Terms Before Prompts: A Contract Audit for AI Tools in UK Professional Services
For a professional-services firm, the most consequential AI decision is rarely which model produces a better first draft. It is whether the terms attached to that tool are compatible with the duty of confidence owed to clients. A free trial can move from a partner's browser to a team workflow in days, often before anyone has read the data processing agreement. In 2026, as generative tools become embedded in drafting, research and client communication, SMEs need a repeatable contract audit rather than an informal nod from the person who signed up.
Begin with the data flow, not the demo
Before reading clauses, map what the tool will actually receive. Will staff paste client names, matter details, financial records or special category data into a prompt? Will the tool connect to email, document stores or practice management systems? Will outputs be saved back into client files?
That map determines which contractual terms matter. It also clarifies the firm's role under the UK GDPR and the Data Protection Act 2018. In most client-facing uses, the firm remains the controller and the AI vendor acts as a processor. If the roles are unclear, or the vendor claims broad rights over inputs, the arrangement needs closer scrutiny.
Professional bodies add another layer. Solicitors, accountants, surveyors and consultants may owe duties of confidentiality that go beyond data protection law. A contract that is adequate for a generic business user may still be unsuitable for a firm handling sensitive client information.
Confidentiality must cover AI processing
Check that the confidentiality clause applies to AI processing, not merely storage. It should prohibit the vendor from using client data for its own purposes, selling it, disclosing it to third parties for marketing, or retaining it after the service ends.
Where the vendor is a processor, a written data processing agreement should address the Article 28 requirements of the UK GDPR: processing on documented instructions, confidentiality of personnel, security measures, subprocessor controls, assistance with data subject rights, breach notification, deletion and audit rights. A link to generic online terms is not enough if those terms reserve flexibility to change subprocessors or processing locations without notice.
For firms with clients in regulated sectors, consider whether the contract supports sector-specific expectations. A clause allowing broad overseas transfers may be workable only with appropriate safeguards and a documented assessment.
Training data is the clause most often missed
Consumer and free-tier terms frequently differ from enterprise terms. A clause permitting the vendor to improve its products and services may allow inputs to be used for model training. That is a different risk from ordinary storage.
The firm needs an explicit position: no training on client data, a clear opt-out, or enterprise terms that contractually exclude training. Ask whether prompts and outputs are reviewed by humans, whether history is retained, and whether a zero-retention option exists. For sensitive matters, consider disabling chat history and using private or dedicated instances where available.
Record the answer. A vendor's sales page may say one thing while the contract says another; the contract governs.
Liability and indemnity need a reality check
AI outputs can be inaccurate, incomplete or infringe third-party rights. Vendor terms often disclaim accuracy and cap liability at a low figure. That may be acceptable for low-risk drafting assistance, but it is rarely sufficient for client-facing work without human review.
Check whether any indemnity covers intellectual property infringement and data breach, and whether it extends to AI-generated output. Review the liability cap against the potential harm and the firm's professional indemnity arrangements. Do not assume that the firm's existing insurance responds to every AI-related claim; if in doubt, ask the broker before rollout.
A practical control is to require human review of all client-facing output. The contract should not be treated as a substitute for professional judgement.
Exit rights and deletion deserve equal attention
Free trials can become orphaned accounts with retained data. Check termination rights, notice periods, data export formats, deletion timelines and backup retention. Can the firm retrieve prompts and outputs? Can it confirm deletion in writing? What happens if the vendor changes its terms or is acquired?
Where client records are held only inside the tool, lock-in becomes a professional risk. Prefer tools that allow export and maintain client files in the firm's own systems.
A short audit checklist before the next sign-up
Use a one-page record for every AI tool, including free trials.
- Data map completed: what client information enters the tool, and why?
- Written data processing agreement in place, with controller and processor roles clear.
- Confidentiality clause covers AI processing and survives termination.
- Training-data position documented: no training, opt-out, or enterprise exclusion.
- Subprocessors and processing locations identified, with change notification.
- Security measures and breach notification terms reviewed.
- Retention and deletion periods stated, with export rights.
- Liability cap and indemnities assessed against the risk.
- Exit rights, notice periods and term-change rights acceptable.
- Human review controls agreed for client-facing output.
- Approval recorded by an owner, with a review date.
- Tool added to a central AI register.
Make the audit an approval gate
Do not rely on individual enthusiasm. Set a simple threshold: tools that handle no client data may be self-certified by a team lead; any tool that touches client information requires review by operations, the data protection lead or external advisers. Keep the register current and review high-risk tools annually or when terms change.
The aim is not to block AI adoption. It is to ensure that the contract matches the confidence the firm promises its clients. Read the terms before the first prompt, and the free trial becomes a controlled experiment rather than an unrecorded transfer of risk.
