When a Client Finds Out You Used AI

When a client finds out you used AI and you did not tell them first, the instinct is to defend the work. Resist it. The work is rarely what they are actually upset about. Underneath that reaction sit three different complaints, they need three different answers, and giving the wrong one makes it worse.

Figure out which one you are facing before you say anything substantive.

What it means when a client finds out you used AI

Listen to how they phrase the objection. The wording usually gives it away.

"Did you put our data into it?" This is a confidentiality complaint. It has nothing to do with quality and everything to do with where their information went. It is also the most serious of the three, because it may be a contractual or regulatory matter rather than a matter of taste.

"So I paid you to press a button?" This is a price fairness complaint. They are not questioning the output, they are recalculating what they think it cost you to produce. Argue about quality here and you will miss entirely.

"Is this even any good, then?" This is a quality complaint, and it is the easiest of the three, because you can settle it with evidence.

The mistake almost everyone makes is answering the second with a defence of the first, explaining your data handling to someone who was only ever worried about the invoice.

Answering the confidentiality objection

Be specific and be fast. Vagueness reads as concealment.

Tell them exactly which tool, exactly what went into it, and exactly what the retention terms are. "I used the API, which does not train on submitted data, and the only client material I sent was the public product copy from your site" is an answer. "I was careful" is not.

Get the detail right, because a client who checks will find the nuance. OpenAI's data controls documentation states that API data is not used for training by default, but also that prompts and responses are retained in abuse-monitoring logs for up to 30 days, and that zero data retention is not a switch you flip, it requires prior approval from OpenAI. Claiming "nothing was retained" when you never applied for that is the kind of overstatement that turns a recoverable conversation into a credibility problem.

Then check the contract before the conversation, not after. If there is a confidentiality clause covering disclosure to third parties, a consumer AI tool may well count as one. If you find you did breach it, say so plainly in that same conversation. A disclosed breach handled badly is survivable. A concealed one found later is not. Handling a client contract with a no-AI clause covers the clause language itself, and whether it is safe to share customer data with ChatGPT sets out which tools treat your inputs how.

If actual personal data went into a tool that retains it, you are into breach-notification territory rather than client-relations territory, and the clock matters. The first 72 hours after an AI data leak sets out that sequence.

Answering the price fairness objection

This is the one people handle worst, usually by apologising.

Do not. The client is applying a cost-plus model to work they agreed to buy on value. You quoted a price for an outcome, they accepted it, and you delivered the outcome. How many hours it took you is not a term of the agreement unless you billed hourly, in which case bill the hours you actually worked and the question answers itself.

What does help is making the invisible work visible. The model produced a draft. You chose the approach, supplied the context, rejected the versions that were wrong, checked the facts, and took responsibility for the result. Say that, concretely, with reference to specific decisions in their project. "The first three structures it generated put the pricing section above the use cases, which would have lost the enterprise readers, so I restructured it" is persuasive in a way that "there was a lot of human input" is not.

If you are billing hourly and AI genuinely halved your time, you have a rate problem rather than a disclosure problem, and it is worth fixing properly rather than per argument. Pricing freelance work when AI makes you faster works through the structures that survive this.

Answering the quality objection

Easiest of the three, because it is testable. Invite the test.

Offer to walk them through the work. Point at the parts that required judgment about their business. Show where you corrected something the model got wrong. If there is a factual claim in the deliverable, show the source you verified it against. Quality objections tend to evaporate the moment a client sees that someone actually read the thing.

They do not evaporate if the work is genuinely thin. Be honest with yourself before you open this conversation. If what you handed over was a lightly edited first draft, the client is right, and the rest of this article is about the wrong problem.

When to offer money back

One case, and it is narrower than guilt will suggest: when you broke a term they are entitled to rely on.

A confidentiality clause you breached. A no-AI clause you ignored. An explicit assurance you gave and then did not honour. In those cases the deliverable's quality is beside the point, because they did not get the thing they contracted for.

A client who is merely annoyed, who received good work, and who has no term you breached is not owed a refund. Offering one teaches them that expressing displeasure reduces invoices, and it implies you did something wrong when you did not. Hold the position politely. If the relationship ends over it, it was going to end over the next thing instead.

Where a refund genuinely is owed, structure it rather than improvising a number under pressure. Refunding a client for a failed AI build covers the calculation and the written settlement that stops the matter reopening.

The version of this that never happens

Say it first. Put one line in your proposal describing how you work, in neutral terms, before anyone has to discover anything: which categories of tool you use, what never goes into them, and what you remain accountable for.

Clients who object to that line object at proposal stage, which costs you a proposal instead of a relationship. Most do not object at all. The discovery is what does the damage, not the practice.

FAQ

Do I have to tell clients I use AI?

Legally, usually only if your contract requires it or your jurisdiction regulates the specific output. Commercially, telling them beforehand costs far less than being found out, and it is cheap to do in a proposal.

What if the contract says nothing about AI at all?

Then you have not breached a term, and the conversation is about expectations rather than obligations. Confidentiality clauses still apply though, and they are often written broadly enough to cover sending material to a third-party service.

Should I stop using AI for that client?

Only if they ask you to, and if they do, price the work accordingly and say so at the time. Silently absorbing a slower process to keep a client happy is a decision you will resent in about a month.

How do I answer if they ask what percentage was AI?

The question has no honest numeric answer, so do not invent one. Describe the division of labour instead: what it drafted, what you decided, what you corrected, and what you are accountable for.