The Schedule Behind the Policy: Writing AI Use Into the Engagement Letter

The Policy Arrives Before the Scope

It is now common for the first substantive document in a competitive tender to be the client's acceptable-use policy for artificial intelligence, appended to the request for proposal and accompanied by a portal acknowledgement. The engagement letter itself, when it finally appears, tends to be silent on which tools may be used, whose data may be entered into them, who reviews the output, and what happens when the output is wrong.

The asymmetry is worth pausing on. The client's policy is drafted for its own staff, addresses conduct rather than contract, and is almost always expressed to be amendable at the client's discretion. The engagement letter is the instrument that allocates risk, prices the work and is read by the firm's professional indemnity insurers. Leaving the first to govern the second is not a compliance posture; it is an uninsured exposure.

The remedy is neither refusal nor capitulation. It is a schedule.

Why an Amendable Policy Is Not a Term

A clause obliging the firm to comply with the client's AI policy "as updated from time to time" transfers an open-ended obligation. The firm cannot price it, cannot insure it with confidence, and cannot know whether a revision issued in the middle of a mandate has rendered an established working practice non-compliant.

That does not mean the policy should be rejected. It means it should be incorporated on defined terms: by reference to a version dated at signature, subject to a change-control mechanism, and expressly subordinate to the engagement letter where the two conflict. A short order-of-precedence clause resolves most of the difficulty before it arises.

Six Questions the Schedule Should Answer

A workable AI-use schedule is rarely longer than two pages. It earns its place by answering six questions with specificity.

Which tools. Resist a brand-by-brand list, which dates within months. Define a class — generative text, image, code, transcription, summarisation — and specify which classes are approved, which are approved only on firm-issued accounts, and which are prohibited outright for client work. A tool register maintained by the firm, available on request, is a lighter control than a contractual annex that must be renegotiated at each product launch.

What data. State plainly which categories of client material may be entered into an approved tool and which may not. Personal data, privileged material, unpublished price-sensitive information and anything subject to a third-party confidentiality undertaking are the obvious exclusions. Confirm the contractual position on model training, retention and deletion, and record the data residency of the service. For regulated clients, the schedule should be capable of being read alongside their own outsourcing and third-party risk obligations without contradiction.

Who verifies. The single most important sentence in the schedule is that outputs are treated as drafts. Work product relied upon by the client should be checked by a qualified individual within the firm before delivery, with factual assertions, citations and calculations verified against primary sources rather than accepted because they read fluently. Where the deliverable turns on a legal authority or a reported figure, the review obligation should be stated rather than assumed.

What is disclosed. Clients differ. Some want a standing note in the file; some want a line in the deliverable; some want disclosure only on request. Fix the position in advance so that the question does not become a negotiation at the point of delivery, when the firm has least leverage.

Who carries the error. An AI-assisted deliverable remains the firm's work. The schedule should say so, tied to the ordinary professional standard of care rather than an absolute warranty of accuracy, and should make clear that no guarantee is given as to the performance of a third-party tool. Where the client has mandated a specific product, the risk allocation should shift accordingly; that is a commercial point, and it belongs in writing.

What is retained. Prompts, outputs and working notes form part of the matter file. Say where they sit, how long they are kept, and how they are deleted at the end of the mandate.

Proportionality Beats Audit

The objection usually raised by partners is that a schedule invites a technology audit at proposal stage. It need not. A tiered structure — prohibited, permitted with conditions, approved — allows the firm to present a single standard position and vary it only where the client's sector demands. Negotiation then turns on one tier rather than on the whole architecture.

It also helps to give the schedule to the client first. Most AI policies are drafted defensively because the client has no evidence of what advisers actually do. A concise, signed statement of permitted tools, data handling and human review frequently resolves the concern before it hardens into an indemnity demand.

Two internal disciplines make the document credible. The first is a tool register with an owner and a review date. The second is a short AI-use log per matter recording which tool was used for what purpose and who reviewed the output. Neither is onerous; both convert an assertion into a record.

The Points That Are Easy to Miss

Professional indemnity proposal forms increasingly ask whether the firm uses generative AI and how output is supervised. A schedule agreed with clients is useful evidence of control at renewal. Data protection documentation should reflect the same position, particularly where a tool provider acts as a processor. And for clients in regulated sectors, the schedule should be capable of being produced to a supervisor without embarrassment.

None of this requires the firm to become a technology specialist. It requires the firm to describe, in contractual language, a practice it already follows.

The Distinction That Matters

The client's policy is a statement of institutional intent. The schedule is a joint record of what will actually happen on the mandate, who will check it, and who bears the consequence if it fails. One is unilateral and revisable; the other is priced, insured and defensible. Firms that treat the policy as sufficient will find, at the point of difficulty, that they have contracted to an obligation they never negotiated.