PII detection and data minimisation

Control objective

PII detection and data minimisation checks that the system reliably finds personal data in model outputs and that it only collects, processes and exposes the minimum personal data needed for the stated purpose — the heart of privacy-by-design and purpose limitation under IBM Sharing PI, NIST Data Privacy and OWASP ASI03. For a bank, an assistant that echoes a customer's PAN, Aadhaar or account number, or that hoards data beyond what a task requires, creates direct privacy and regulatory exposure. The control is measured by PII-detection recall on outputs, with a threshold of recall ≥ 0.95, alongside qualitative requirements that data minimisation is applied and purpose limitation is enforced. Recall here is the fraction of true PII instances the detector actually catches, so the target means missing no more than 5% of real PII. To implement, run a PII detector (NER plus pattern/checksum rules for identifiers like PAN and Aadhaar) over generated outputs, validate its recall against a labelled test set, and pair it with redaction; enforce minimisation by stripping fields not required for the purpose and binding data use to a declared purpose. Log detection results, recall measurements and minimisation decisions. If output recall falls below 0.95, or minimisation/purpose limits are not enforced, the control breaches, triggering detector retuning and remediation before release.