Back to Guides
Enterprise Software 15 min read

Should you tell clients you use AI? The disclosure decision

Should you tell clients you use AI? The disclosure decision

Yes in the cases that matter, and it will cost you something measurable. Across 13 experiments published in 2025, people who disclosed their AI use were trusted less than people who stayed quiet. Disclose anyway: the same research found that being exposed by somebody else does more damage than telling the client yourself.

AI client disclosure gets argued as a question of values, which is why it never gets settled. Priced as a decision, it resolves in about an hour: sort which of your AI uses are in scope, say so early enough that it reads as method rather than apology, and make sure every client-facing person gives the same answer when asked on a call.

The evidence nobody in this argument cites

Almost every published page on this subject asserts that transparency builds trust. The strongest study available found the opposite.

Oliver Schilke and Martin Reimann ran 13 experiments, published in Organizational Behavior and Human Decision Processes in May 2025 (volume 188, article 104405). They tested disclosure across written work, analysis and creative output, and across both people and organizations. Those who disclosed AI use were trusted less than those who said nothing. The driver was perceived legitimacy rather than dislike of the technology: disclosure made the discloser look like somebody who had done a thing that needed explaining. The effect survived changes in wording, survived cases where the audience already knew AI was involved, and held whether disclosure was voluntary or required.

Then comes the finding that settles the decision. The trust penalty for disclosing is weaker than the penalty for being exposed by a third party. Telling your client costs you something; having them work it out, or hear it from a competitor or a departing employee, costs more.

A second study explains how exposure happens. Angel Hsing-Chi Hwang and colleagues, in work on freelance engagements posted in March 2026, found that workers disclose only when asked, assuming clients can already tell, while those clients are much less sure they can spot AI-assisted work and would rather be told first.

So the question is where the disclosure sits, not whether to be transparent. Early, inside a conversation about how the work gets done, it reads as method. Late, after a client raises it, the same sentence reads as a confession.

Where the choice has already been made for you

Four things can take this decision out of your hands before you make it.

What binds you What it asks for Status as of October 2026
Your client’s own contract Notice before use, written approval of named tools, or a flat ban. Services agreements and vendor codes increasingly carry one In force wherever you signed one: the most common binding instrument, and the least often read
EU AI Act, Article 50 A person interacting directly with a system must be told they are dealing with AI at the start, unless it is obvious to a reasonably well-informed person Applicable 2 August 2026, with Commission guidelines published 20 July 2026. Penalties reach €15 million or 3% of worldwide turnover
State transparency law Utah’s AI Policy Act asks for disclosure in state-licensed occupations during high-risk interactions, and on a clear consumer request Utah’s act took effect May 2024; S.B. 226 narrowed it from 7 May 2025. Several states carry chatbot rules
Your professional body ABA Formal Opinion 512 (29 July 2024) asks a lawyer to get informed consent before putting client-representation information into a self-learning tool Binding on licensed practitioners. Accounting has no equivalent mandate, though Gramm-Leach-Bliley and Treasury Regulation 301.7216 reach the same material

Two of the four catch people out. The EU regulation reaches deployers outside the Union whose system output is used inside it, so a US firm running a client-facing assistant for a European account can sit inside Article 50 with no office there. And if you signed anything carrying an AI clause in the past two years, your customer wrote your disclosure policy and nobody has read it since.

Four different things “we use AI” can mean

The phrase is doing too much work. Pull it into parts and the duty sorts itself out.

What is happening What the client is owed Why
A bot or voice agent speaks to the client directly Tell them at first contact, every time, in the interface itself The one case with a hard legal duty behind it. Nobody should have to work out whether they are talking to a person
Client material goes into a tool Consent rather than notice, and check what you signed first A confidentiality question wearing a disclosure costume, and the two have different answers
AI output ships as a deliverable with little human change Tell them, and name who reviewed it The client bought your judgment. If none was applied, what they bought was not delivered
Staff draft internally and a person rewrites the work and owns the result Usually nothing specific, plus a plain answer whenever asked The tool is a method here, like a research database

Collapsing those rows into one question produces both failure modes. Firms that treat everything as row one disclose so much that clients wonder what else is automated. Firms that treat everything as row four ship unreviewed output under a human byline and lose an account over it.

Row two holds most of the real exposure, and disclosure is the wrong instrument for it: notice is something you give, consent is something you get. What can go into a tool at all is a separate question of client confidentiality.

The five questions that settle it

These five decide what to say, and they work without a professional body behind you.

  1. Does anything we signed already answer this? The clause you find outranks your preference. Start with what your existing agreements say about client material and AI tools.
  2. Does client material leave accounts we control? If yes, this is a consent question, and consent is asked for in advance rather than mentioned afterward.
  3. Could the client learn this from anyone but us? A subcontractor, a departing employee, a competitor pitching against you, a document’s own metadata. Where the answer is yes, you are better off first.
  4. Would the deliverable be the same if a person had produced it alone? When the honest answer is no, the client is buying something other than what they think, and that is disclosure territory regardless of any rule.
  5. Does our fee assume hours the tool has removed? Then the disclosure conversation is already a pricing conversation.

Four noes describe most internal AI use at a services firm, where the right disclosure is an honest answer on request rather than a banner on every invoice. One yes changes the shape of the engagement.

Why the clause in your contract is the weakest version

Most advice on this question ends at adding a clause to the engagement letter. The clause is worth having and will not do the job alone.

ABA Opinion 512 is blunt about why: a boilerplate acknowledgment or notice provision does not achieve informed consent, which requires that the client got enough explanation to make a real decision. The Journal of Accountancy carried the same reasoning into accounting practice in April 2025, in a piece by CNA risk consulting director J. Michael Reese that found no federal law or CPA standard mandating disclosure, then recommended plain-language disclosure anyway.

That logic does not depend on holding a license. A sentence in clause 14.3 of a document a procurement manager signed eleven months ago is a record, and records are useful. It is not a shared understanding with the person who will be annoyed in March. Keep the clause and date it, then put the substance in the kickoff conversation, from the person who owns the relationship.

What the sentence should say

Three patterns fail predictably. The tool list (“we use ChatGPT, Claude, Gemini and Microsoft Copilot”) gives a client nothing to act on and dates within a quarter. The apology (“we do use some AI, but rest assured every word is checked by a human”) is the exact register the trust research says gets punished. The hedge (“we may use AI tools from time to time”) says nothing and still carries the full cost of raising the subject.

What works is three facts in plain words: where the tool sits, who is accountable for the output, and what never goes near it.

We use AI assistants for first drafts, research summaries and analysis. Every deliverable is reviewed and signed off by the named lead on your account, who is accountable for it. Your confidential material stays inside our business accounts, configured so your content is not used to train anybody’s model, and nothing goes into a personal account.

Under sixty words, no apology, and it answers the three questions a client has. The third sentence earns the most trust and is the one most firms cannot say truthfully, because it depends on account tiers somebody has to verify. The gaps between training, retention and human review commitments are rarely what the sales conversation implied.

The fee question hiding inside the disclosure question

One trigger turns a quiet internal decision into a client conversation, and it is missing from nearly everything written on the subject. ABA 512 holds two things about money: a lawyer cannot bill for hours the tool removed, and cannot pass AI subscription costs through as a disbursement without disclosure and agreement. Both generalize into any services business.

If you bill hourly and a tool cut a ten-hour task to four, the invoice already discloses something. If you bill a fixed fee scoped on the old assumption, you are keeping the gain, which is defensible and gets harder to defend once a client asks. And if you have told a client that AI makes you faster while your rate card has not moved, expect the next renewal to open there.

None of that argues for giving the margin away. It argues for deciding your position first, because the difference between “our pricing reflects outcomes, not hours” and an improvised answer from an account manager is the whole of your credibility. Procurement teams and insurers now ask in writing, the same shift visible in the questions clients put about AI controls.

Forty people, one answer

Disclosure does not happen in your contract. It happens at 3pm on a Tuesday, when a client asks “be honest, did a person write this?” and a project manager with no script answers on instinct.

That makes this a training problem rather than a drafting problem. At a firm with forty client-facing people the policy can be excellent and still produce four different answers across four accounts in a week, which damages trust more than any single answer would, including a blunt one. Inconsistency is what reads as something hidden.

In the sessions we run, the question that produces the longest pause is never what the policy says. It is who signs off. Three things have to live in people’s heads, not in a document:

  1. The sentence. The short version above, in their own phrasing, delivered without flinching.
  2. The boundary. What never goes into a tool, named specifically: this client’s unredacted files, anything under a named NDA, personal data of their customers.
  3. The escalation. Who to route a client to when the question goes past the sentence, plus permission to say “let me get you the precise answer” without treating it as a failure.

Half a day of practice gets a client-facing team there, and a written policy alone never does. It is the same gap that shows up across this field, where the document exists and the behavior does not, the subject of the safe AI use playbook. Market rates for hands-on AI enablement programs run roughly $2,000 to $15,000 by scope and headcount, a modest line next to one lost account. Licenses, clauses and settings are purchasable inside a quarter; the judgment to use them well is built, the argument in why buying licenses is not the same as building capability.

One sitting gets you a defensible position: search your largest agreements for “artificial intelligence” and “automated”, list every place AI touches a client directly including the meeting recorder everyone forgets, have whoever runs delivery write the three sentences, brief the team, then date it and diary a quarterly review.

The firms that handle this well are not the most transparent ones. They are the ones whose answer is the same in the proposal, on the kickoff call, and in month seven when something goes wrong. For a second read on how your own AI use would stand up to a client asking directly, book a free AI-readiness call.

Frequently asked questions

Do we have to tell clients we use AI?

Only sometimes. A hard duty applies when AI interacts with the client directly, under EU AI Act Article 50 and several state laws. Beyond that it turns on your contracts, your profession, and whether unreviewed output reaches the client. No general US rule covers internal drafting a person rewrites and owns.

Does a clause in our contract count as telling them?

As a record, yes. As consent, no. ABA Formal Opinion 512 states that a boilerplate acknowledgment in an engagement letter does not achieve informed consent, and the same logic holds outside the law.

What do we say when a client asks on a call?

Three facts in plain words: where the tool sits, who is accountable for the output, and what client material never goes near it. Do not lead with a tool list and do not apologize. Twenty seconds, in the register of someone describing a method.

Does using an enterprise account mean we do not have to disclose?

It removes one reason to, not all of them. A business tier that excludes your content from training weakens the confidentiality argument for consent, which is what ABA 512 keys on. It does nothing about a bot that talks to your client, an unreviewed deliverable, or their own contract.

Should we name the specific tools we use?

Be ready to, and do not lead with it. Clients asking this want to know whether their material is safe and who stands behind the work; a tool list answers neither.

Does AI use change what we can bill?

It can. ABA 512 holds that a lawyer may not bill for hours a tool removed, and may not pass AI subscription costs through without disclosure and agreement. No equivalent rule applies outside licensed practice, but the same conversation arrives at renewal.

What if our client’s contract bans AI use outright?

Then you comply or you renegotiate, rather than quietly interpreting your way around it. Flat bans are often broader than the client intended. Ask for a scoped amendment naming what is allowed.

Does the EU AI Act apply to us if we are a US company?

It can. The regulation reaches providers and deployers outside the Union whose system output is used within it, so a client-facing assistant serving European users can fall inside the Article 50 duties, applicable since 2 August 2026, with fines up to €15 million or 3% of worldwide turnover.

Who should own this internally?

One named person in delivery or operations, not a committee and not legal alone. Legal reviews the wording; delivery owns the answer, because the answer has to match how the work is produced.

Key takeaways

  • Voluntary disclosure carries a real trust cost. Across 13 experiments published in 2025, disclosers were trusted less than non-disclosers, driven by perceived legitimacy rather than algorithm aversion.
  • Being exposed by a third party costs more, which is why silence is a bet rather than a policy.
  • Check what already binds you first: client contracts, EU AI Act Article 50 since 2 August 2026, state transparency law, and professional rules if you hold a license.
  • “Using AI” covers four situations with four duties. A bot talking to your client is near-absolute; internal drafting that a person rewrites and owns is close to zero.
  • A clause in the engagement letter is a record, not consent, and the substance belongs in the kickoff conversation.
  • The policy is worth nothing until every client-facing person can say the same three facts without flinching, which is a training outcome rather than a drafting one.

Last Updated: Nov 16, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles