Decide when to tell a client about AI assistance by checking the contract, material contribution, data exposure and review, with clear disclosure examples.
Direct answer: Disclose AI use when the contract or applicable requirements call for it, or when the contribution would reasonably affect the client's assessment, approval or use of the work. Explain what AI did, what information it received and what a competent person checked. Trivial assistance may not justify a separate notice when no requirement applies, but confidentiality, substantive generation and automated decisions need an explicit conversation before the work is done, not a vague label afterwards.
Disclosure is not a contest to count every machine-assisted keystroke. It should give the client information that changes a decision: whether to approve the method, trust a claim, check rights, or permit a particular data flow.
Nor does a disclosure repair an unauthorised upload or make weak work acceptable. You remain responsible for the agreed deliverable. This article develops the accountability question in adopting AI without losing trust into a practical decision about client communication.
Applies to: client-service work using a UK starting point. Legal duties vary by jurisdiction, sector and contract. For a specific dispute, regulated assignment or uncertain obligation, obtain advice from an appropriately qualified local professional.
Apply the material-contribution disclosure test
The material-contribution disclosure test is an editorial method, not a legal definition of materiality. It asks whether the AI use changes the client's understanding of the agreed work, its evidence, its risks or the process used to create it.
Start with the contract and project instructions. Read the actual wording on AI, subcontracting, confidentiality, information processing, originality and approval. Do not assume a clause permitting ordinary software use permits every external service or automated action.
Next, describe the contribution without promotional language. Did the system correct spelling, propose an outline, produce a substantive first draft, analyse records or generate an asset? Identify what survived into the deliverable, but also what data or systems were exposed during the process.
Finally, ask whether the client would need different checks or permission if they understood that contribution. An AI-generated illustration used as decoration raises different questions from an image presented as documentary evidence. The same amount of generated material can have very different consequences.
My default recommendation is to disclose substantive contributions briefly and specifically, even when the contract is silent. The case against routine disclosure is strongest for minor assistance that does not change meaning, evidence, rights or data handling. Flooding a client with meaningless labels can obscure the use that actually matters.
Establish the requirement before choosing the wording
Contract-specific disclosure is not hypothetical. UK central-government procurement guidance includes example questions about suppliers' use of AI in tender responses and service delivery. Its scope is defined public bodies and procurement arrangements, not all private client work. Read the requirements in your own tender or contract rather than copying a universal rule from that example. Cabinet Office PPN 017.
If the wording asks whether AI assisted any part of a submission, answer that question accurately. Do not substitute your preferred threshold of substantial use. If a clause is ambiguous, clarify it before relying on a favourable interpretation.
For public-facing work, also check the destination's requirements. A client's approval does not automatically settle a publisher's rules, licensing conditions or professional obligations. Where the outcome depends on law or contractual interpretation, identify the relevant jurisdiction and get qualified advice.
Keep the agreed answer with the project records. It should remain available to the person who delivers the work if the original account manager is absent.
Distinguish contribution from exposure
Use two separate descriptions: what AI contributed to the output and what it received during production. A small final contribution can still involve a substantial data-handling decision.
| Illustrative use | Contribution question | Conversation needed |
|---|---|---|
| Spelling suggestions on public text | Did meaning or evidence change? | Check whether the agreement requires all assistance to be declared |
| Generated draft of a client recommendation | Which reasoning and claims came from it? | Explain substantive assistance and independent review |
| Analysis of confidential project records | What records left the approved environment? | Establish permission and applicable terms before processing |
| Generated image for a factual case study | Could a reader mistake it for real evidence? | Agree presentation, labelling and rights review before use |
This is not a claim that one label resolves each situation. If you cannot identify the data terms, do not upload the records while you work out how to describe the result. If a client forbids a proposed method, use an allowed method or renegotiate before proceeding.
Likewise, “reviewed by a human” must mean something concrete. State the checks actually completed: figures reconciled with supplied records, references opened and matched to claims, or generated imagery checked against the intended representational limits. Do not imply technical validation or rights clearance that nobody performed.
The Home Office's engineering guidance offers an institutional example of making AI-assisted changes traceable alongside normal review processes. It is specific to its engineering context, but illustrates the distinction between declaring assistance and verifying the work. Home Office AI guidance.
Write a disclosure that supports a decision
Use plain information, not a defensive disclaimer. A disclosure should identify the part of the work, the role of assistance, the data boundary and the review that actually happened.
For a hypothetical project, wording might be:
> An AI writing assistant helped create the initial outline and alternative section headings. The analysis and final recommendations were checked against the supplied source documents by the project editor. No client records were submitted to that assistant.
Use that wording only if every statement is true. It is an illustrative example, not a claim about a completed project. If client records were supplied under an approved arrangement, describe that arrangement accurately rather than retaining the final sentence.
For a proposed use, speak before the action: identify the service, account arrangement, information involved and intended review, then request the required approval. Do not present a past upload as a future option.
Avoid a blanket statement that the team “may use AI” if the client needs to approve a specific process. Also avoid suggesting that the client now bears responsibility for errors because assistance was disclosed. The disclosure should make your accountability clearer.
Budget a proportionate review across a project batch
Imagine an agency preparing seven deliverables. These categories and timings are illustrative assumptions, not legal classifications or measured results.
After reading the agreements, it identifies three deliverables with an explicit disclosure requirement, two where substantive assistance warrants a voluntary explanation, and two involving only trivial help with no identified requirement. The total is 3 + 2 + 2 = 7.
Suppose checking each deliverable's method and agreement takes six minutes: 7 × 6 = 42 minutes. Preparing a specific note for the five disclosures takes seven minutes each: 5 × 7 = 35 minutes. Two cases need an additional 12-minute internal clarification: 2 × 12 = 24 minutes.
The planned effort is 42 + 35 + 24 = 101 minutes.
That is coordination effort to include in the job, not time saved by AI. It excludes external legal advice if a case requires it. A shared record of already agreed methods may reduce repeated clarification on future deliverables, but you still need to check whether this work fits that agreement.
The calculation also shows why counting only drafting speed is incomplete. Any evaluation of the method should include approval, verification and communication, not treat them as invisible costs someone else absorbs.
Resolve the next deliverable before handover
- Before starting the next assignment, read the AI-related and data-handling instructions. Record unresolved requirements and ask the appropriate person to clarify them.
- List the actual or proposed contribution and information exposure for that deliverable. Stop unapproved processing rather than relying on a later disclosure.
- Before client review, verify the work and prepare a short, accurate account of the assistance where required or material. Have the project owner check that every stated review occurred.
- At handover, give the client the information needed for their next decision and retain the agreed method with the project record. Review the arrangement when the service, data or use changes.
If you discover that the work already breached an agreed restriction, escalate promptly through the relevant contractual or incident process. Do not disguise the problem as an ordinary editorial note.
Related guides
Frequently asked questions
Do I have to disclose ordinary spelling correction?
Check the actual requirement first. If a contract asks for any AI assistance to be declared, do not decide unilaterally that spelling correction is too minor to count. Where no such requirement applies, minor assistance that changes neither meaning nor evidence may not need a separate client notice under this article's editorial recommendation. The data-handling question still exists: processing confidential text in an unapproved service is not made trivial by the small correction it produces. If the wording or applicable professional rule is unclear, clarify it before using the service.
Should I name the model or just the application?
Name the information that is relevant and that you can establish accurately. The application and account arrangement may matter more to data handling than the underlying model name. If a model identity is visible and relevant to a technical record, include it with the date, but do not guess because a product is commonly associated with a particular provider. A client requesting more detail may need the actual supplier terms or processing arrangement. Avoid a long technical inventory when a clear description of the contribution and checks answers the real question.
What if the client says they do not care how the work is made?
That can simplify the editorial conversation, but it does not remove every other condition. Confirm what their statement covers, particularly confidential information, third-party processing, rights and any downstream publication requirements. Keep the agreement with the project record rather than relying on a casual remark remembered differently later. You should still verify the deliverable and remain accurate about the method if asked. If the client lacks authority to approve a particular data flow or regulated process, their indifference does not supply that missing authority. Use the appropriate approval route for that decision.
Can I disclose AI use after delivering the work?
Sometimes a final method note is the agreed timing, but it is too late when permission was required before processing or production. Separate approval from reporting. If you already used an unapproved method, tell the responsible project or compliance person promptly and establish the appropriate response rather than backdating permission. The consequences depend on the agreement, information and jurisdiction. For future work, agree disclosure points during planning so the client can make a real choice. A late label should not be used to avoid a conversation that might have changed the method.
Does disclosure mean I should charge the client less?
Not automatically. Pricing depends on your agreement and what you are being paid to provide, such as time, a defined deliverable or an outcome. Do not bill for hours that were not worked under a time-based arrangement, and do not claim a review or specialist contribution that did not happen. A fixed-price assignment may include preparation, verification and responsibility beyond drafting. If the method materially changes an agreed staffing or production commitment, discuss that rather than assuming the price is unaffected. Obtain qualified advice where the contractual interpretation is disputed.
How should I handle AI-generated images or audio?
Explain the generated nature when it affects how the client or audience will interpret the asset, and check the rights and destination requirements before use. A decorative illustration is different from an apparent photograph of a real event, and a synthetic voice must not be presented as an actual recording without an appropriate basis. This article does not establish licensing or consent for a particular asset. Review the applicable service terms, source permissions, representations and jurisdiction. If a claim of authenticity matters to the assignment, do not rely on a generic AI-assisted label to resolve it.
Sources and verification
- Cabinet Office: PPN 017 on AI transparency in procurement, checked 11 September 2026 for its scope and example disclosure questions. No universal private-sector disclosure duty is inferred.
- Home Office: Use AI, checked 11 September 2026 for an institutional example of traceability and review.
- The parent was read in the supplied website source because its public route could not be retrieved. The decision method and agency scenario are editorial examples, not legal advice or reported client experience.
This article is practical guidance. Apply it in proportion to your tools, evidence, risks, and responsibilities.
