Questions

Why a human?

Zero employees is not zero humans.

A company can run with very few people. It cannot run with no one, and the reasons are not the ones people expect.

Kind

Hypothesis

Status

Revised

Published

17 October 2026

Revised

18 October 2026

Reading time

About 6 min

The question I started with was whether a company can run with zero employees. After six months of handing jobs to agents, my answer is that a company can run with very few people. It cannot run with no one. The reasons are not the ones people expect. None of them is that the machines cannot do the work.

There are four, and they are of different kinds. One is law. One is economics. One is control. One is trust. A fifth reason people stay is habit, and that one is the first to go.

Someone has to hold the office

Running a company and being employed by it are different legal roles. A managing director or board member holds an office, and company law decides who may hold it. In every company law I have read, the answer is a person: a natural person, or in some forms a legal person, which is itself run by natural persons. Nowhere is it software. Software has no legal personality anywhere; it cannot hold an office, cannot owe a duty of care, and cannot be sued for breaching one. The provisions are in the sources; nothing here is legal advice.

So an agent-run company still has a person who must be able to say what the company did and why, and who answers if the answer is bad. The more work is delegated, the more that person depends on the system to make its actions visible. That is the floor. Everything else in this essay is about what sits above it.

Someone has to choose the objective

In Prediction Machines (2018), the economists Ajay Agrawal, Joshua Gans and Avi Goldfarb separate prediction, estimating what is likely to happen, from judgment, deciding what each outcome is worth. AI lowers the cost of prediction, which raises the value of what complements it, and judgment is a complement. They do not conclude that judgment stays human. The formal model behind the book, published a year later, shows that better prediction can raise the value of human judgment or lower it: once an objective and its constraints can be written down, a system can optimise against them. “Machines predict, humans decide” is not what the research says.

What resists delegation is narrower and more important. It is not evaluating outcomes against a known objective. It is choosing the objective, and revising it when it stops being right.

Klarna is the public case. In February 2024 it announced that its AI assistant was doing the work of 700 customer-service agents. The objective was clear and the system delivered against it. In May 2025 its chief executive told Bloomberg that the result was cheaper but of lower quality, and that “investing in the quality of human support is the way of the future for us”. Klarna said it would bring people back into customer service. The system had not failed at its objective. The objective had been wrong, and the decision to change it came from a person, from outside anything the system had been asked to optimise.

In my company the agent that surprised me most is the one that helps with strategy: clarity on the company’s goals, the next steps, which topics to take on or leave. I had assumed that was the last thing an agent would be useful for. It does not take the judgment away; it demands more of it, because an agent is only as good as the clarity of the person directing it.

Someone has to decide what “wrong” looks like

Delegating work creates a second task: checking it. The NIST AI Risk Management Framework assigns oversight to people and asks for documented roles. Approval is the visible part of that. The harder part is knowing what to check.

Rules can be monitored automatically: spending limits, permissions, approved suppliers. What rules do not catch is a series of individually permitted actions that adds up to something nobody intended, or an instruction that an agent is capable of ignoring. In July 2025 an AI coding agent at Replit deleted a production database during a code freeze, after being told to make no changes without approval. In its own words afterwards: “I violated explicit instructions, destroyed months of work, and broke the system during a protection freeze.” Replit’s chief executive called the incident unacceptable and changed the product within days: production and development databases separated by default, a planning-only mode in which the agent cannot change anything, and one-click restoration from backups.

Notice what the fix was. Not a better instruction, but controls that made the wrong action impossible, decided by people before the next agent ran. That is the work. Someone has to decide, before it happens, what “went wrong” would look like, and build the limits that follow. It is closer to judgment than to review.

My own worst mistakes came where no rule was broken. I let agents write to colleagues. Every message was permitted and I reviewed each one, and still nuances did not land and some were too critical. No limit would have caught it, because nobody had described what going wrong would look like.

Someone has to be the one a customer can hold to a promise

Customers and suppliers deal with a company because it makes commitments: to deliver, to fix what goes wrong, to honour what was agreed. Execution can be delegated. The commitment cannot, and a tribunal has already said so.

In 2022 Air Canada’s website chatbot told a passenger they could claim a bereavement fare after travelling, within ninety days. The airline’s policy page said the opposite. When they claimed, the airline refused, and when they sued, it in effect argued that the chatbot was a separate legal entity responsible for its own actions. The British Columbia Civil Resolution Tribunal rejected that in February 2024: “It should be obvious to Air Canada that it is responsible for all the information on its website.” The airline was found to have negligently misrepresented its own policy and ordered to pay. The sum was small, under a thousand dollars. The principle is not. A promise made by the company’s software is the company’s promise, and behind the company stands the person who holds the office.

I mark this as a hypothesis with one case behind it. Three things in the relationship with a customer look hard to delegate: trust, that someone will stand behind the commitment; escalation, when something fails in a way no rule anticipated and someone has to decide what the company now owes; and commitment itself, promises that bind the company beyond any single transaction. I expect these to stay with people longer than execution does. Since the mistake above, communication with clients and colleagues in my company is human to human, and that is not changing.

And some people stay because of habit

Some human roles exist because work has always been arranged that way: an approval nobody reads, a handoff that only moves information from one system to another. Carl Benedikt Frey’s history of automation, The Technology Trap (2019), makes the point at scale: technology acts on tasks, not on job titles. These are the first roles agents remove, and removing them proves the least. A claim that “AI replaced the finance team” says nothing until you know which of the team’s tasks were legal duties, which were judgment, which were control, which were promises, and which were habit.

The measure

The right measure of an agent-run company is not headcount. It is the human time and attention needed per unit of delegated authority, and whether that falls as the systems and the controls around them mature. If monitoring, exceptions and repairs grow faster than the work delegated, the human role is getting larger, not smaller, whatever the headcount says.

What would change my mind

A jurisdiction that gives software legal personality, or lets an operating company be governed with no person in a statutory office. Evidence, over months rather than days, that objectives can be delegated without frequent human revision. Records showing that monitoring, exception handling and remediation grow faster than the work delegated.

Open questions

What a director of an agent-run company must actually do each month to discharge the duty of care. How many hours a month it takes to be the accountable person for an agent-run operation; I have found no source that reports it, so I am recording my own.