Practice
Anatomy of a Review
When a senior reviews an associate’s work, they do not “check.” They perform an invisible legal act that current AI makes impossible.
Let us film a senior reviewing a note produced by an associate, and slow the sequence down. In appearance, they read. In reality, they do five things simultaneously. They check legal accuracy, of course, but that is the simplest and fastest part. Above all they evaluate whether the reasoning follows the firm’s usual logic on this type of question. They spot what was not said and should have been, which is often more important than what appears in the note. They replace the analysis in the client’s history. And they gauge conformity to a tacit methodology no one has written down.
These five operations together form what is called, by a deceptively simple word, review. It is what transforms a production into a firm deliverable. And it is exactly this that current AI tools make impracticable. Not because they produce badly, but because they produce without letting anything be seen of how they produced.
One must take the time to slow down further, because each of these five operations is a distinct act, and none is mere reading. Checking accuracy is confronting a legal knowledge. Evaluating the firm’s logic is confronting an internal culture. Spotting the omission is confronting an expected exhaustiveness. Replacing in the history is confronting a memory. Gauging the methodology is confronting a way of doing. Five confrontations, five different reference frames, mobilized at once by a trained eye. It is this simultaneity that makes review so fast in appearance and so dense in reality.
Review is not a verification. It is a legal act in its own right.
What the senior no longer receives
When an associate produces this note with a copilot, the senior receives a deliverable detached from all traceability. They do not know which precedents the AI consulted. They do not know whether the firm’s methodology was applied or merely plausibly imitated. They do not know which trade-offs were taken into account. They do not know what was asked of the model, in what order, with what context.
Review then turns into archaeological reconstruction. Instead of validating a chain, the senior must reconstruct it, fragment by fragment, questioning the associate who, themselves, often cannot answer, because the model showed them nothing of its own path. The time saved at production is lost again, and beyond, at review. What was to lighten the senior’s burden displaces it toward them in a heavier form still, that of investigation.
There is worse than lost time. Faced with an opaque but plausible deliverable, the rushed supervisor is tempted to validate on the strength of appearance. The text holds, the references seem correct, the conclusion appears reasonable: one signs. It is exactly the scenario review exists to prevent, and it is the one AI’s opacity makes probable. A plausible and untraceable deliverable is not a productivity gain, it is a risk disguised as a gain.
An AI that produces without a trace does not save time. It displaces time toward review, or worse, discourages it.
This discouragement of review is the gravest point, because it works against the very culture of the craft. A conscientious senior, swamped with opaque but plausible deliverables, cannot reconstruct everything; they end up, by necessity, granting appearance a trust they would never grant to work whose chain they could see. It is not an individual slackening, it is a mechanical consequence of volume and opacity combined. The faster AI produces without a trace, the more it pushes review toward abdication, exactly where the craft demands it strengthen.
It is often answered that recent tools cite their sources, and that traceability is therefore solved. This confuses two very different things. To cite a source is to indicate where a piece of information comes from. To trace an execution is to make visible the whole reasoning: what was queried, what was compared, what was retained and above all what was discarded, according to what methodology, with what human validation, at which step. The first informs about a point; the second gives a path to see.
Citing is not tracing
The distinction deserves a pause, for it is at the heart of the problem. A citation answers the question “is it true?”. A trace answers the question “how was this decision built, and can I defend it?”. The first is a property of the text, attached to a sentence. The second is a property of the chain, attached to a process. They are two different orders, and one never replaces the other.
No assistant that works in isolated interactions can produce the second, because it has no chain, only answers. It can back a sentence with a source, which is useful, but it cannot account for how it reached that sentence, because that how was not kept: it dissolved with the session. The citation is necessary; it is not sufficient. And confusing the two is believing solved a problem that remains whole, that of the defensibility of the reasoning.
This confusion has a practical consequence for the senior. Faced with a deliverable that cites its sources but does not trace its execution, they have the illusion of verifiability without verifiability itself. They can check each reference taken in isolation and remain unable to say whether the whole was built according to the firm’s method, or plausibly assembled around correct citations. The citation reassures about the bricks; it says nothing of the edifice.
Citing a source answers “is it true.” Tracing an execution answers “can I defend it.”
One can measure the gap by a simple experiment. Take a deliverable that perfectly cites its sources and ask, not whether its references are accurate, but how it reached its conclusion: which leads were explored then abandoned, why such an interpretation was preferred to another, at what moment a human validated what. A tool that only cites stays mute on all these questions, because it never kept its own path. Yet these are exactly the questions a senior asks when reviewing, and they are the ones a judge, a colleague or a client asks when the decision is contested.
The real risk: the loss of transmission
The strategic risk is therefore not that a tool produces imperfect work. It is that a firm loses, progressively and without noticing, visibility over how its own reasoning is transmitted, supervised and reproduced. A firm’s methodology has never been a separate document; it lives in the way each deliverable is produced and reviewed, in the correction a senior brings to a junior, in the motive of an added reservation.
If the AI layer produces outside methodology and outside review, it is the chain of transmission itself that atrophies. Juniors learn less, because correction now passes through a black box. Seniors validate less well, because they no longer have the chain before their eyes. After a few years, the firm has produced much and transmitted little, and its most precious asset, its way of reasoning, has grown poorer without anyone having decided anything.
One must measure the slowness and silence of this impoverishment, for that is what makes it dangerous. It does not happen at once, it has no culprit, it triggers no alert. Each deliverable taken in isolation seems correct; it is the reproduction of the know-how, from one generation of practitioners to the next, that breaks in silence. And when a firm notices that its juniors reason less well than before, the cause is already old and diffuse, lodged in years of opaque production no review could reread.
This is why MAX was conceived as a layer that makes methodology active in generation, that traces each step, and that makes the whole chain legible again. Not to replace review, but to make it possible again, and with it, the transmission that makes a firm endure. An AI that traces does not merely produce; it lets the firm keep learning from what it produces.
A firm that produces much and transmits little grows poorer without seeing it. The trace is what gives it transmission back.
