Practice
Why Lawyers Drop AI Midway
They readily adopt it to start a deliverable. They almost always set it aside as it matures. The reason is not distrust.
There is a behavior observed in almost every organization that has adopted legal AI, and that should intrigue more than it does. Professionals readily use AI to start a deliverable: a first draft opinion, a first sketch of a contract, a first version of pleadings. Then, as the deliverable matures and moves from one version to the next, they turn away from it. By the third or fourth iteration, the AI has quietly vanished from the process, and the work finishes as before, by hand.
This behavior is too regular to be chance, and too widespread to owe to individual preferences. One could read it as distrust, as if confidence eroded as the stakes rose. That would be a convenient explanation, but a false one. What one observes is not a prudent withdrawal before risk; it is a practical abandonment before a growing uselessness. The AI is not set aside because one stops trusting it, but because it stops, literally, being useful as the work advances.
They do not set the AI aside because they distrust it. They set it aside because it stops being useful.
Why the work turns away from it
To understand this dropping off, one must look at how a legal deliverable is really built. It proceeds by successive layers. The first version comes out, it is commented on, reworked. The second integrates the feedback and raises more. The third arbitrates between trade-offs already made. The final version is often very different from the first, not because the first was bad, but because legal work advances this way, by sedimentation, each version answering the previous one and carrying the trace of all that was settled along the way.
This is where the dropping off plays out. Today’s AI tools are almost all designed for the first version, and for it alone. The copilot excels at producing a first draft; but from the second iteration, it drops off. It does not know what was corrected between the two versions, nor above all why. It ignores which of the reviewer’s remarks were taken up and which were deliberately discarded. When called on again at version three, it starts as at version one, blank-minded, ready to reopen questions already settled.
The user feels this dropping off immediately, before even being able to formulate it. To keep using the AI would mean re-explaining to it, at each round, all it has forgotten since the previous round: the corrections made, the reasons for them, what was validated, what remains open. The math is quickly done. At the first version, the AI saves time; at the third, it costs time, because the cost of bringing it back up to speed exceeds what it brings. It is faster to abandon it and finish oneself.
A tool designed for the first draft abandons the lawyer at the moment the work becomes truly legal.
The paradox of the best tool abandoned
This dropping off produces a paradox worth pausing on. The more demanding a firm, the earlier it abandons the AI. A less exacting practitioner may settle for the first draft and barely polish it; that one will keep the AI to the end, because their work goes scarcely beyond the first version. The demanding practitioner, however, does precisely the work the AI cannot follow: they iterate, rework, refine, and it is in this refining that they lose the tool. In other words, the AI serves all the less as the work is of higher quality.
This paradox is a signal, and an alarming one. A tool that fades as the demand rises is a tool that helps where help matters least and disappears where it would matter most. It gives the illusion of being useful, because it shines on the visible, easy part of the work, while being absent from the difficult part where the deliverable’s quality is decided. A firm that measures the AI’s contribution on the first draft therefore systematically overestimates its real value, because it measures at the only place where the tool still works.
One then understands why satisfaction surveys on legal AI give such contrasting results. Asked in the moment, about starting a deliverable, users are enthusiastic; asked about their work as a whole, they are disappointed. It is not a contradiction, it is the exact reflection of the dropping off: the tool is excellent where it is tested and absent where it would really be used. As long as the evaluation bears on the first draft, it misses the precise moment the tool stops being useful.
A mismatch of rhythm, not a failure of trust
It is therefore not a rejection, and above all not a judgment on the tool’s quality. It is a mismatch of rhythm. The AI works in isolated interactions, with no tomorrow; legal work advances by accumulation, where each step remembers the previous ones. The two cadences do not agree, and in this disagreement, it is always the human who ends up absorbing the gap, by becoming again the memory the tool lacks. The lawyer who drops the AI at the third version merely observes this gap and draws the rational consequence.
One might think a little training would suffice, that teams must be taught to keep the AI in the loop. But that takes the problem backward. One does not train professionals to compensate for a structural amnesia; one builds a tool that does not have it. As long as continuity rests on the user’s discipline alone, on reinjecting the lost context at each round, it gives way under pressure, and pressure is a firm’s normal state. To ask the user to carry the memory the tool lacks is to ask them to do the work one promised to spare them.
One must measure what this dropping off costs, for it is graver than it seems. The AI is abandoned precisely at the moment the work becomes most demanding, most legal, most value-creating. It helps with the draft, which anyone could sketch, and disappears at the refinement, where the craft plays out. In other words, it assists the least difficult part and leaves the human alone on the most difficult, which is exactly the opposite of what one would expect from a tool meant to multiply the expert.
Lawyers do not distrust AI. They find it incompatible with the way the work advances.
This mismatch of rhythm also illuminates why managerial remedies fail. One tries to mandate keeping the AI in the loop, one sets up usage metrics, one worries about the abandonment rate. But the practitioners who drop the tool at the third version are not disobeying; they are making the most efficient choice given the tool they were handed. As long as that tool forgets at each round, keeping it in the loop amounts to imposing a handicap on oneself, and no directive durably convinces professionals to work less efficiently. The behavior will not change because it is disapproved of; it will change when the tool, itself, can follow.
It is this continuity, from one version to the next, that MAX was designed to carry. A layer that follows the deliverable through its iterations, retains what was modified and why, knows what was validated and what remains open, and feeds each new version with all the previous ones produced. No longer a succession of independent generations the user must piece together, but a followed work in which the AI stays useful until the final deliverable, because it lost nothing of what played out along the way.
This shift changes the AI’s place in the work. As long as it assists only the first draft, it remains a starting gadget, useful but marginal, abandoned as soon as serious matters begin. The moment it follows the deliverable all the way, it becomes a work tool in the full sense, present where value is created. The first generation of tools optimized the first draft. The next will optimize everything that comes after, that is, the essential.
The first generation of tools optimized the first draft. The next will hold the deliverable to the end.
