Intelligence, by design

The status of this text, stated at the outset because it governs the reading: it is a theoretical demonstration. The scene that follows is a situation type, inspired by real situations, and its quantities are illustrative. This document invokes no engagement result, no client case, no proprietary data.

By the end of an assessment engagement on an executive committee, the machine has read everything. Fourteen hours of interviews, transcribed to the word. The strategic plan and its three successive revisions. Two years of executive committee minutes. The official organisation chart, and the one nobody has ever drawn but which the exchanges make visible. The instrument reports. The documentation of the project this team is supposed to carry. It holds all of it in memory, simultaneously, without fatigue and without forgetting.

It would take a few seconds to produce what is asked of it everywhere else: a ranking of the five executives, a score per person, a flight-risk index, a replacement recommendation.

It will not do so. Not because it is incapable — it is perfectly capable, and that is precisely the problem — but because the chain in which it works was built so that it never would, and because that prohibition can be verified line by line.

That machine did not exist three years ago, and the fact that it does is the subject of another of our documents: The World That Is Coming argues that intelligence has always been rationed, that the scarcity is lifting, and that what then opens is not, in the first place, an age of replacement but one of widening. What remained to be settled was what one does with it when what one investigates is people. That is the object of this one.

An artificial intelligence system is ordinarily judged by what it can do. In a profession that bears on people, whose conclusions decide careers and commit capital, it has to be judged by something else: by what it has been forbidden to do, by the reasons for that prohibition, and by the ability to prove that it holds when nobody is watching. That is the subject of this document. It is not a brochure: it is the description of an architecture, delivered with its limits and exposed to contradiction. A reader who wants to find it at fault will find here what is needed to do so — that is the condition of its being worth anything at all.

Every leader who has sought help on a difficult decision about people — recruiting a member of the executive committee, arbitrating a succession, knowing whether the team in place will carry the phase now opening — has met two families of answer, and has found both unsatisfying without always being able to say why.

On one side, practice by hand: an expert judgment expressed with assurance. Four interviews conducted, what can be read is read, and a view delivered whose quality rests on personal acuity. The approach is sincere and often right. But it stands alone in front of a volume no human attention can cover: the strategic plan skimmed, retained from Tuesday's interview whatever chimed with Thursday's, triangulated whatever happens to be remembered. Confidence in an expert judgment, as research has long established, correlates very weakly with its accuracy. This is not a failure of competence: it is the normal condition of a human brain faced with too much material.

On the other side, the software. A platform that administers tests, aggregates, and delivers a dashboard: scores, a fit index, sometimes a traffic light. Volume, this time, is no longer the problem. But the object produced is signed by nobody. It cannot be explained to the person concerned except by reading her own score back to her. It cannot be defended before a board, because no challenger can interrogate the chain that produced it. And it silently shifts a legal risk onto whoever uses it — we shall return to this, it is the least understood point in the whole subject.

The error lies in believing that one must choose. Both answers fail on the same point, by opposite routes: they confuse the volume of the investigation with the quality of the judgment. The craftsman has the judgment and denies himself the volume that would make it solid. The software has the volume and believes it produces judgment.

If these two offerings survive side by side, it is not through market inattention: each answers a real budget line. The craftsman is bought when a decision is too sensitive to be entrusted to a tool, and he is often bought almost confidentially, by a person rather than by an organisation. The software is bought when volume is too great to be handled any other way, and it is bought by a function that must account for a process. Both needs are legitimate; they are simply never served together. And it is at the levels where decisions are at once rare and heavy — the executive committee, the board, the team of a portfolio company — that the absence of a third answer costs most dearly, since both are needed at once: the depth no single person produces, and the commitment no software carries.

The times, moreover, have just reshuffled this false dilemma, and they reshuffle it on both sides at once. The cognitive work of analysis has been repriced downwards, definitively: what was sold yesterday in man-days is produced today in machine-hours — which takes from the artisan his excuse. And what becomes scarce in proportion is precisely what the machine will not carry: judgment that commits a responsibility, consented access to people, the confidence of a committee, the signature at the foot of a conclusion — which takes from the software its pretension. We have argued this movement for its own sake in The World That Is Coming; here it is enough to retain that neither of the two existing offers comes out of it intact.

Hence the frame of this entire document: frontier intelligence widens what a practitioner can investigate; governed judgment is what turns it into a decision. The gap can no longer be closed by hand. The software cannot produce what a leader actually buys, which is not a probability but someone who answers for a reading. The differentiator is therefore neither the tool nor the expert: it is the design that articulates them — and a design, unlike a talent, can be described, verified and contested.

Two practices can employ the same models, read the same corpus, mobilise the same computing power, and produce objects of opposite natures. One delivers a score on a person. The other delivers an investigated file that a human being concludes and signs. The difference does not lie in the technology: it lies in where the line has been drawn, and in the rigour with which it is held.

Our line is the following, and it is invariant: the machine never assesses a person · no decision is automated · the process is consented to and its findings returned · every conclusion is signed.

This is not moral caution added after the fact. Three necessities of different orders converge on this rule, and it is that convergence that makes it solid.

The condition of the material. A leader does not open up to a system that scores him. He opens up to a practitioner who listens to him, within a process whose rules he knows, and from which he knows he will receive what concerns him. This obvious point has an architectural consequence that is rarely measured: the prohibition is not laid on top of the collection, it is what makes the collection possible. A chain that scores dries up its own source — it obtains front-of-house interviews, calibrated answers, everyone's presentable version. Put differently, the quality of what the machine will have to investigate is a function of what it has been forbidden. This is the most counter-intuitive reversal in the whole subject: the restriction does not amputate the system's performance, it produces it.

The legal protection of everyone. This is the subject of the preceding section and it need only be recalled: the line keeps the service on the right side of two regimes — that of high-risk systems, and that of automated decision-making — and it protects the client as much as the service provider, since exposure travels with whoever relies on the result in a determining manner.

The definition of what is being bought. An executive committee, a board, an investment committee do not buy a score. They buy a conclusion that someone signs and for which that someone will answer — before them, before the person assessed, before a court if it comes to that. A machine cannot sign, not for technical reasons but because responsibility is not a feature: it is a legal and moral relation between persons. One does not buy a probability; one buys someone who answers for a reading. The red line therefore defines the product.

It remains to translate all of this into a real chain, for a principle that does not descend into the specifications is no more than an intention.

The grammar of verbs. A prompt is not a way of talking to a machine: it is a specification. The chain is therefore written with the verbs of investigation — monitor, prepare, transcribe, structure, cross-check, corroborate, extract, verify, format — and never with those of verdict: assess, score, rank, predict, qualify, recommend. This is not a precaution of vocabulary. An instruction asking a model to "assess the chief executive's robustness" has built an assessment system, whatever the sales brochure may say, and the legal qualification will follow the reality of the instruction, not its dressing. An instruction asking it to "extract every passage in which the chief executive describes how he decides, with references" has built an investigation system. Same model, same corpus, two products, two legal regimes. This is why the commercial formulations themselves carry consequences: marketing that implies the AI assesses can requalify a service the practice had, for its part, built correctly.

The strictest rule in the practice: no emotional inference by the machine. Here we deliberately place ourselves above the legal minimum. What is permitted is the analysis of content: what the person says, the facts she reports, the structure and coherence of her account, the convergences and contradictions between the interviews, the instruments and the documents. What is forbidden to the machine is the inference of an affective or psychological state — anxiety, tension, enthusiasm, dejection — from the voice, the face, the rhythm, the hesitations, from any para-verbal signal, and, under our house rule, from any channel whatsoever, including the choice of words alone. In practice: the transcript is verbatim and nothing else, with no annotation of tone and no affective marker; no prosodic, facial or video analysis; no sentiment analysis module anywhere in the practice, and any third-party tool that would embed one has it contractually disabled or is ruled out.

And when a model spontaneously produces a qualification of that nature — "seems uncomfortable", "defensive tone" — the passage is neither retained, nor reworded, nor quietly recycled: it is discarded, and the incident is recorded in the engagement log, with its date, the station of the chain concerned and the remedy applied. This detail looks minor; it is the only one that proves anything at all. A rule whose breaches are not recorded is not a rule, it is an intention.

Where, then, does human observation pass? Any observation of an affective or psychological order appearing in a deliverable is an observation made in interview by the practitioner, written and signed by her, and the deliverable documents its provenance. The machine never proposes a first draft in that register — there is no first-draft exception, because the first draft is precisely where the machine's judgment would enter the human's without a sound, under cover of saving time.

The serious objection: what if the human merely signs? This is the only objection that truly counts, and it must be faced, because the entire sector now invokes the human in the loop. The law itself does not ask for human intervention, it asks for meaningful human intervention — the qualifier does all the work. Research on automation bias establishes that an operator tasked with validating a machine proposal tends to accept it, all the more so as that proposal is well written, confident, and costly to contradict. A well-formed synthesis is a remarkably effective anchor. Asserting that "it is the human who decides" therefore proves nothing: it is precisely what has to be demonstrated, through the construction of the chain.

Three provisions see to it. The order of operations, first: the practitioner forms and writes her own reading before receiving the machine's structuring, never after. A judgment formed first can be corrected by the material; a judgment formed second is almost always a ratification. The nature of what reaches the concluding stations, next: the machine does not hand up syntheses but passages, with their references. A synthesis invites assent; a set of sourced quotations forces one to read again. The design deliberately makes the convenient path a little more costly than the right one. The exposure of contradictions, last: where sources diverge, the machine's task is to report the divergence intact, never to harmonise it. A chain that smooths contradictions produces a comfortable document and an unusable diagnosis.

The control that follows is verifiable: the sentences that conclude must attach to sources the signatory has read again. That is what meaningful human intervention concretely means — not a box ticked at the end of the process, but an order of work in which ratifying costs more than judging. One reservation, in honesty, to finish: no design eliminates automation bias, it can only make it costly and visible. Whoever claims to have removed it has not looked for it.

The red line is the architecture. It is read in the instructions, in the logs and in the checkpoints, where any serious challenger can go and verify it.

What the practice does not solve. The frontier shifts the cost of investigation; it does not shift the limits of the profession, and a document that kept quiet about them would not deserve to be read. It does not replace presence: what a person agrees to say, and what she lets be seen, depends on who is sitting opposite her — a badly conducted interview will be instrumented to perfection, and it will remain poor material. The chain amplifies the quality of the collection, in both directions. Nor does it correct a missed scoping: if the engagement has been scoped on the wrong question — and the presenting subject is rarely the real subject — the whole apparatus will investigate the wrong question with exemplary rigour. Depth of investigation is a multiplier of scoping, never a substitute for it. It does not make prediction certain: any reading of a human being remains probabilistic, and will remain so. A structured assessment improves the quality of a bet; it does not turn it into a certainty, and the service provider who suggests otherwise is preparing a disappointment, sometimes a dispute. Nor, finally, does it dispense with consented access to people: where that access does not exist — and there are situations where it cannot exist — the honest answer is that the engagement does not happen, not that it would be done on documents alone.

These limits are not usage caveats relegated to a footnote. They belong to the design on the same footing as the prohibitions: a system that claims to have lifted them has simply stopped looking at them.

A leader who commissions an assessment of his team sooner or later receives a question from his own data protection officer, or from his general counsel. It comes in two parts: on what basis are these people being processed, and does an artificial intelligence decide anything about them? Most providers answer with reassurance. Reassurance is not an answer — the answer is an architecture, and European law has written its specification with a precision few professionals in the sector have taken the time to read.

The category. Assessing people in a work context by means of an artificial intelligence system falls, under the AI Act, within Annex III — the list of uses qualified as high-risk, whose point relating to employment and workforce management explicitly covers the evaluation of persons. The timetable of this regime has moved: the package known as the Digital Omnibus, adopted in 2026, defers its application to December 2027. That deferral changes a deadline; it does not change a specification. A system designed to be defensible in 2027 is a system that is defensible today; the converse — building fast on a bet on the delay — amounts to scheduling a rebuild for the end of the year in which the market will begin to ask for accounts.

The hard prohibition, already in force. One provision benefits from no deferral and has applied since February 2025: the prohibition of artificial intelligence systems intended to infer the emotions of a natural person in the workplace, save for medical or safety reasons. It sits among the prohibited practices, those carrying the highest penalty ceiling in the regulation. For our profession it is the sharpest line there is: any chain that records interviews and entrusts them to a machine is working at the exact edge of that frontier. The question is not whether one approaches it — one approaches it by construction — but whether one can demonstrate which side one stands on, and demonstrate it otherwise than by assertion.

Transparency. The same regulation has required, since the summer of 2026, that people be informed when they interact with an artificial intelligence system or when content is produced by a machine. Translated into the profession: the person being assessed must know what the machine does with her words. This obligation costs something only to those systems that had an interest in its being ignored.

And the point almost nobody anticipates. The General Data Protection Regulation — the GDPR — prohibits a decision producing legal effects, or similarly significant effects, from being based solely on automated processing. This rule is ordinarily read as a constraint bearing on the provider of the system. The Court of Justice of the European Union has decided otherwise in its case law on scoring: where a third party produces a score and the recipient uses it in a determining manner to decide, it is the score itself that constitutes the automated decision. The reasoning transposes without effort to the assessment of executives. A service provider who delivers a score on a person to a client who will use it to decide is not merely taking a risk: he is transferring one to his client, generally without telling him. Conversely, a service provider who delivers a human conclusion, reasoned, signed, and who stipulates by contract that the deliverable informs the decision without making it, protects both parties at the same time as the person assessed.

This symmetry has a practical consequence that management teams generally discover too late: in the economy of the regulation, whoever deploys an artificial intelligence system answers for it, and not only whoever supplies it. A company that commissions an AI-equipped assessment is therefore not merely buying a deliverable — it enters, as far as it is concerned, a regime of obligations of its own: informing the persons, documenting the use, ensuring the competence of those who interpret the results. The good service provider is the one who arrives with those documents already written and hands them to his client; the bad one is the one who leaves his client to discover them six months later, through a question from his own lawyer. The test is simple, and it discriminates very quickly.

One question remains that cannot be settled in a general document, and that it would be dishonest to treat as closed: the legal basis for the processing. In a relationship of subordination, an employee's consent is a legally fragile basis, because the imbalance between the parties weakens the freedom to refuse. Two things must therefore be distinguished that the sector's marketing readily conflates. Fair consent — the person knows what is being done, accepts it, and receives her debrief — is a requirement of method and of professional ethics: without it, the material gathered is worth nothing, and we shall return to this. The legal basis for the processing is a distinct legal question, assessed deployment by deployment, with counsel, and documented in the record of processing activities and in the impact assessment. A serious service provider handles both; he does not mix them in order to make the second fit inside the slogan of the first.

The state of the law summarised here is that of the summer of 2026 and it will continue to move; dates and scopes must be verified at source before any decision. What will not move, however, is the direction: the European regulator has written, in substance, that a system must not conclude alone about a person, that it must be transparent about what it does, and that whoever employs it answers for it. That is, word for word, the professional ethics of the trade — with penalties added. Documented compliance is therefore not a cost of getting into line: it is the form taken by a pre-existing requirement, and, for whoever has built it before the others, a barrier to entry.

Where to go next. The general thesis of which this document is the technical application is set out in The World That Is Coming (white paper no. 1). Its application to decisions of capital is set out in The Blind Spot (white paper no. 2), which also contains the full development of the scientific apparatus drawn on here.