Keeping Customer Data Safe Inside an AI App

These are design decisions, not paperwork

Privacy questions asked at the end of an AI project turn into rework. Asked at the start, they cost a meeting.

The reason is architectural: whether personal data leaves your infrastructure determines which providers are viable, which affects cost, capability and latency.

1. Send less than you can

Most AI features need far less personal data than the first design assumes. A support assistant needs the account's plan and order status; it rarely needs the customer's full record.

• Identifiers rather than identities where possible
• Redact names, addresses and payment details that add nothing
• A summarised context rather than a whole document
• Special-category data excluded unless there is a documented reason

2. Where processing happens

Major providers offer regional processing. For UK and EU businesses that is usually the sensible default — it simplifies transfers and is far easier to explain to a customer or an auditor.

Get it in the contract rather than the marketing page, and confirm it applies to every component: the model, the vector store, the logs and any monitoring service. Logging is the one people forget, and logs contain prompts.

3. Retention and deletion

Decide what is kept, where and for how long, for each of prompts, outputs, retrieved context and conversation history. Then implement deletion rather than intending it.

The design test: if a customer exercises their right to erasure tomorrow, can you remove their data from every store the feature touches, including logs and embeddings?

4. Training on your inputs

Enterprise API terms from major providers generally exclude your inputs from training; consumer tiers often do not. The difference matters and the terms change, so verify against current terms at contract time.

Whatever the answer, write it into your privacy notice in plain words. Customers increasingly ask.

What we produce during scoping

1. A one-page data flow: what leaves your systems, to whom, processed where, what comes back, what is stored
2. Retention periods per category, implemented rather than documented
3. Access control by role, with an audit trail on sensitive records
4. A note on whether a DPIA is warranted

Frequently asked questions