Before the Pilot Becomes Policy: An AI Procurement Discipline for London Firms

The first AI tool most London firms bought was never really procured. It arrived on an expenses claim, behind a personal subscription, and it was useful enough that nobody asked an awkward question. Two years later the awkward question is due. What happens when the trial becomes a recurring line in the budget, a tab on a hundred laptops and a processing activity that a client's compliance team asks about in a tender questionnaire?

The answer for 2026 is unglamorous: a file. Not a strategy paper, not an innovation committee, but a procurement record with a checklist inside it — one a partner can produce at short notice and that an insurer, auditor or client will read as evidence of control rather than evidence of panic.

The demonstration is not the decision

Fee earners evaluating tools this year have largely seen the demonstrations. The novelty has worn off and the questions have become narrower and more commercial: whose data is this, where does it go, what does the contract say when something goes wrong, and how do we get out? A firm that answers those questions before signing finds the pilot easier to extend. A firm that answers them afterwards usually finds itself renegotiating from a weaker position, often with a vendor that already holds twelve months of its drafts.

Start with the data map, not the model

Any tool that touches client material — a draft lease, a settlement position, a payroll run, a set of management accounts — is a data processing arrangement. In practice that means a written contract meeting Article 28 of the UK GDPR, a named list of sub-processors, and a clear statement of where processing and storage occur. 'Cloud-based' is not a location, and a support desk in one jurisdiction with a training cluster in another is a detail worth knowing.

Firms handling special category data, or advising clients in regulated sectors, should also be able to point to a data protection impact assessment. It need not be long. It does need to exist, be dated, and be revisited when the vendor changes its model or its terms.

Confidentiality, privilege and the client's silent consent

A solicitor's duty of confidentiality, an accountant's professional obligations and a client's own supplier terms can each be compromised by a tool the client never agreed to. Litigation privilege is the sharper edge: material fed into a third-party system may be argued to have lost the protection it would have had in an internal file.

Two safeguards do most of the work. First, a contractual prohibition on the vendor accessing client content except to provide the service, backed by access controls and logging. Second, a short, plain statement in the engagement letter or terms of business telling clients which categories of tool are used and what is never put into them. Clients rarely object to a firm that is explicit. They object to discovering the position later.

The clause that decides everything

'We may use customer data to improve our models' is the most consequential sentence in most AI contracts and the one least likely to be read. Enterprise tiers will often remove it. Consumer and team tiers frequently will not. Where a vendor insists on retention for model improvement, the honest options are a private instance, a contractual carve-out, or a decision not to route confidential material through that tool at all.

The same clause usually governs retention periods and the deletion timetable at the end of the relationship, which is where diligence tends to run out of energy. Persist.

Indemnity, insurance and a call to the broker

Several vendors now offer intellectual property indemnities for generated output. Read the exclusions: they typically fall away where the firm supplied the infringing material, ignored usage limits, or used a preview feature outside the supported product. Note also which entity is giving the indemnity, and whether liability is capped at fees paid.

Then telephone the professional indemnity broker. Insurers increasingly ask, at renewal, whether AI tools are used in client work and how outputs are reviewed. A firm that can describe its register of approved tools, its human review requirement and its contract terms will have a shorter conversation than one encountering the question for the first time in a proposal form.

Twelve questions for the vendor file

1. Where are data stored and processed, and which sub-processors are involved?
2. Is client content used for model training, and can that be contractually disabled?
3. What are the retention periods, and is deletion evidenced in writing?
4. What security certification applies — ISO 27001, Cyber Essentials Plus, SOC 2 — and when was it last assessed?
5. How quickly must a breach be notified to the firm, and in what form?
6. Who owns the output, and what IP indemnity is offered?
7. What accuracy and human-review expectations are written into the terms?
8. Can the firm audit the vendor, or obtain an independent assurance report?
9. How are model changes and feature retirements communicated in advance?
10. What are the termination rights, notice periods and exit assistance?
11. In what format and timescale is data returned, and how is deletion certified?
12. Which entity contracts with the firm, under which governing law?

Exit is a design requirement, not a rainy day

The test of an AI contract is not the pilot. It is the Tuesday in 2031 when the firm needs its matter histories back in a usable format and the vendor has been acquired. Insist on export in a documented format, a defined notice period, a deletion certificate and a named contact for transition. Keep the register of what was used where, so that the question of which matters were touched by a given tool has an answer that is not a guess.

The staff-use policy does the real work

Technology controls rarely survive contact with a busy week. A short policy, signed annually, does more: approved tools only, no client identifiers in free-text prompts, human review before any output reaches a client, and a clear route for reporting a mistake. Pair it with a named owner — a practice manager or operations director is usually the right person — and a quarterly review of the tool register. That is proportionate governance for a firm of twenty, and it is precisely what a client's procurement team will want to see.

A timetable that fits a small firm

Three weeks is enough. Week one: inventory the tools already in use and classify the data they touch. Week two: send the twelve questions to the two or three vendors that matter, and speak to the broker. Week three: issue the policy, agree the register and put a review date in the diary before the renewals season arrives.

None of this makes the technology better. It makes the firm harder to surprise, which in 2026 is the more valuable property.