“How was this made?” is a practical question about a result: what shaped it, what you contributed, and what someone should know before relying on it. A handover can answer that question without pretending there is one universal disclosure rule or one record that fits every project.[1][2][3][7]
Use the client’s terms and the rules that apply to your work as the starting point. Then keep a concise, retrievable account of the process, your review, and any limits that matter for the next decision.[3][12][17]
Disclose where AI contributed, while keeping legal and sector guidance in scope
Guidance on AI disclosure varies by context, so explain your process clearly and follow the instructions that apply to your engagement.

Several sources discuss disclosure, but they address distinct settings. The American Bar Association’s ethics opinion concerns lawyers working under the ABA Model Rules. It says disclosure may be required when a client asks how the work was conducted or whether generative AI was used, or when engagement terms or client guidelines require it. The opinion is not a general rule for independent consultants.[1]
The UK Government’s framework recommends appropriate disclosure when AI produces an output and recording which parts of a process are automated and which are controlled by people. It is guidance for government and other public-sector organisations, not a universal private-consultant requirement.[2]
U.S. Copyright Office guidance has a narrower purpose: applicants seeking U.S. copyright registration must disclose AI-generated content in a submitted work and briefly explain the human author’s contributions. That registration requirement does not establish a general client-handover rule.[4]
These sources do not establish one disclosure rule for every independent consultant. In practice, read the engagement terms and applicable client instructions, then describe AI’s role in the work in a way that fits those requirements.[1][2][4][12]
Give the client a traceable account of the inputs and tools behind the result
A concise, work-specific record can show what systems and inputs shaped the result and what changed along the way.

NIST describes provenance metadata that can include information about a model or content creator, the time and location of creation, modifications, and sources. It presents these as information that can be included, not as a fixed, mandatory consultant template.[3]
OECD principles also discuss explaining sources of data and inputs, and making datasets and processes traceable. The appropriate level depends on context, feasibility, usefulness, and the state of the art.[6][7]
For a particular project, you might record the AI system used, relevant source material or inputs, the prompt or instructions, when the work occurred, and material edits or selections you made. Treat these as adaptable fields: include what helps explain or continue the work, while respecting client instructions and the needs of the project.[3][6][7][8]
For example, NIST’s health-data guidance describes a fuller provenance record for AI involvement in healthcare data exchanged using FHIR. It identifies the AI system, human and automated participants and their roles, inputs and prompts, and a link to the system’s model card. That guidance is specific to this health-data context; it is not a general handover standard for other work.[5]
Digital asset credentials offer another example of recording origins, modifications, and AI use. C2PA explains that this kind of provenance can describe an asset’s history, but records may be incomplete and do not establish whether the asset’s content is true or accurate.[8]
Describe your contribution without treating prompting as proof of authorship or quality
Say what you actually supplied, selected, arranged, edited, or changed; do not use a process record to claim more than it can establish.

A useful handover separates AI-generated material from the human work that shaped the deliverable. You can identify what you supplied, which outputs you selected, what you rearranged or edited, and which parts you wrote yourself. This describes the process without making a broad claim about authorship or ownership.[4][9]
The U.S. Copyright Office’s report discusses prompting, expressive material authored by people and included in a work, and modifications or arrangements of AI-generated outputs as different types of human contribution. Its scope is copyrightability under U.S. copyright law; it does not create a universal rule for documenting client work.[9]
A prompt may be part of the record, but by itself it does not explain every human contribution or prove the quality of the delivered work. Describe the choices and edits that actually affected the result, and keep any legal conclusions within the relevant jurisdiction and context.[4][9]
Tell the client what you checked and what the checks covered
Describe the review you performed and its limits; a record of where content came from cannot, by itself, prove that it is accurate.

When a deliverable contains claims, citations, or other source-dependent material, identify what you checked and how far the review went. NIST recommends reviewing and verifying sources and citations in generative AI outputs as part of risk measurement and ongoing monitoring. This is voluntary cross-sector guidance, not proof that every output has been checked or that every check will catch an error.[3]
A Stanford-hosted study gives a bounded example of manual checking in legal work. Researchers built a preregistered dataset of more than 200 legal queries and manually reviewed outputs from three tested systems for accuracy and fidelity to legal authority. The query count describes that research dataset, not all legal questions or systems.[10]
Keep two questions separate: what the record says about the output’s origin and history, and what evidence supports its factual accuracy. C2PA says provenance can provide evidence about creation and movement over time, but provenance alone cannot tell you whether content is true, accurate, or factual.[8]
For a client, this distinction makes a review note more useful. Name the sources or claims you checked, the method or scope of review, and any important areas you did not verify. A traceable process helps someone inspect the work; it is not a substitute for evidence supporting the result.[3][8]
Make assumptions, limits, and client choices visible
State the intended use and the conditions that affect how the client should rely on the result.

NIST recommends documenting the expected and acceptable context of use for a generative AI system. Its guidance includes considering assumptions, limitations, the intended environment, observed usage, potential impacts, and expectations. For a client handover, the practical point is to name the context in which you prepared the work and the assumptions that matter to its use.[3]
Tests and reviews have conditions. NIST advises documenting limits on generalising results beyond the conditions under which a technology was developed, and cautions against extrapolating from narrow or anecdotal assessments. A handover should therefore avoid implying that a check under one set of conditions guarantees performance in another.[3]
Client instructions may also shape whether and how AI is used. California State Bar guidance says lawyers must follow applicable client instructions or guidelines that limit AI use in a representation. That guidance concerns legal practice in California; it is not a rule for every profession or jurisdiction.[12]
Singapore’s Ministry of Law guide says clients may be offered an option to opt out of generative AI being applied to their legal matters. The guide is non-binding and applies to legal work in Singapore. It notes that opting out may mean losing AI-enabled efficiencies or capabilities, using traditional processes, and experiencing slower turnaround or higher costs; these are possible effects, not guaranteed outcomes.[11]
Record the client’s relevant choices and instructions, the intended use, and any assumptions or limitations that affect the result. Keep jurisdiction-specific guidance clearly labelled, and do not generalise a narrow review or legal-sector rule to every client engagement.[3][11][12]
Leave a practical trail for whoever uses or maintains the work next
For technical deliverables, give the next person enough information to set up, inspect, and continue the work.

For a software project, a handover can include the setup steps, dependencies that need to be installed, one-off commands, and a clear way to run tests or linting. GOV.UK’s README guidance offers these as practical topics for public application repositories and notes that its page may be out of date.[13]
Record meaningful code changes and the reasons for them so a future developer can understand the work. For npm projects, the package lockfile records the generated dependency tree so later installs can reproduce it despite intermediate dependency updates.[14][15]
Document deployment configuration separately from code where the deployment setup calls for it. The Twelve-Factor App guidance describes storing deployment-specific configuration in environment variables; it distinguishes that configuration from internal application settings that do not vary between deployments.[16]
Where AI documentation matters to the work, make it retrievable. NIST’s AI RMF Playbook lists topics such as scope, uses, assumptions, limitations, test results, dependencies, and plans for deployment, monitoring, and change management. It also says that storage and access procedures can help retrieve critical information during an incident. These are voluntary suggested practices, not a universal checklist.[17]
For work that continues in production, define who reviews it and when. NIST’s AI RMF Core describes planning ongoing monitoring and periodic review of risk-management outcomes; it does not present its actions as a required checklist or fixed sequence.[18]
If prompts are part of an application, keep prompt changes reviewable and recoverable. OpenAI’s API guidance recommends using version-control history, pull-request review, release tags, and feature flags for production prompt changes. This is provider-specific guidance for work using the OpenAI API.[19]
A practical handover can be short. Include the details that help this client understand the work, make a decision, or continue it safely; make the scope and limits of those details clear.[3][7][17]
- 01AI’s role. Identify where AI contributed, in line with the client’s instructions and applicable guidance.[1][2][12]
- 02Inputs and tools. Note relevant systems, sources, prompts, and material changes when they help explain the work.[3][6][8]
- 03Your contribution. Describe what you selected, arranged, wrote, or modified without claiming more than the record shows.[4][9]
- 04Review and limits. State what you checked, how far the review went, and what remains uncertain.[3][8]
- 05Next steps. Record intended use, relevant client choices, and practical setup or maintenance information for the next person.[3][11][13][17]
How do you explain AI’s role when a client asks how the work was made?
Take the survey →References
- American Bar Association, Standing Committee on Ethics and Professional Responsibility2024-07-29ABA Formal Opinion 512
- UK GovernmentUpdated 2025-12-18Data and AI Ethics Framework
- National Institute of Standards and Technology, U.S. Department of Commerce2024-07NIST AI 600-1: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- U.S. Copyright Office2023-03-16Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence
- National Institute of Standards and TechnologyCreated 2026-08-31; updated 2026-09-15HL7 AI Transparency: Fast Healthcare Interoperability Resources (FHIR)
- OECD.AI, Organisation for Economic Co-operation and DevelopmentDate not statedOECD AI Principles: Transparency and Explainability
- OECD.AI, Organisation for Economic Co-operation and DevelopmentDate not statedOECD AI Principles: Accountability
- Coalition for Content Provenance and AuthenticityVersion 2.3; date not statedC2PA and Content Credentials Explainer
- U.S. Copyright Office2025-01Copyright and Artificial Intelligence, Part 2: Copyrightability
- Stanford Law School hosts the paper; authors affiliated with Stanford University and Yale University2024-05-30Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools
- Ministry of Law, Singapore2026-03-06Guide for Using Generative Artificial Intelligence in the Legal Sector
- State Bar of California, Standing Committee on Professional Responsibility and Conduct2026; exact release day not statedPractical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law
- GOV.UK Developer DocumentationLast updated 2023-12-07READMEs
- GOV.UK Service Manual, Software engineering communityPublished 2016-05-23; updated 2016-12-02Maintaining Version Control in Coding
- npmnpm CLI v11.21.1 documentation; accessed 2026-11-01package-lock.json
- The Twelve-Factor AppDate not stated; accessed 2026-11-01Config
- National Institute of Standards and TechnologyDate not stated; accessed 2026-11-01AI Risk Management Framework Playbook
- National Institute of Standards and TechnologyAI RMF 1.0 (2023); page date not statedAI RMF Core
- OpenAIPage publication date not stated; accessed 2026-11-01Prompting