Can you automate 100% of a company?

No. And the reason that matters is not the usual one: it is not that the models are not good enough yet.

The state of play first, because it is unambiguous. No fully self-directed company exists in 2026. Organisations presenting themselves as autonomous sit in practice at levels 2 or 3 of a five-level scale, a few reaching level 4 in a narrow domain. The spectacular gains: on the order of ten to one, are real and documented, but in specific verticals: code migration, legal document research, tier-one support.

Automation always breaks in the same four places

What is striking, when you look at where autonomous loops stop, is the regularity. Never the same tasks across sectors, but always the same categories.

Legal accountability. Signing a contract, committing to a deadline, issuing a refusal to somebody, disputing an invoice a client is defending.

Brand judgement. What you accept to say in your own name, to whom, and in what terms when the situation is delicate.

Irreversible decisions. The ones you cannot undo the next morning with an apology, because the other party has already acted on them.

Regulated relationships. The candidate, the employee, the patient, the policyholder, every relationship where a third party holds enforceable rights.

A company with those four categories removed keeps running, and a company reduced to those four categories is still a company. The reverse is not true, and that asymmetry is the whole subject.

The real ceiling is called attribution

Here is our conviction, and it explains the regularity described above. A company is not a set of processes. It is a legal person that answers for its acts, before a judge, a client, an employee, an administration. Answering assumes assets, a mandate, the capacity to be sanctioned. A model has none of that and will have none of it: this is not a technical gap, it is a category difference.

Hence the rule we apply in building Balt: automation stops exactly where somebody has to answer. Not where the model fails, since often it would not fail, but where the act commits. That boundary does not move with model quality. It moves with law, and law moves slowly and in the other direction: works council consultation, Article 22 GDPR, Annex III of the AI Act are all texts that name a responsible person rather than dispense with one.

There is a counter-intuitive consequence. Delegating an act to a machine never moves the liability: it stays entirely with whoever delegated. A company “100% automated” would therefore not be a company with nobody responsible: it would be a company where one person answers for everything the machine does, with no way to check any of it. Nobody signs that.

The right question is not a percentage

“What percentage of my company can I automate” is a question that produces no decision. It mixes hours and commitments, which are not counted in the same unit.

The question that produces decisions is: which decisions have a named owner, and which do not?

Run the exercise over one week in a staffing firm. You will find three piles.

Reversible internal acts: updating a record, writing a summary, preparing a skills file, spotting a roll-off. Many hours, no decisions. This is what you delegate first, and it is where most of the real gain sits.

Outbound acts: a message to a candidate, a reply to a client, a job ad published. Few hours and a great deal of commitment, which is why the machine drafts and a person releases.

Door-closing decisions: rejecting, arbitrating, giving up. Very few hours and almost all of the risk, which is why they stay human even when they never leave the company.

The ratio between the three piles is instructive: the first may be 70% of the time and 0% of the decisions. That is why a massive time saving is perfectly compatible with near-zero decision automation, and it is that sentence, rather than a percentage, that is worth holding on to.

The counter-argument, taken seriously

There is an opposing case, argued by serious people, Andrew Yang and the economist Simon Johnson publicly maintain that AI will make work obsolete, and Goldman Sachs estimates roughly a quarter of the global labour market is exposed to it. It would be dishonest to wave it away.

The argument runs: legal accountability is a convention, and conventions change. We did eventually work out how to insure autonomous vehicles; we will work out a regime where the machine, or its vendor, answers in the company’s place.

Two things make that unlikely at the scale that concerns us.

First, that regime already exists and it runs the other way. The AI Act does not dilute accountability, it names it: a provider, a deployer, an identified human overseer. The regulatory movement of the last five years consists of finding somebody responsible, not doing without one.

Second, the automotive parallel misleads. An autonomous vehicle operates in a closed domain, with a written highway code, enumerable situations and an accident report. A company operates in an open domain where “who decided this?” has no mechanical answer.

What the thesis does capture well is that the volume of human work will fall sharply. On that point it is probably right, and we do not contradict it. Our disagreement is about what follows: fewer hours does not mean fewer people responsible. A company doing the same revenue on a third of the hours still needs as many people to say yes: it just needs far fewer people to prepare the yes.

Guardrails first, agents second

One observation recurs among those actually running agents in production: organisations that build the guardrails before the agents move faster than those doing the reverse.

That is not caution, it is speed. An unbounded autonomous loop runs until the first incident, then stops for good, and the stop is permanent because trust does not restart. A bounded loop runs less far and is still running in six months.

So the sequence that works is the opposite of the intuition. You do not ask “what can the agent do” and then “how do we contain it”. You establish what it must be structurally incapable of, and give it everything else.

What this means concretely for a staffing firm

Three things, none of them a retreat.

Aim for autonomy on whole loops, not fragments everywhere. An agent that handles roll-off tracking end to end, detect, ask, chase, write to the ATS, notify, beats six half-automations none of which finishes without intervention. A loop that closes is a gain; a loop that stalls is one more task.

Count decisions, not tasks. You will find there are few of them, and that they are exactly what your clients pay you for.

Name an owner per category of outbound action. Not “the team”: a person. That is what makes delegation extensible, because you stop debating each case.

What we actually think

The fantasy of the company without humans looks to us like an error of analytical level. It treats a company as a stack of tasks, when it is a stack of responsibilities that happen to produce tasks.

What will disappear is the preparation work: and that is the majority of hours in many jobs, including ours. What does not disappear is that somebody has to say yes, and mean it.

We would rather build for that company: the one where a person decides more often because they spend less time preparing the decision. That is not a moral position; it is the only one that describes what actually exists in production today.