How should you answer a customer who asks if AI handled their data?

Read the question before answering it, because it is usually a proxy for a different one. Almost nobody objects to a supplier drafting an email with a tool. What people are actually asking is whether their material left the business, whether it went somewhere that keeps it, and whether the supplier knows the answer. The last part is the one that decides how the relationship goes: a specific answer that includes something the customer would rather not hear lands better than a comfortable answer that is visibly not based on anything.

So the reply that works has four parts and is short. Which tools received information relating to them. What those tools received — the whole file, a summary, a name and a date, nothing. What the vendor's terms say happens to it, particularly retention and whether inputs are used for improvement. And whether a person reviewed the output before it reached them. Four sentences, and every one of them is checkable.

The obligation to answer at all varies and is often stronger than businesses assume. Where the customer is a consumer in a jurisdiction with data rights, a request about processing may be a formal one with a deadline attached. Where the customer is a business, the contract frequently contains an audit or information right, and increasingly a clause about subprocessors that a new tool has quietly breached. Where the customer is in a regulated relationship — a patient, a client, a tenant — the professional rules may set the standard rather than the contract. Checking which of these applies takes minutes and changes what the answer has to contain.

The most common failure is not evasion but genuine ignorance, and it is worth being honest about internally. Most businesses cannot answer this question on the day it is asked because nobody keeps a list of which tools are connected to what. That is a records gap rather than a policy gap, and it has a cheap fix: a single sheet naming each tool, what it can reach, which plan it is on, and who owns it. Businesses that build it after the first customer asks describe the exercise as uncomfortable and short.

There is a version of the answer that makes things worse and it is worth naming. Assuring a customer that their data is never used for AI training, when the business has not read the terms of the tier its staff are actually on, is a statement about the world that may be false — and it converts a question about process into a potential misrepresentation. If the honest answer is that you are checking, that is a better answer than a confident one you cannot support.

Finally, the answer should say what happens next, because a question like this is usually a signal rather than an audit. Offering the customer a choice — to have their material excluded from these tools, or handled in a specified way — turns the exchange into a decision they made rather than a disclosure they extracted. Most decline the offer. The ones who accept were going to be a problem later, and finding out now is worth more than the accommodation costs.