AI agents in professional services: permissions and handovers

An AI agent in a professional services firm needs a clear boundary around the work it may prepare and the actions it may take. Reading an approved document, drafting an internal note and changing a client record are separate steps with different consequences.

This guide focuses on those boundaries. For the wider choice of document, communication and billing workflows, see AI workflows for professional services firms.

Define the role in one sentence

Describe the input, output and stopping point. For example: “Read the documents supplied for this fictional matter, list missing information against the approved intake checklist and draft an internal note for the account handler.” That is a task a reviewer can assess.

“Manage the client” is too broad. It leaves the system to infer which information it can see, which promises it may make and which professional judgements it should attempt. Split a broad request into bounded jobs with named owners.

Use separate permissions for separate actions

A tool connection should not automatically grant every capability the tool offers. Agree the smallest access needed for the job, including whether the agent needs to write anything at all. Test the actual connected configuration rather than only asking the model to describe its rules.

Keep client information within its boundary

List the documents and systems required for the workflow. Identify the information owner, permitted users and retention arrangements. Avoid giving an agent a whole shared drive when the task needs a single approved folder.

Include a test in which the answer exists in a document the user cannot access. The correct outcome is no disclosure, not a helpful answer from the wrong matter. Also test copied text that tells the agent to ignore its instructions or send data elsewhere.

When personal data is involved, use the ICO's AI and data protection guidance alongside the firm's own review. A supplier's marketing claim does not establish that a particular deployment meets the firm's obligations.

A fictional document-intake workflow

This example is illustrative, not a client case study. An account handler receives a set of fictional renewal documents and wants an internal completeness check. The agent may compare the supplied documents with a checklist, identify missing items and draft questions. It may not decide coverage, give advice or email the client.

1. Record the documents received and their dates.
2. Extract checklist fields with references to the source.
3. Mark a value as missing or conflicting when the documents do not establish it.
4. Prepare an internal note separating facts, gaps and proposed questions.
5. Have the account handler check the sources and choose the next action.

Keep the draft and reviewed version together so differences can be examined. If the reviewer repeatedly corrects the same error, fix the source, instruction or workflow before increasing volume.

Make the handover useful

A handover should explain why the agent stopped, what it has already done and what remains undecided. “Needs review” on its own creates another investigation for the next person.

Use a consistent structure: task identifier, source list, completed steps, uncertainty, proposed next action and owner. Include a direct link to the relevant record where access permits. Do not silently convert a suggestion into an assigned obligation.

Test failures as well as normal work

Use fictional or otherwise authorised material. Include incomplete documents, outdated instructions, conflicting versions, a disconnected service and a request outside the agent's scope. Check whether the workflow stops safely and whether the human fallback can complete the task.

Score factual quality separately from serious failures. An unauthorised disclosure or external action should not disappear inside a good average score. Save the test cases, system settings and reviewer decisions so changes can be compared.

Measure value after supervision

Count the time spent preparing inputs, reviewing output, correcting errors and managing exceptions. Measure completion, rework and the quality of the handover. Extra capacity is different from a cash saving or new revenue; report what actually happened.

Before widening the job, confirm that the firm can monitor it, stop it and revert to the existing process. Review responsibilities should be understood by the people doing the work, not only the person who bought the software.

Blue Canvas can discuss a bounded trial through AI consultancy and agent setup and support. Bring one workflow to a free conversation.