking AI agents. I let one in.

The one that surprised me most is the strategic advisor. It helps me get clarity on the goals of the company, the next strategic steps, which partnerships to pursue, which topics to take on and which to leave alone. I had assumed that was the last thing an agent would be useful for. It turned out to be the thing I needed most, because an agent is only as good as the clarity of the person directing it.

Then the others, each with its own story, which I will tell one at a time: the website, which an agent rebuilt and now keeps current; the proofs of concept we build for clients; IT security, which checks our systems every day and proposes fixes but never applies them alone; market discovery, which finds the people our work is for and drafts every message I then send myself; ads for the exhibitions we attend; and offers, where every new deal gets its own project space and the offer is generated from what we know.

Nine jobs. In every one of them the same line holds. The agent drafts, proposes, builds and checks. A person sends, pays, and talks to people.

What went wrong

The worst mistakes came when I let agents talk to colleagues. Even with my review, there were nuances that did not land. Some messages were too critical. Agents are good at writing tickets. Letting one write the review email after a deploy was a mistake.

So communication with clients and colleagues stays human to human. That is not changing. I have a handful of other rules now. They come later, one at a time.

The question underneath

In 1937 the economist Ronald Coase asked why firms exist at all. If you can buy anything on the market, why gather people inside one company? His answer was that organising work inside a company is sometimes cheaper than buying it outside. Agents change that cost, and so the old question is suddenly practical: what still needs to be inside a company, and who still needs to be a person?

I have spent the last months reading what already exists on this, from Coase to the recent work on prediction and judgment and the frameworks for overseeing AI in organisations, and putting it next to what I see in my company every week. Three questions keep coming back. Why a firm? Why a human? And where exactly does the line go between what an agent may do and what a person must?

Where this leaves me

The company these agents work in turns over about €1M a year. People still create the value and talk to our clients and colleagues. Many of the enabling roles are now agent-based.

I set out to find whether a company can run with zero employees. What I found is harder to describe and more interesting. The world around us is changing faster than most of us can follow, and I have been living inside one small piece of that change. nullfte.com is where I try to make sense of it, in plain terms: what actually happens when agents take over work, how it works, what still needs a person, and what I do not yet understand. Two of those open questions: whether the line between proposing and acting can move, and what it would take; and how much of what stays human stays so because of the law, and how much because of judgment.

The site is live today. If you read one more thing, make it Why a human?: why zero employees is not zero humans, and what still needs a person when agents do the work.

If you want the next pieces by email, subscribe here: zambori.substack.com