Do I Need People, or Will AI Agents Run My Business?

A founder asked me last month whether he should fill two vacancies or wait eighteen months, on the basis that he might not need either of them by then.

It is a serious question and he was right to ask it. The answer required taking the two jobs apart, and once we had, the picture stopped resembling the argument that is being had in public. One of the roles was about sixty per cent work that an agent could do today. The other was about ten per cent. Neither of them disappeared, and both of them changed.

Will AI agents replace employees in your business is the wrong shape of question, and the reason matters. Agents do not consume roles. They consume tasks, and then they produce a category of work that did not exist last year and that almost nobody has staffed.

01What can an agent actually do without supervision?

Bounded, repetitive work with a checkable output and a low cost of being wrong.

That covers a great deal: drafting, summarising, classifying, extracting structured information out of documents, first-pass research, routine code changes, reconciling one list against another. In these tasks the current tools are genuinely good, frequently better than a distracted human, and cheap enough that the cost barely enters the calculation.

Two things sit outside that boundary, and both are about failure rather than capability.

The first is sequence length. Take a step that is right ninety-five per cent of the time, which sounds excellent, and chain ten of them together with no checkpoint. The arithmetic gives you roughly a sixty per cent chance of a clean run. That is illustrative rather than measured, and the shape of it is what matters: unsupervised chains degrade much faster than intuition suggests, which is why long autonomous processes disappoint while short ones with a human at the end do not.

The second is the cost of being confidently wrong. An agent that produces a plausible but incorrect answer, in a domain where the reader cannot immediately tell, is worse than no answer. Where the output touches money, safety, regulation or a customer commitment, the requirement is not better output. It is somebody accountable for it.

02Which jobs will AI agents actually replace, and which employees are safe?

Very few whole roles, and fewer than the discussion implies, because a job in a business of this size is a bundle of about a dozen tasks and only some of them are the kind an agent handles. The roles most exposed are the ones where most of the bundle is bounded and repetitive. The ones least exposed are built on exception handling and judgement, which is why they often look the most mundane from outside.

Take the two vacancies at the top of this piece. The first was a research and proposal role: gathering information, assembling documents, formatting, chasing the missing pieces. Genuinely most of that is automatable now, and the residue is judgement about what to include and a relationship with the people who own the content.

The second was an operations coordinator. On paper it looked similar. In practice the work was almost entirely exception handling, chasing people who had not done things, and making small judgement calls about which customer to disappoint on a bad week. Nothing in that bundle is repetitive, bounded or checkable, which is exactly why it looks unglamorous and is very hard to automate.

The useful exercise is not to ask whether a role can be replaced. It is to list what somebody actually does across a fortnight, mark each item as bounded or unbounded, and see what proportion falls where. Founders are consistently surprised in both directions.

An agent does not remove a job. It removes about a third of several jobs, and then quietly adds a new one that nobody has written a description for.

03What new work does an agent create?

Four things, and none of them are free.

Somebody has to own the output. Not review every item, but be accountable for the fact that it is right, in the way a manager is accountable for their team's work. This role is real, it requires domain knowledge, and it is the single most commonly omitted line in any business case I am shown.

Specification. Getting useful work out of these systems is largely a matter of being able to say precisely what good looks like, including the edge cases. That is a skill, it is scarce, and it is the same skill that makes somebody good at briefing a junior colleague or writing a decent contract.

Verification. You need a way to know whether the output is any good that does not involve a person reading everything, because if it does, you have not saved anything. In practice this means samples, checks and a small set of known cases you re-run when something changes.

Maintenance. Integrations break, vendors change models, prompts that worked stop working, and the thing you built in March behaves differently in September. This is ordinary software maintenance and it should be budgeted as such rather than treated as a surprise.

Put a number on those four and the comparison against a salary becomes honest. The token cost of the work is frequently trivial. The cost of owning it is not.

04Should you stop hiring?

The case for pausing is stronger than I would like it to be, so it is worth stating properly rather than dismissing.

It goes like this. Headcount is the hardest thing to reverse in a business of this size. Every person added raises the fixed cost base, and reducing headcount later is expensive, slow and damaging to everybody who stays. If a meaningful share of a role's work is going to be absorbed within two years, hiring now creates a commitment you may spend three years unwinding. Waiting is cheap and reversible. That is a real argument.

What it misses is the timetable and the compounding. Nobody controls when a capability becomes reliable enough to depend on commercially, and the gap between an impressive demonstration and something you would run unattended in your business has consistently been measured in years rather than months. A business that runs understaffed for eighteen months waiting for a capability does not simply stay still. It loses deals, burns out the people who covered the gap, and arrives at the future with a damaged team.

The defensible middle is not a hiring freeze. It is a change in what you hire for. Be slower to add capacity in work that is clearly bounded and repetitive. Be no slower at all in hiring for judgement, specification and customer relationships. And re-examine each vacancy when it falls open rather than filling it by reflex, which is a discipline worth having whether or not any of this happens.

05What does the reshaped team look like?

Flatter in the middle, and thinner in a specific place that should worry you.

The pattern I see is fewer people doing throughput work, the same number doing judgement work, and one or two new positions that sit between the two, owning automated processes and answering for them. Titles have not settled and it does not matter. The function is clear enough: somebody who understands the domain well enough to know when the output is wrong, and is accountable when it is.

The part that deserves more attention than it gets is what happens underneath. In most businesses, the route to a senior person ran through the junior work. You learned the domain by doing the tedious parts of it, badly at first, and the judgement you now rely on was built out of several hundred small mistakes nobody remembers. Automate all of that and you have improved this year's margin by removing the mechanism that produced your next generation of seniors.

That is a real trade, it is slow-acting, and it is invisible for about four years. The businesses that handle it well are keeping some of the junior work deliberately, as training, and being explicit that it is training rather than throughput.

06The version of this that matters

Take your two most expensive roles apart, task by task, and mark each one bounded or unbounded. That afternoon will tell you more than any amount of reading about what is coming.

Then staff for what is left. Fewer people doing repetitive work, the same number exercising judgement, someone accountable for anything running unattended, and a deliberate decision about how the next generation in your business learns their trade now that the work they would have learned it on has become cheap.

The businesses that get this wrong will not be the ones that adopted too slowly. They will be the ones that removed a third of the work from six jobs, congratulated themselves on the saving, and never noticed they had also removed the reason anybody stayed long enough to become good.