Supervision and Sign-Off When AI Produces the First Draft

An intellectual property practice that has begun using a large language model for first drafts has changed something more consequential than its drafting speed. It has introduced a step into the file that produces work product without producing a person who is accountable for it. The professional responsibility has not moved. It remains attached, exactly as before, to the signature at the bottom of the document.

That is the entire problem, and it is not a technology problem. It is a supervision problem, and the profession already has a well-developed body of practice for supervising work produced by someone other than the signing attorney. What it does not yet have is a settled answer for supervising work produced by something that is not a someone.

The distinction that will matter in the next twelve months is not whether the firm reviewed the draft. Every firm will say it reviewed the draft. The distinction is whether the firm can demonstrate who reviewed what, on which version, and at what moment. A claim of supervision that cannot be evidenced is, for the purposes of a malpractice inquiry or a professional liability renewal, indistinguishable from an absence of supervision.

What Supervision Actually Covers

When a junior associate produces a first draft, the supervising attorney is doing four things at once, usually without naming them. They are checking the substance against the law and the file. They are checking the completeness of what was considered. They are checking that nothing was asserted that the record does not support. And they are forming a judgement about how much of the draft can be trusted without independent verification, which is a judgement built from knowing the associate.

Three of those four transfer straightforwardly to a model-produced draft. The fourth does not, and its absence is what makes the review harder rather than easier. A supervising attorney knows that a particular associate is strong on claim construction and weak on prior art characterisation, and reads accordingly. No equivalent knowledge exists for a model, whose failure modes are not stable across matters and are not visible in the output.

The practical consequence is that a model-produced draft requires more review attention than an associate draft of apparently similar quality, not less. Firms that adopted the technology expecting a reduction in senior time have generally found the opposite in the first year, and the ones that did not find it are usually the ones not looking carefully.

Three Failures Observed in Practice

The first is cosmetic review. The draft is read for tone, structure and obvious error, and the reader forms an impression of competence from the fluency of the prose. Fluency is precisely what a language model produces most reliably, and it correlates poorly with accuracy. A draft that reads well and cites a case that does not stand for the proposition asserted will pass a cosmetic review every time.

The second is review that arrives too late. The draft is produced early in the matter, circulated, incorporated into other documents, discussed with the client, and reviewed properly only shortly before filing. By that point the cost of a substantive correction has risen enough to bias the reviewer toward accepting what is there. The review happened. It happened at the moment when it was least able to change anything.

The third is review that leaves no trace. The supervising attorney read the draft carefully, corrected three material points, and did so in a document that was subsequently overwritten. Nothing in the file records that a review occurred, what it found, or who performed it. Twelve months later, when the question is asked, the honest answer is that the firm believes the review took place and cannot show it.

The third failure is the most common and the least discussed, because it does not feel like a failure at the time. The work was done properly. Only the evidence of it was not kept.

The Minimum Evidence File

A firm does not need a governance programme to address this. It needs four records, and all four are byproducts of work that is already happening.

The first is the fact of model involvement, recorded per document rather than per matter. Which documents in this file had a machine-produced first draft, and which did not. This is a single field, and its value is that it tells a later reader where to concentrate.

The second is the identity of the reviewer, named, and the version reviewed. Not the matter partner in general, but the person who read this draft. Where more than one person reviewed, all of them.

The third is the timestamp, which is what converts the first two records into evidence. A review recorded without a date establishes that someone read something at some point, which is close to establishing nothing.

The fourth is the substance of what the review changed, at the level of a sentence rather than a redline. Three or four lines are sufficient. This record is the one that distinguishes a review from a signature, and it is the one that a professional liability underwriter will find most persuasive, because it is the hardest to produce retrospectively.

What Will Be Asked, and by Whom

Two audiences will ask, and they will ask different questions.

Corporate clients will ask first, and they will ask in procurement language rather than in professional responsibility language. The question will appear in an outside counsel guideline or a panel review as a request to describe the firm's use of generative AI in the production of client work product, and to confirm the review controls that apply. Firms that have the four records above will answer that question in a paragraph. Firms that do not will answer it with an assurance, and the difference between a paragraph and an assurance is visible to the person reading both.

Professional liability insurers will ask second, and more slowly, because underwriting practice moves at the pace of claims experience rather than at the pace of technology. When the question does appear on a renewal form, it will not ask whether the firm uses AI. It will ask what the firm can demonstrate about supervision, which is a different and considerably harder question.

Neither audience is asking a firm to stop using the technology. Both are asking a firm to show that the accountability structure survived its introduction.

Where This Sits in a Wider Position

Supervision is one instrument within the governance layer of an AI-ready practice, and it is the instrument that becomes urgent first, because it attaches to work already going out the door. The broader question of which workflows a practice should open to AI at all, in what order, and against what readiness, is addressed in our white paper The AI-Ready IP Firm, which sets out the shadow stack diagnostic and the two instruments that convert it into a position.

The related question of what practitioners are already doing without sanction, which is where most supervision gaps originate, is treated in Shadow AI in IP Departments.

Starting This Week

The first action costs nothing and requires no decision about tooling. Take one matter that is currently active and in which a model produced at least one draft. Ask, for each document in that file, who reviewed it, against which version, on what date, and what the review changed.

If the file answers all four questions, the firm's practice is already sound and the remaining work is to make it consistent. If it answers two, which is the common result, the gap is now specified rather than suspected, and specifying it is the part most firms have not done.

Riseon Advisory works with intellectual property practices on the governance layer of AI adoption, as part of the Passage stage of its transformation sequence.