What is left to humans when agents do the work

What commits. And the wording matters more than it looks, because “which jobs will disappear” is the wrong question: it is not a job that goes, it is a tier.

The numbers first. A quarter of companies using generative AI had deployed agents in 2025, with a projection of half by 2027. The movement is real. But in consulting, despite billions invested, the pyramid has not yet moved: the talk runs far ahead of the structure, which should make anyone cautious about one-year predictions and worried about five-year ones.

Three things that become scarcer

When a capability becomes abundant, value moves to what stays scarce. In our industries three things do, and none of them is a technical skill.

Putting your signature on something. Agreeing to be the person who answers for a decision before a client, a candidate or a judge. It is the fundamental scarcity, the one from which the impossibility of fully automating a company follows, and it has nothing to do with the intellectual difficulty of the decision: a rejection is easy to word and impossible to delegate.

Holding a relationship over time. A client who calls you back after three years does so because they remember a moment when you told them no. That reciprocity assumes continuity, a personal reputation and an accepted risk, three things a system can assist and never carry.

Judging a case nobody anticipated. Not applying the rule, but knowing it does not apply. It is the exact inverse of what a system does well: it generalises from the past, and the case that matters is precisely the one that resembles nothing.

The tier that disappears

Here is the part few people say plainly.

In a staffing firm as in a consultancy, the pyramid rested on a useful conflation: volume work produced and trained at the same time. The junior searched for profiles, prepared files, took interview notes. The firm billed that work; the junior learned by doing it.

Automation brutally separates the two. It takes the productive half and leaves the formative half with no support. So what is threatened is not “the junior recruiter job”: it is the mechanism by which a junior became a senior.

The danger is not immediate, and that is what makes it serious. A firm that removes its junior tier tomorrow will see no problem for three years. It will see an enormous one in five, when the departing seniors need replacing and there is nobody behind them. The pyramid does not narrow: it empties from the bottom, and you only notice once it cannot be fixed.

What to do instead

The answer is not to keep pointless work around in order to train people. It is to make explicit what used to be implicit.

Have beginners review the agent rather than redo its work. A beginner asked to review twelve of the agent’s proposals and explain, for each, why it holds or does not, learns faster than by spending ten hours searching. They see twelve lines of reasoning instead of producing one. It only works if somebody decides that is training time rather than wasted time, and writes it down somewhere.

Hand over small real decisions early. Not simulations: genuine arbitrations, small but committing. The capacity to sign is not acquired by watching somebody else sign.

Name what was never written down. A senior’s tacit knowledge, why this client will refuse this profile, why this availability is optimistic, used to travel through shared repetition. Without it, it has to be written, which is incidentally what makes an agent useful: what you teach it is written down, therefore transmissible to humans too.

What it changes for anyone hiring today

Three immediate consequences, and they concern the job description more than the tooling.

Hire on judgement, not on execution speed. The criterion “handles high volume” has a short shelf life. What replaces it is hard to assess in an interview but testable: give a case where the usual rule produces a bad outcome and watch whether the candidate follows it or questions it.

Redefine the junior role by what it decides, not what it produces. A title starting with “source” describes work on its way out. A title starting with “arbitrate, under supervision” describes a role that will still exist.

Accept paying more, to fewer people. That is the economic consequence of everything above, and avoiding it does not make it go away. A smaller, more senior team costs the same budget and produces more, provided the succession question has been solved, which brings you back to the previous point.

There is a pleasant and little-noticed side effect. When preparation work leaves, what remains is almost exclusively the part of the job people liked, talking to candidates, understanding a client need, arbitrating. The part that made recruiters quit after eighteen months was precisely the volume. A properly equipped team should, if our reading is right, keep its people longer; and that is measurable, unlike most promises made on this subject.

It assumes, of course, that the freed time goes somewhere. It could equally be spent doing three times the volume with the same headcount: a legitimate choice, and not the same job. That is a management decision rather than a consequence of the tool, and it is what decides what delegation actually produces.

What will change less than people say

Two notes of caution, because a vision piece with none is a brochure.

The pace. The consulting pyramid is holding despite considerable budgets. Organisational structures move at the speed of careers, not of models. An eighteen-month prediction about employment is very likely wrong in both directions: too pessimistic in the short term, far short of reality at ten years.

Demand. The reasoning “AI does the work so you need less of it” assumes a fixed quantity of work. Historically, an abundant resource has created uses rather than freed up time. A recruiter who can properly follow three hundred candidates instead of fifty is not working less: they are doing a job that did not exist, because it was materially impossible.

One last observation, less comfortable than the rest. Everything above assumes firms notice the problem in time, and the incentive structure argues against it. A manager judged on this year’s margin has every reason to cut the junior tier and no reason to fund a training mechanism whose payoff lands in five years, under somebody else’s name. The pyramid emptying from the bottom is not a failure of understanding; it is a predictable outcome of how services firms are measured. Which is why we think this will be decided by owners rather than by operators, and why it is worth writing down now, while the choice is still open.

Our conviction

We build on the following hypothesis: preparation work will largely disappear, and the number of people who decide will not fall.

That is not a reassuring position: it implies that roles defined by volume will grow scarce, and it would be dishonest to dress it up otherwise. But it also implies that a team’s value will be measured less and less by its size and more and more by what it is prepared to answer for.

That is why we build an agent that asks rather than one that decides. Not out of regulatory caution, but because the only work that will remain is precisely the work we refuse to take, and a tool that took it would destroy the competence its customer will need in five years.