Governance · 7 min read · Aug 12, 2026

Who sees it, and who keeps it.

This is the question that stalls more AI projects than cost, and it is usually asked too late and answered too vaguely. Here is what actually happens to a request, which parts are contractual rather than technical, and the specific questions that produce an answer you can act on.

Follow one request

A single call travels further than most people picture. Your application sends a prompt — which may include retrieved documents, customer records or a system instruction — to a provider endpoint. It is processed on the provider's infrastructure, which is frequently a hyperscaler's, in a region that may or may not be the one you assumed. A response comes back. Somewhere in that path, the request may be logged.

Every meaningful question about data protection sits in that last sentence. Not whether the request happened, but what persisted afterwards and who could read it.

The four things that actually matter

Strip away the marketing and there are four properties. They are independent of each other, and a reassuring answer on one tells you nothing about the other three.

PropertyThe real questionWhere the answer lives
TrainingIs my data used to improve their models?Contract / tier
RetentionHow long is it stored, and can that be zero?Contract / config
ResidencyWhich country is it processed and stored in?Region setting
AccessWhich humans can read it, and under what process?Policy / audit
Four independent properties of LLM data exposure: whether content is used for training, how long it is retained, which region it is processed and stored in, and which humans can read it. Each is settled by contract or configuration rather than by the model.
Four independent properties. A good answer on one says nothing about the others.

Note the third column. Almost none of this is a property of the model — it is a property of the agreement and the configuration. The same model, from the same provider, can have entirely different answers depending on which tier you bought and which region you deployed in.

Consumer terms and business terms are not the same product

This is the single most common source of confusion inside organisations, and it usually surfaces as a policy argument based on a headline about the consumer product.

Consumer chat products and paid API or enterprise tiers generally operate under different terms, frequently with different defaults on training and retention. An employee pasting a client contract into a personal account and an application calling a business API under a signed agreement are two genuinely different risk events. Treating them as the same thing produces a policy that blocks the safe path and does nothing about the unsafe one.

The realistic leak in most organisations is not the API you procured. It is the browser tab someone opened because the procured tool was too slow to arrive.

Zero retention is a real setting, and it has consequences

Several providers offer configurations where request and response content is not persisted. Where available this is the strongest technical answer to the retention question, and for regulated work it is often what makes a deployment approvable at all.

It is worth knowing what you give up. Without retained logs, some abuse-monitoring features and provider-side debugging become unavailable, and you cannot ask the provider to help reconstruct what happened during an incident. That is usually the right trade, but make it deliberately — teams that enable it without noticing find out during their first production incident.

Residency is not solved by picking a region

Selecting a region controls where inference runs. It does not automatically tell you where logs go, where abuse-monitoring data goes, where the support team sits, or which sub-processors are involved. Those are separate answers and they are frequently different.

If a regulator or a client contract constrains you to a jurisdiction, region selection is the beginning of the answer, not the whole of it. Ask about each path independently.

The questions that get a straight answer

Vague questions get vague answers. These do not, and every one of them should be answerable in writing before anything sensitive is sent:

  • Is our content used to train or improve your models under our current tier? Point me at the clause.
  • What is the retention period for prompts and responses, and can it be set to zero on our account?
  • In which region is inference performed, and in which regions is any logged content stored?
  • Who are your sub-processors, and how are we notified when that list changes?
  • Under what circumstances does a human at your company read our content, and is it logged?
  • What happens to our data if we terminate — what is deleted, and on what timeline?
Six questions to put to a model provider: training use, retention period and whether zero is available, processing and logging regions, sub-processors and change notification, when a human reads content, and what happens to data on termination.
Every one of these should be answerable in writing before anything sensitive is sent.

A provider selling to businesses will answer all six in writing. If any answer is a reassurance rather than a clause, treat that as the finding — it is the same signal as a vendor who will not quote running cost before signature.

What you control on your side

Provider terms are one half. The other half is what you chose to send, and it is the half you can change without a negotiation.

Most systems send far more than the task requires, because it was easier to pass the whole record than to select fields. Redacting or tokenising identifiers before they leave your boundary, sending the minimum context that answers the question, and keeping a log on your side of what was sent — not just what came back — cost little and change your exposure materially. The strongest position is not a provider promise about data they hold. It is not having sent it.

Dealing with this in your own group?

We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.

Questions

Short answers,
in full.

The questions this article gets asked most, answered so each one stands on its own.

Talk to us
Do AI providers train on business data sent through their API?

It depends on the tier and the contract, not on the model. Paid API and enterprise tiers commonly differ from consumer chat products, frequently with different defaults on training and retention. Do not reason from a headline about the consumer product — ask for the clause that applies to your account and get it in writing.

Can we stop an AI provider from storing our prompts?

Often yes. Several providers offer zero-retention configurations where request and response content is not persisted. The trade-off is that some abuse monitoring and provider-side debugging become unavailable, and the provider cannot help reconstruct an incident after the fact. It is usually the right choice for regulated work, but it should be a deliberate one.

Does choosing a region guarantee data residency?

No. Region selection controls where inference runs. Where logs are stored, where abuse-monitoring data goes, which sub-processors are involved and where support staff sit are separate questions with frequently different answers. If a contract or regulator constrains your jurisdiction, ask about each path independently.

What is the biggest data risk in adopting AI?

In most organisations it is not the procured API — it is unmanaged consumer tool use by employees, which operates under different terms and outside any logging you control. That risk grows when an approved internal option is slow to arrive, so shipping a sanctioned path quickly is a data-protection measure, not just a productivity one.