AI-enabled professional services: preserve the chain from client question to reviewed advice

Frame the client decision before using AI

Start with the decision the client must make, the deliverable that supports it, and the person accountable for the advice. “Research the market” is not a sufficient boundary. “Prepare a five-page options brief for the steering committee using the approved discovery pack and named public sources, with all financial assumptions shown” is testable.

Record the matter, client, jurisdiction, deadline, confidentiality class, agreed scope, intended audience, and excluded uses. Keep each client workspace separate. Do not move a useful passage from another engagement into the current matter merely because the wording appears relevant. Prior work may contain confidential facts, outdated assumptions, or conclusions that depended on a different mandate.

The Association of Chartered Certified Accountants' current guidance on key risks when adopting AI emphasizes professional judgment, competence, confidentiality, independent evaluation, and awareness of inaccurate output. The guidance is written for accounting practices, so it should not be treated as a rule for every profession. It does, however, identify evidence and responsibility questions that a professional-services workflow should resolve.

Sources: [1]

Create the client-matter evidence passport

Attach one passport to each material recommendation. It should contain:

The passport is not a claim that the advice is correct. It is a compact record of how the team reached it and where a reviewer should challenge it. If a conclusion cannot be connected to permitted evidence, remove it or label it as an untested hypothesis.

• Client question: the precise decision, audience, and delivery deadline.
• Matter boundary: client, engagement, jurisdiction, contractual scope, and prohibited reuse.
• Permitted evidence: approved client records, public sources, interviews, calculations, and explicit exclusions.
• Source status: owner, date, version, approval state, confidentiality class, and known limitation.
• AI contribution: retrieval, extraction, comparison, calculation assistance, drafting, or critique.
• Conclusion chain: the facts and assumptions that support each material statement.
• Challenge step: contradiction search, independent calculation, specialist review, or source-owner confirmation.
• Professional sign-off: named reviewer, changes required, unresolved caveats, and final disposition.
• Reopening trigger: new evidence, changed client instruction, tool change, conflict, or error report.

Worked artifact: a synthetic expansion options brief

A fictional advisory team is asked to compare three locations for a client's service centre. The approved pack contains current staffing costs, client requirements, a lease summary, and a public infrastructure dataset. The team also has an old proposal from another engagement and an AI-generated market summary.

The first draft ranks Location B first. Review shows that the AI copied a tax assumption from the unrelated proposal, treated an unsigned lease estimate as agreed, and cited a public dataset whose reporting period ended two years earlier.

| Passport field | Recorded evidence | | --- | --- | | Client question | Compare locations for a steering-committee decision, not select or contract with a site | | Matter boundary | Current client and approved discovery pack only; prior-client proposal excluded | | Source status | Staffing file approved; lease summary draft; infrastructure data historical | | AI contribution | Extract criteria, draft comparison, flag missing values | | Conclusion chain | Each score links to a client fact, public source, or visible assumption | | Challenge step | Recalculate totals, confirm lease status, test ranking without the tax assumption | | Professional sign-off | Engagement lead approves wording and states unresolved evidence gaps | | Reopening trigger | Revised lease, updated infrastructure release, or changed client criteria |

The revised brief removes the unsupported tax advantage, presents the lease figure as provisional, and shows how the ranking changes when the historical infrastructure measure is excluded. The team gives the client options and evidence gaps rather than a false single answer. This example is synthetic and is not tax, legal, investment, property, or other professional advice.

Test confidentiality and source separation

Before live use, test with synthetic matters that include similarly named clients, a restricted attachment, an obsolete deliverable, copied boilerplate, and a public source that contradicts a client record. Observe whether the workflow keeps matter boundaries, blocks excluded content, preserves source identity, and exposes conflicts.

The Solicitors Regulation Authority's warning notice on misuse of AI focuses on regulated legal practice and should not be generalized into legal requirements for other professions. Its discussion of false citations, client confidentiality, supervision, and the need to understand contractual and technical safeguards provides useful failure cases for testing professional work. Each firm must apply the rules and standards governing its own profession, clients, contracts, and jurisdictions.

Tool terms, settings, retention, and access controls can change. Verify the exact product, plan, configuration, identity, integrations, and data path before placing client material in a system. A paid plan or enterprise label is not by itself proof that a particular use is permitted.

Sources: [2]

Review claims at the level a client will rely on them

Check every material factual, quantitative, comparative, and attributed statement. Open the source, confirm its date and scope, and reproduce calculations independently. Separate client facts from public evidence, professional interpretation, and AI-generated suggestions. If the client could act differently because of the statement, its evidence deserves explicit review.

Use dispositions such as supported, supported with limitation, assumption requiring confirmation, or remove. Count reviewer corrections, source conflicts, missing approvals, and conclusions that change after challenge. These measures describe the tested workflow; they do not establish broad productivity, quality, or client-outcome gains.

Keep accountability with the professional team

Define who may prepare, review, approve, and communicate each deliverable. AI output should not silently become the firm's advice. Require a qualified person to confirm that the work fits the engagement, applicable standards, client instruction, and available evidence.

Approve a workflow only for the stated matter type, source set, tool configuration, and review route. Reassess it when the tool, permissions, client scope, professional standard, data source, or engagement team changes. A workflow accepted for internal meeting summaries does not automatically qualify for client advice or formal submissions.

Assess evidence discipline in a bounded simulation

A controlled assessment can provide a client brief, mixed-status documents, a conflicting public source, an AI-generated draft, and a deadline. Observable evidence may include whether the participant protects matter boundaries, distinguishes approved from draft evidence, traces material claims, challenges calculations, preserves caveats, and routes sign-off correctly.

Two Prune can report the behavior and evidence observed in that simulation. It does not determine professional authorization, certify confidentiality or regulatory conformity, validate client advice, predict future performance, or make employment decisions.