Note · law · technology
Law-As-A-Service: AI adoption in the legal industry
Law is unusually exposed to generative models, and the reason is structural rather than technological. Most legal work is written, research-heavy, and repetitive. That is precisely the profile of task where a modern language model produces usable output quickly, and it describes a large fraction of what junior lawyers are paid to do.
Adoption is already well underway and already uneven. In the American Bar Association’s 2024 technology survey, roughly a third of respondents reported using AI tools, rising to nearly half among firms with a hundred or more attorneys. The gap between those two numbers is the more interesting figure. It suggests the constraint is not the technology, which is equally available to everyone, but the capacity to govern it: procurement, vetting, insurance, and someone whose job is to decide which tool is approved. Large firms have that capacity and small ones do not.
My expectation is that integration becomes the industry norm, and that it transforms lawyers into supervisors of AI in much the way they already supervised junior attorneys. The relationship is familiar: work is delegated to something faster and less accountable than you, and your value lies in framing the task, catching the errors, and signing your name to the result.
The second-order effects follow from throughput. If output per lawyer rises, client expectations about speed rise with it, and the pressure lands on billing. The billable hour is a strange instrument to keep once the time spent stops tracking the value delivered, and industry reporting already describes both the billable hour and the staffing pyramid as norms under erosion. Firms that price by the hour are, in effect, selling the inefficiency that the tools remove.
None of this is unmanageable, and the mitigations are not exotic. Firms should use approved, contractually governed tools with clear rules on privacy and auditing rather than whatever an associate has open in a browser tab. They should require human verification of any AI-assisted research and treat model output as a draft rather than an authority. Professional bodies and courts, for their part, can clarify standards for disclosure, competence, and responsibility, so that accountability stays attached to a licensed practitioner instead of dissolving into the tooling. Inaccuracy and confidentiality are consistently the top-ranked concerns in adoption surveys, and both are addressable by process.
The harder problem is the one that no procurement policy touches, and it is the reason I think the optimistic framing is wrong even if every prediction above holds.
The staffing pyramid is not only a way to make money. It is the profession’s training mechanism. Junior work is tedious because tedium is where judgment is acquired: you read the hundred cases badly before you learn to read the ten that matter. Document review, first-pass research, and routine drafting are simultaneously the least valuable work in the firm and the apprenticeship through which everyone above learned what good work looks like.
Automating the bottom of the pyramid removes the rungs. The immediate effect is a loss of employment, which is bad and visible. The delayed effect is worse and much less visible: a profession that has stopped manufacturing the senior people it depends on. Firms will feel this roughly a decade after the decision that causes it, at which point the training cohort that would have supplied the supervisors has not existed for years. Supervision requires having done the thing you supervise. An institution cannot promote its way out of a gap in that experience, and it cannot buy the missing cohort from anyone else, because everyone made the same decision at the same time for the same reasons.
Nobody inside a firm has an incentive to solve this. The cost of automating junior work is borne by the partnership of 2040; the benefit is booked this quarter. That is the classic shape of an externality, and it usually needs something outside the firm to correct it.
The most plausible correction is in legal education. If the apprenticeship is disappearing from practice, the training has to move upstream, and it has to change what it trains for. Explicitly teaching the management of AI systems, how to specify a task, how to audit output you did not produce, how to know when a fluent answer is wrong, would substitute deliberate instruction for an apprenticeship that firms no longer have a reason to provide. That is a real curriculum change, not a seminar. It is also cheaper than the alternative, which is discovering the gap when the people who should have filled it are already forty.
A note on provenance: this piece began as a short class assignment whose research was AI-assisted, and the original file ended with its own prompt log. I have rewritten the argument and hedged the survey figures, which should be checked against the ABA’s published report before anyone cites them from here. Given the subject, publishing it without saying so would have been its own small joke at my expense.