What should a UK law or accountancy firm verify before an AI agent processes client documents?

What should a UK law or accountancy firm verify before an AI agent processes client documents?

An AI agent may search, summarise or act on documents, but a firm should not connect client files until it can explain what information the system receives, where it goes, who can retrieve it and how a professional checks the result. The decision is whether the proposed tool and safeguards fit the engagement, not whether the demonstration looks useful.

The selected discovery source asks whether professional-services firms use generative AI or agents on client documents, whether prompts and attachments are retained or used for training, and where processing and logs are located. GreenNode's article on deploying AI agents in consulting, audit and law also discusses project-level data separation and access controls. It is an industry source, not UK legal guidance; the checklist below uses its operational questions and does not rely on its Vietnam-specific legal examples.

The checks to complete before documents enter the system

Before using an agent with client material, document the intended task and data, verify the provider's written retention and training terms, establish processing and logging locations, test access boundaries between engagements, and define who reviews outputs before they affect advice or client work. If a material answer is unknown, keep that data out of the system until the firm resolves it.

• Purpose and data: list the task, document categories and client or engagement boundary. Exclude information the task does not require.
• Provider terms: check what happens to prompts, uploaded files, outputs, logs and backups, including retention, deletion and any use for model training. Record the evidence and contract or configuration owner.
• Processing and access: identify the services and locations involved, then test that permissions limit retrieval to the right people and engagement. Ask how cross-client separation is enforced and monitored.
• Human review: name the professional responsible for checking material outputs, correcting errors and deciding whether they may be used in client work.
• Evidence: retain the approved use case, settings, test results, accountable owner and review record so the firm can explain the decision later.

Assess the use case and controls with an AI Risk & Efficiency Audit

Check the full data path, not only the chat setting

An agent can call models or connected services, retrieve files and create logs as part of one task. Ask the provider to map the components that receive client information, including the model, storage, monitoring and supporting services. Confirm which settings are contractual and which depend on an administrator configuration. Keep a dated record of the answers and test the configured behaviour with synthetic documents before any client data is considered.

Access should follow the engagement boundary. A user who can ask the agent a question should not automatically gain access to another client's material. Test permissions with accounts representing different roles and engagements, and confirm that an out-of-scope request is blocked and recorded. Review these controls when the model, integrations, permissions or provider terms change.

Set review and stop conditions before a pilot

Define which outputs are drafts, which require a second professional check, and which actions the agent may never take without approval. For example, a pilot might allow document classification while forbidding external transmission, client-facing advice or updates to a matter record. Log the input source, agent and model configuration, output, reviewer and any correction. Start with a limited, low-sensitivity test set and pause if the system exposes another engagement's data, uses an unverified service or produces an output that cannot be checked.

A governance retainer can help maintain ownership, review and change controls as tools evolve. Set up continuing AI governance and review

Choose a route based on the evidence

Proceed only when the task is bounded, provider terms and data paths are understood, access tests pass, a named professional owns review, and the firm can keep an audit trail. Restrict the agent to non-client or synthetic data while evidence is incomplete. If the controls need technical work, scope that implementation before connecting client systems.

Plan secure AI implementation around the firm's controls