What a small firm's AI policy covers

Why a policy

The PCRT AI guidance, published in January 2026, says an AI usage policy can set out acceptable use and require staff to disclose when they have used an AI tool, which helps protect a firm against unknown or inappropriate use. It also says members are responsible for work done by staff under their supervision, and that staff need training on any tool used (PCRT AI topical guidance, January 2026).

A firm supervised for anti-money laundering also has to assess the risks of new technology before and while adopting it (MLR 2017 regulation 19), and the ICO expects a data protection impact assessment for most AI use on personal data (ICO: accountability and governance in AI).

Scope and responsibility

• Who the policy applies to: partners, staff, subcontractors and anyone working on client files.
• What counts as an AI tool, including AI features inside other software and browser extensions.
• The person responsible for the policy, for approving tools and for answering questions.
• The date of the policy and when it will next be reviewed.

Approved tools

• A list of approved tools, each with what it may be used for and on which accounts or plans.
• How a new tool is approved: the supplier questions, the contract check and the data protection impact assessment. See confidentiality and AI suppliers.
• A ban on using unapproved tools, or personal accounts on approved tools, for firm or client work.
• Who reviews a tool when its supplier changes terms or releases a new version. The PCRT guidance says an updated tool should be approached with caution until it is shown to work at least as well as the earlier version (PCRT AI topical guidance).

Data rules

• Which kinds of data may go into which tools, for example public material only, anonymised data, or client data in a tool under a processor contract.
• Data that never goes into an AI tool, such as passwords, Government Gateway details, identity documents, and anything about suspicious activity reports.
• How to anonymise, including removing details that could identify a client by other means (PCRT AI topical guidance).
• Where outputs are saved, and that they are kept in the client file and not left in the tool.

Clients and engagement letters

• A point in the engagement letter that AI-enabled software may be used in providing the services, as the PCRT guidance suggests (PCRT AI topical guidance).
• A point on how client data may be processed by suppliers, linked to the privacy notice.
• When the client is told about actual use, for example when AI is central to a deliverable.
• What happens when a client asks that AI tools are not used on their affairs.
• See the engagement letter outline.

Review

• The review standard: AI output is reviewed as the work of a junior colleague, by someone competent in the subject.
• The checks expected: source every figure, check every rule, re-perform calculations, agree to the ledger and confirm citations exist. See checking AI output.
• Which work needs a second reviewer, such as anything sent to HMRC or a client.
• A rule that automated filing steps keep a human review before submission.

Records

• A record of which tool was used on which piece of work, and by whom.
• Prompts, inputs and the sources checked, kept with the working papers where they support the work.
• The reviewer's sign-off.
• How long these records are kept, in line with the firm's retention policy.

Training

• Training before anyone uses an approved tool on client work, covering how it works, its limits and the review standard.
• Refresher training when a tool changes, recorded as continuing professional development where it qualifies.
• Awareness of prompt injection and other risks the NCSC describes (NCSC: AI and cyber security, 13 February 2024).

Incidents

• What counts as an incident: client data put into an unapproved tool, an error reaching a client, a supplier breach, or a tool acting on hidden instructions.
• Who to tell inside the firm, and how fast.
• Assessing whether a personal data breach must be reported to the ICO, which is within 72 hours of becoming aware of a notifiable breach, and recording every breach (ICO: personal data breaches).
• Telling the client, the professional indemnity insurer and the professional body where required.
• Changes the firm makes afterwards to stop it happening again.

Questions