Maeliom’s position on AI at work fits in one sentence, and it is neither cautious nor enthusiastic: in favour, on condition that you are already able to do it without.
This is not a reservation of principle. It is an observation about what delegating means. To delegate is to hand over execution while keeping the ability to assess the result. Without that second half, you are not delegating: you are deferring.
A conviction, and its condition
There is nothing to hold against a tool that saves time. The tipping point is not use, it is order: AI should act on a skill that exists, not in place of one not yet acquired.
Put that way, the rule seems obvious. It is nonetheless broken systematically, because nothing in the tool signals that it is being broken. The output looks the same either way.
Three cases where the difference shows
Writing. Delegating an article assumes you have settled what it must say, to whom, and what it refuses to say. Without that prior decision you get a competent text that advances nothing — and you have no way of noticing, since the text is well written. The skill at stake is not writing: it is editorial strategy, and it does not delegate.
Coding. Delegating implementation assumes you understand the architecture — what depends on what, what can be swapped out, what will be expensive to undo. A developer who holds that map saves considerable time. A developer who does not gets code that works and a debt they cannot name, until the first change that will not go through.
Deciding. Delegating a market analysis assumes you know your market. Without that you do not read an analysis: you believe it. And the weakness of an analysis produced by a machine is almost never in its figures — it is in its framing, that is, in the question it was given, which nobody will read again.
In all three cases, what is missing is not the ability to execute. It is the ability to recognise that the result is wrong.
The test is one question long
The rule would be useless if it could not be checked on the spot. It can, and the question is short: if the answer were wrong, would I see it?
If yes, the tool accelerates: it does quickly what you would have known how to do slowly, and you keep hold of the result. If no, the tool replaces — and not noticing is precisely the problem, not a mitigating circumstance.
The question has a practical merit: it is not asked of a whole profession but of one precise act, at one moment. The same professional will answer yes for one task and no for the next. That is what makes it usable, where “should we allow AI” is never settled.
The chicken-and-egg paradox
The rule then meets a serious objection: if AI should only come in after a skill has been acquired, how does anyone acquire a skill in an environment where the friction of learning has become optional? You cannot ask someone to earn a tool everybody already has.
The question is not new. In the Phaedrus, Socrates reproaches writing with deskilling memory: entrusting your recollections to an external support excuses you from exercising them. The objection was sound — we genuinely no longer memorise as we did — and it did not stop writing from being progress.
The journal AI & SOCIETY takes up that lineage to move the problem: deskilling is not a matter of individual will but of structure. It proposes the notion of capacity-hostile environments — settings where machine mediation prevents human capacity from being cultivated, whatever the good resolutions of those working in them (AI & SOCIETY, Springer).
That shift is what counts. As long as the question is treated as a matter of personal discipline, each person is asked to resist an environment designed so that they will not. The difference from writing is not one of nature, in any case: it is one of speed. It took centuries for memory to reorganise. Here, the medical study discussed in the first part of this series measures a degradation within months.
The real risk is not the beginner
The worry is about the junior using AI before knowing how. The worry is legitimate but misplaced: it is about the visible link. The link that gives way quietly is elsewhere.
Jarrahi describes the mechanism in one sentence: juniors no longer build foundational craft, seniors become approvers rather than decision makers, and mentoring becomes harder. The third term follows from the first two. An expert reduced to approving no longer has a natural occasion to explain why they approve — and transmission that has lost its occasion is not replaced by a willingness to transmit.
BCG documents the same mechanism seen from the organisation: less hands-on practice feeds less ownership, which feeds less scrutiny of the machine’s answers. And because AI absorbs entry-level tasks first, organisations expect newcomers to perform without the on-ramp that used to build seniority — the career ladder is being restructured faster than the learning paths that made it workable.
One leader surveyed by BCG sums the loop up in a sentence: less hands-on practice feeds less ownership, which feeds less scrutiny of the answers. Each turn makes the next more likely, and no single turn is dramatic.
So this is not a question of generations. It is a question of chain: the first link no longer forms, the last has no reason to lean in, and nobody took a decision.
One avenue, and its objections
One possible answer is to grade the autonomy given to AI according to a person’s demonstrated skill, rather than their role or seniority. In practice: on the critical acts of a trade, access to delegation opens once you have shown you can do without.
The idea is coherent. It restores a link between capability and delegation, it gives the expert an occasion to transmit again — assessing means explaining — and it treats skill as what it is, a verifiable acquisition rather than an attribute of rank.
It raises at least three objections, and Maeliom has no settled answer to offer against them:
- Who assesses, and by what right? A “demonstrated” skill assumes a judge. In an organisation that judge is a superior — you will have replaced a hierarchy of rank with a hierarchy of certification, with no guarantee it is any fairer.
- The risk of elitism. A graded access scheme favours those already trained, that is, those who came in before the change. It can close the door in the name of protecting it.
- Friction for friction’s sake. Banning an effective tool in the name of learning is defensible for a critical act, absurd for the rest. The line between the two is not obvious, and it moves.
The avenue is therefore offered as an avenue. What does look solid, on the other hand, is the shift in the question: stop asking whether AI is allowed, and start asking on which acts an organisation wants to remain able to do without it.
What that requires you to settle
Answering that question is not a matter of tooling. It requires knowing which acts constitute the trade, which make the difference against a competitor, and which are mere execution. A company that has not settled it will apply the same rule everywhere — and it will be the wrong one, in both directions.
That is the work Maeliom Consulting does: separating what constitutes an organisation’s own competence from what can be delegated without consequence, before a single rule of use is written.
One question no internal rule will settle
If friction disappears from work, how will anyone still know, from the outside, what a thing cost to produce? The industry has begun to answer the question of authorship. It has not begun to answer the question of effort — that is the subject of the third part.
Common questions
How do you use AI at work without losing skills?
By only handing it what you could do yourself. Delegating means entrusting execution while keeping the ability to judge the result; without that second half, you are not delegating, you are deferring. AI should speed up a skill that exists, not replace it.
How can I tell whether AI is helping me or replacing me?
One question is enough: if the answer were wrong, would I notice? If yes, the tool is accelerating. If not, it is replacing — and not noticing is precisely the problem. The question is asked task by task, not for a whole job.
What is the real risk of AI for teams?
Less the beginner who uses AI than the expert turned validator: they no longer have a natural occasion to explain why they approve, and knowledge transfer stops. Juniors no longer build the basic craft, and nobody has taken a decision.