advanced 2 min answer

Legal asks whether customer personal data is being sent to your model provider. What must you be able to answer?

privacyaicompliance
Show the full answer Hide the answer

What the interviewer is testing

Whether you have thought about an AI feature as a data processing activity with a data flow.

What must be answerable

What data is sent. The prompt, the retrieved context, conversation history and any tool results all leave your estate. Teams frequently account for the user's question and forget that the retrieved context — which may be an entire customer record — goes with it.

Where it is processed. Which region, and whether that constitutes a cross-border transfer needing a mechanism.

Whether it is retained, for how long, and whether it is used for training. Enterprise agreements typically exclude training use and set retention to zero or short, but this must be verified in the contract rather than assumed.

Who the sub-processors are, since the provider's own infrastructure providers are in the chain.

What the lawful basis is, and whether the processing is compatible with the purpose the data was collected for. This is the question that most often has no good answer, because the feature was built without asking.

The architectural controls

Minimise what is sent. Retrieve and pass only the fields the task requires, not whole records. This is the highest-value control and it usually reduces cost as well.

Redact or pseudonymise identifiers before the call where the task does not require them — names, account numbers, addresses replaced with placeholders and restored in the response.

Regional endpoints where the provider offers them, so the processing stays in the required jurisdiction.

Log what was sent, so a subject access request or an incident can be answered accurately.

Prevent accidental inclusion: guard against personal data reaching the prompt through logs, error messages or debug context.

What a strong answer adds

Recording the AI feature in the register of processing with its own entry — purpose, categories, recipients, transfers, retention — because it is a distinct processing activity and it will be asked about. Doing this at design time is trivial; reconstructing it later is not.

Common weak answers

Pointing at the provider's certifications. Assuming the enterprise agreement covers it without reading the retention and training terms.