Consulting AI Governance Policy: A Team Template for Client Work

A consultant can turn a confidential meeting note into an artificial intelligence (AI) prompt in seconds. The risk begins before the answer appears: the client may not have approved the tool, the data transfer, or how its information will be stored.

An AI governance policy gives your team a usable boundary for that moment. It should say which tools and data are allowed, who approves exceptions, and what a person must check before work reaches a client. The clauses below are written for adaptation to your firm and its engagements.

Key Takeaways

• Approve the tool and the use case before approving client data for a prompt. A paid account alone doesn’t establish permission.
• Keep a named consultant responsible for every AI-assisted deliverable, with human oversight of its sources, calculations, and recommendations.
• Record approvals and incidents where project teams work to support transparency and accountability. A policy nobody can apply during delivery offers little protection.

Set the boundary before adopting the template

This AI governance framework covers drafting, research, analysis, coding, transcription, and other AI-assisted work. It applies to employees and contractors working on behalf of [FIRM NAME]. Replace each bracketed field, then have the policy owner check it against client agreements and existing security procedures.

A firm-wide approval does not override an engagement restriction. If a statement of work limits subcontracting or external processing, the project team must resolve that restriction before using a third-party AI service. For sensitive work, the right choice may be a client-approved environment or no generative AI at all.

This is general operational guidance, not legal advice. Counsel should review contractual, data privacy, professional, and jurisdiction-specific obligations before the firm adopts the policy.

Policy template: scope and approved tools

Purpose and authority

Policy clause: [FIRM NAME] permits AI to assist with approved consulting tasks when its use follows client commitments, data-handling rules, and the review requirements below. AI output is working material. The assigned consultant remains responsible for work delivered to the client.

This policy applies to personnel, contractors, AI accounts, integrations, and automated workflows used for firm or client work. Where an engagement’s written terms are stricter, the engagement terms control. The policy owner is [ROLE]; its effective date and next review date are [DATE] and [DATE].

Tool approval

Policy clause: Personnel may use only AI tools and account configurations listed in [APPROVED-TOOL REGISTER] for client work. Approval is limited to the recorded purposes and data classes. Personal accounts, browser extensions, meeting bots, and connected assistants require separate approval, even if the underlying model is approved.

Before listing a tool, [SECURITY OR OPERATIONS ROLE] reviews its data-processing terms, retention and training settings, security controls, access controls, deletion process, integrations, and relevant third-party vendors and subcontractors. The review considers available model lifecycle documentation; material model changes trigger another review. [LEGAL OR PRIVACY ROLE] checks compliance controls, client consent, and contractual requirements. A material change to the product or its settings triggers another review.

Policy template: client data and acceptable use

Input rules and confidentiality

Policy clause: Under the firm’s data governance rules, personnel shall use the least client information needed for an approved task. They shall remove unnecessary names, personal details, credentials, pricing, proprietary files, and identifying metadata before submission. Sanitized text may still identify a person or client when unusual details appear together, so the submitter must inspect it.

Confidential or restricted information may enter an approved environment only when [DATA OWNER] confirms that the engagement, provider arrangement, and data-processing terms permit it under data privacy rules. Personnel shall not enter client material into public or unapproved AI tools. If approval is unclear, work stops until the project lead resolves it.

Permitted and prohibited work

Policy clause: Within approved tools and data limits, acceptable use includes outlining proposals, organizing permitted research, drafting nonfinal text, suggesting code, or identifying questions for review. Personnel may not present generated facts as verified, fabricate citations, upload prohibited data, or let AI make an unreviewed client recommendation.

AI shall not independently send client communications, alter client systems, or publish deliverables. Recording or analyzing a meeting requires the permissions specified by [CLIENT AGREEMENT AND MEETING POLICY]. A meeting invitation by itself does not establish permission to process its transcript with AI.

Policy template: risk tiers and deliverable review

Classify the task before use

Policy clause: The project lead records each recurring AI use case in [USE-CASE REGISTER], including its owner, tool, data class, client, intended output, and approval status. The lead assigns a risk classification using the following criteria. The risk management process moves uncertain use cases to the higher tier pending review.

Tier	Consulting example	Approval before use
Low	Drafting from public sources without client identifiers	Project lead, using an approved tool
Elevated	Analyzing permitted internal project material	Project lead and data owner
High	Processing sensitive records or supporting a material client decision	Project lead, data owner, and [LEGAL/SECURITY ROLE]

These are firm workflow tiers, not legal classifications under the EU AI Act. They do not authorize unreviewed automated decision-making. The firm considers its ethical standards as an additional review consideration. [POLICY OWNER] documents additional restrictions when an engagement or applicable law requires them.

Check client-facing output

Policy clause: A named reviewer provides human oversight and is accountable for checking each AI-assisted deliverable before release. The reviewer compares material claims with original sources, checks calculations and quotations, examines confidentiality risks, considers bias mitigation, and assesses whether conclusions follow the evidence. AI-generated citations alone do not verify a claim.

For high-tier work, [SENIOR REVIEWER ROLE] approves the recommendation and records unresolved limitations. The project lead also checks disclosure requirements in the client agreement. If the work calls for disclosure, the team describes AI’s role accurately without suggesting the tool independently validated the result.

Policy template: ownership, records, and incidents

Assign decisions to people

Policy clause: The project lead owns use-case approval and delivery quality. The data owner authorizes access to client information. Security approves technical controls; legal or privacy staff assess relevant obligations; the policy owner maintains the register and exceptions. [EXECUTIVE OR BOARD COMMITTEE] receives periodic reports on high-tier uses, incidents, and overdue actions as part of board oversight.

One person may hold more than one role in a small firm, but every approval must name its decision-maker. The board or its delegated committee oversees whether management’s controls work; it does not replace project-level review.

Keep a usable record

Policy clause: For elevated and high-tier use, the team retains the use-case decision, tool approval, input classification, source references, reviewer, exceptions, and final approval in [APPROVED RECORD LOCATION] for [RETENTION PERIOD]. It follows client deletion instructions and restricts record access to authorized personnel.

Personnel use [INCIDENT CHANNEL] to report suspected data exposure, unauthorized tool use, or harmful output within [REPORTING WINDOW]. This starts the incident response process. The incident owner preserves relevant records, limits further use, assesses notification duties with counsel, and documents corrective action. Exceptions require a named approver and expiry date.

Make the policy work during a live engagement

Catch unapproved use and excessive access

An approved-tool list needs a matching access process. Use access controls to give staff only the AI tools and integrations their projects require. Where available, use managed accounts, role-based permissions, multifactor authentication, and logs that show relevant activity without creating unnecessary copies of client content.

Project leads should reconcile their use-case register with procurement records and available account or network activity. This continuous monitoring can reveal unreviewed extensions and personal-account use, including shadow AI, but won’t catch every instance. A fast approval path lets teams raise new use cases before deadlines push them toward shortcuts, supporting policy enforcement.

Watch what enters and leaves connected systems

Connected assistants can retrieve files, read web pages, and sometimes act through other tools. Treat retrieved text as source material, not instructions. Limit what an assistant can access, test suspicious inputs, and require human approval before it sends messages or changes records.

OWASP’s prompt-injection guidance recommends measures including least privilege and human approval for high-risk actions. Sample outputs for unsupported claims and sensitive-information disclosure, then log failures against the affected use case. Monitoring reduces risk; it does not guarantee that every unsafe prompt or answer will be caught.

Use standards as a map, not a substitute for decisions

Connect risk work to NIST and ISO

The voluntary NIST AI Risk Management Framework (NIST AI RMF) organizes work into Govern, Map, Measure, and Manage. A consulting team can govern through named owners, map tools and use cases, measure output failures, and manage them through restrictions and incident review.

NIST’s Generative AI Profile addresses risks such as inaccurate output and sensitive-information disclosure. ISO/IEC 42001:2023 provides a management-system structure for maintaining AI controls. Neither reference replaces an engagement-specific confidentiality decision.

Check regulatory duties by use case

The EU AI Act has staged obligations, not one rule for every consulting task. For regulatory compliance, check your role, system category, and applicable dates against the official implementation timeline.

Have counsel assess a regulated use before deployment. A firm’s internal “high” tier prompts closer review, but it doesn’t establish whether the system meets a legal definition of high-risk AI.

Short implementation checklist

1. Name the policy owner and replace every bracketed field, including decision-makers, record location, and reporting window.
2. Inventory current tools, integrations, accounts, and recurring client use cases. Remove or pause unapproved access.
3. Review provider terms and settings alongside client agreements. Record which data classes each approved configuration may handle.
4. Test the approval, output-review, and incident processes on a real engagement before rolling them out firm-wide.
5. Train staff with examples of safe prompts, prohibited inputs, unsupported claims, and suspicious retrieved instructions. Include periodic policy reviews in board oversight, especially after material tool, contract, or regulatory changes.

FAQ

Can consultants use a business AI account for confidential client work?

Only if the specific account and workflow are approved, and the client agreement permits the processing. Confirm provider terms, retention, permissions, integrations, and any required client consent. Business-account settings don’t cancel confidentiality duties or replace human oversight of client work.

Does every AI-generated sentence need a source citation?

No. Material facts, figures, quotations, and recommendations need evidence appropriate to the deliverable. Keep source references during drafting, and have a person check them against the original documents before circulation.

Conclusion

That meeting note should reach an AI tool only after the team knows the client permits the workflow and the data belongs there. A usable AI governance policy makes that decision clear before anyone pastes it.

Keep approvals close to the work, verify client-facing claims, and give each exception a named owner. The policy earns its place when a consultant can follow it under deadline pressure.