How to Tell Your Client You Used AI

The disclosure nobody prepares for

There is a good body of writing on whether you must declare AI use to a journal or an HTA agency. We have written some of it ourselves. What almost nobody addresses is the conversation people actually dread, which has no rulebook at all: the client asking "did you use AI on our dossier?" And "our policy is not to use AI" does not cut it any more, because a large proportion of professionals admit to using it in a shadow form. Nor does "we only use Copilot on redacted data", which is not a final claim if it cannot be demonstrated.

A journal, and some agencies, tell you what to include as a disclaimer. A client just asks. And because there is no rule to point at, the answer tends to be improvised at the worst possible moment. We simply need a routine way of presenting the reality: AI is here to stay, so let us make it safe, and keeping a proper trace with a deep audit is the least we can do.

This page is about the wording. Which bodies require a declaration, and where it goes in a submission, is a separate guide.

The agencies asking you to declare are disclosing too

This is the part that changes how the conversation feels. Disclosure reads like confession mainly because people imagine they are the only ones doing it. They are certainly not, the bodies reviewing their work are using AI and publishing their own use, sometimes in their own words and on their own sites.

The pattern in all of them, which is your template

Read those disclosures side by side and they share a structure, or at least a common direction. Each one names tools, states the scope of use, states the limits, and keeps a human accountable. None of them apologises for using productivity tools.

We can use that. Be conservative, plan for a more regulated future, and build the template out of the regulators’ own words, which gives it a property no wording you invent can have: nobody can argue it is insufficient, because the body that would judge you wrote it. And it makes the point we would rather the field settled on: judging the user of AI should not be anybody’s goal. Preventing misuse, and the disclosure of secret and sensitive data, should be.

Four worked examples

Each of these answers the same four questions, which tool, used for what, what it did and did not receive, who is accountable, in the register the moment calls for. And each of them has to be backed by a complete audit; ours is the Untraceable Protocol.

1 · In the engagement letter, the SOW and if possible the MSA, before work starts

The best way to start is to prepare the field, and it is also the cheapest. Consultants are used to declaring subcontractors in a SOW, and AI is not so different: it sees the data when you do not redact or cloak it, it contextualises your request, and it stores, memorises and screens the content for abuse. At this point the disclosure is a description of your methodology that the client accepts, so it never becomes a disclosure at all.

"[Firm] uses AI tools for templating, drafting, summarisation, augmentation and analysis support. Client-confidential values are not transmitted to any third-party AI service. Where AI is used on material derived from Client information, confidential values are replaced before transmission, and automated breach and coverage tests are passed to confirm that no sensitive value or secret was transmitted. All AI-assisted output is reviewed and approved by the named author, who remains accountable for it. A record of AI use is retained and available to Client on request."

2 · When a client asks mid-project

The question the client is actually asking is about the possible disclosure of their information, not about whether you used AI. It sounds like "did our data leave your computer" and gets phrased as "did you use AI". Clients know by now that when a firm bans AI, people find a way to use it anyway, so their concern is the use of proper tools, not prohibition.

"Yes. We used [tool] for [drafting the background section / restructuring the economic narrative]. Your confidential values were not transmitted, because we replace every secret and sensitive value before anything is sent. Everything was reviewed by [name], who is accountable for the content. I can send you the audit of what was submitted if that is useful."

3 · To a subcontractor or partner, which is where the exposure hides

Your obligation to your client covers everyone you delegate to. A disclosure you make confidently is worth very little if a subcontractor pasted the same document into a consumer chatbot.

"AI tools may be used for drafting and analysis support on this engagement. Client-confidential values, sensitive values and secrets must not be transmitted to any AI service, including enterprise services, unless we have authorised that specific system in writing. You must keep an audit of AI use on deliverables you provide to us for at least [##] years after the engagement, recording the tool used, the purpose, the risk assessment (which values were removed from the document) and the reviewer."

4 · The version where the architecture writes the answer

This is the disclosure we recommend when working with Untraceable and the Untraceable Protocol. When the model never receives a confidential value, secret or sensitive value, the declaration and the reassurance become the same sentence.

"AI assisted the drafting of this document, including templating, augmentation and analysis support. Confidential, secret and sensitive values were replaced with semantic and sequential markers, cloaks, on our own systems before any text was transmitted, so no confidential value was disclosed to the AI provider. The author reviewed and approved all output and remains accountable for it. A deep audit of what was cloaked is accessible through direct and free access to the Untraceable Protocol, which includes the tooling and providers used, the risk assessment, the privacy impact assessment and the signature page."

Note what these sentences do not do. They do not claim AI was not used. They do not ask anyone to trust a third party’s promise never to train on your data, or to delete its memory. They disclose the use and answer the objection in the same breath, which is the whole argument for building the record before anyone asks for it.

A view of my own: the saving should be shared

I will go further, at the risk of sounding ahead of my time. I think the disclosure should come with cost-sharing between the consultant and the client. If AI is used, the consultant gains productivity and delivers more, at better quality, and the client should see not only the quality but a quantifiable share of the gain. Where a client does not accept AI use, they should be able to say so and pay a premium for a fully human deliverable, 100%, no shadow AI, as in the old days. Where they accept it under scrutinised conditions, part of the saving should come back to them. I have put this on proposals myself, as a cost saving line. "If we are allowed to use AI, you should get some discount" is, I think, the right posture towards a client.

There is a reason I think this arrives whether anyone likes it or not, and it is not goodwill. For a large consultancy the revenue is either the hours (time and materials) or a fixed bid, with a pre-specified amount and scope, so sharing a saving means changing the way we work, and partially the way we bill. Adjusting fixed-bid projects is a no-brainer: the margin on the project rises with AI, so that difference can be shared. For time-based consulting it can be both an institutional blocker and, in some cases, feel like a cultural one, but it is fully related to the business model. Smaller boutiques do not carry that conflict: the saving is their margin, and giving part of it back is how they win the work. So the gains get taken at the small end first, quietly or openly, and the large firms end up adapting rather than choosing. Fewer hours per deliverable stops being a philosophy and becomes a condition of competing.

The part I think is clearly misread is what happens to the hours. People hear "fewer hours" and assume a smaller amount of work, a weaker pipeline, a rough year in consulting. What I see is a different type of work, an evolving one. The hours AI actually removes are typically the least valuable ones in the engagement, rereading the same deliverable because one input moved, reconciling versions, reformatting to match a template, finding the typos, rewriting to satisfy strict internal guidelines, templating the skeleton of the report, early pearl growing as you iterate through the literature to find inputs. Removing those does not thin the work, it concentrates it in a very interesting way: more time on strategy, deeper argument, and a deliverable carrying more human intelligence than ever, because the consultant was able to focus on depth instead of formatting.

Which is also why I do not think the budget shrinks in the long run. It simply moves. If production costs less, the client can finally fund what kept getting cut, the quality-of-life study nobody had room for, the ex-payer discussion that would have pressure-tested the strategy before it was submitted rather than after, the manuscript or white paper instead of the budget poster. Total spend will hold in the long run, probably with more deliverables and more quality; what falls out of it is the reformatting and the iterating. That is a better engagement for everyone in the room, and it is the version of this future I would like us to have helped along.

One aspect deserves its own guide, which we have not written yet: training and the retention of junior staff. The reduction in hours falls disproportionately on them, and that has to be faced honestly rather than waved through, though I believe the evolution ends up positive for them too.

What not to do

• Do not say "we used AI" and stop. An unscoped disclosure invites every question a scoped one answers.
• Do not offer the vendor's terms as your answer. A no-training commitment binds the vendor to you; it is not your client's permission. A real AI compliance officer will know the difference.
• Do not disclose and then hedge. "Only for minor edits" invites an audit of what counts as minor, and of whether there were confidential values in the minor thing.
• Do not assume redacting or cloaking excuses the declaration. It removes the data disclosure, not the duty to declare AI assistance. Those are different obligations to different parties, and the second one still stands. We are not here to reduce the burden of declaration. We are here to make it safe.
• Do not write a disclosure you cannot evidence with a deep audit. If you cannot produce the record when asked, the disclosure becomes the problem rather than the solution.

Related reading

• Do you have to disclose that AI wrote it? , which bodies require a declaration, and where the duty hides.
• Does AI break your NDA? , the data disclosure, which is the other half of this.
• An AI use policy you can adopt , what you decide internally, before any of this is said out loud.
• AI in US market access: what binds you today , why the duties arriving are documentation duties.