“Is My Customer Data Training Someone’s Model?” The Question to Ask Before You Sign

Why this lands on you

The information is not abstract. A dental practice holds what people have said about their teeth. An attorney holds the outline of somebody’s divorce. A guesthouse holds names, dates and who is travelling with whom. A plumber holds home addresses and the times nobody will be in.

When that goes into a tool, your customer’s privacy now depends on a company they have never heard of. South African businesses also have obligations under POPIA for the personal information they collect, and the short version is that handing information to a supplier does not hand over the responsibility for it.

Inference, retention and training are three different things

Worth separating, because vendors blur them. Inference is the tool doing its job: information goes to a model, an answer comes back, the model is unchanged. Retention is a provider holding those inputs and outputs for a limited period, usually for security and abuse prevention. Training is using what you put in to change the model itself, permanently, in a way that can surface for somebody else.

Inference is the product working. Short, disclosed retention is ordinary hygiene. Training on customer information without clear agreement is where confidentiality stops existing. A vendor worth signing with separates the three in writing and commits on the third one in words that cannot wriggle.

What a usable commitment looks like

A good one does four things. It covers both what you put in and what comes out. It binds the AI providers behind the product as well as the company selling it to you. It deals with retention separately instead of hiding behind it. And it says the same thing in the privacy policy, the terms and any data processing agreement, because a commitment that changes wording between documents is a commitment with an exit.

Amysa’s published position is that customer content, enquiry data and AI output are not used to train models, that the providers it contracts with are not permitted to train on data submitted through the service, and that those providers may hold inputs and outputs for a limited period for security and abuse prevention. It names Anthropic’s Claude models as the models doing the work. The wording is at PLACEHOLDER-DOMAIN.tld/legal/amysa, which is the point: you should be reading it not take a salesperson’s word for it.

Four more questions for the same conversation

Who is responsible for what? There should be an agreement that says you are responsible for the customer information and the vendor processes it on your instructions, with the other companies in the chain listed.

Which AI models, and whose account? A vendor should be able to name the provider. If you bring your own key to a tool, you are operating under your own agreement with that provider, which is a different arrangement with different terms.

Who inside the vendor can see your data, and under what circumstances? Honest tools say so plainly instead of leaving you to discover it.

What happens when you leave? Look for a stated export window and a deletion timetable, and for honesty about the edges, such as records that have already been written into systems you control and are therefore yours to keep.

The audit takes one afternoon

Most small businesses have fewer tools than they think. List them: the booking system, the email marketing account, the chat widget, the accounting package, whatever the website runs on, and anything a staff member signed up for on their own.

Add three columns. Does it hold customer information. Does it have AI features. What does its published policy say about training. The third column is the afternoon’s work, and it means finding the policy page not the marketing page. If you cannot find a sentence about training, write “no published position”, because that is the answer.

Grade what you find, then act on it

Green is an explicit commitment that covers your content and the output, binds the model providers, and deals with retention separately. Amber is a commitment with room to move, and the classic phrase is “we may use data to improve our services”, which can mean anything at all. Red is silence, or training switched on by default with an opt-out hidden in a settings page.

Ambers get an email asking for clarification in writing, which is a reasonable request and tells you a lot by how it is answered. Reds get either a replacement or a deliberate decision to accept the risk, made by you not by a default.

Then tell your own customers

The audit protects the business. The part that earns you something is putting the answer where customers can see it, in the privacy notice on your own website, in language a person can read: what you collect, what it is used for, who else touches it, and that it is not used to train AI models.

Very few businesses in your street can say that sentence and point at evidence. It costs you nothing, and the first time a customer asks, you answer with a link instead of a scramble.

The takeaway

Send the headline question to every tool in your stack that touches customer information this month. Is our data, or our customers’ data, used to train AI models, yours or anyone else’s, and where is that published? Vendors with good answers reply with links. Vendors with bad answers reply with adjectives.

Amysa publishes its answers at PLACEHOLDER-DOMAIN.tld/legal/amysa and will put the data processing terms in front of you before you sign not after. That order tells you something on its own.