A benefits office avoids putting full bank details into Copilot when drafting payment-delay notices. What practice does this demonstrate?

Minimizing sensitive financial data in generative AI workflows is the right municipal practice because payment-delay notices rarely need full bank account or routing numbers in the prompt itself. Benefits staff can instruct Copilot to draft polite, plain-language notices while keeping complete financial identifiers out of the chat history, grounding context, and any downstream saved prompt. The correct approach works by shrinking the sensitive-data surface: if the account number never enters Copilot, it cannot be echoed, summarized into the wrong thread, or retained in conversation residue after the notice is finished.

Pasting complete account numbers into every Copilot prompt fails because accuracy for a rewrite or tone task does not require full banking identifiers; it only increases exposure if the chat is shared, misfiled, or later reused. Disabling municipal data protection so Copilot can invent bank routing numbers fails conceptually on two fronts: inventing routing data is fabrication, and turning off protection removes the controls that keep resident financial information from traveling beyond authorized pathways. Publishing resident banking details in Agent Store templates fails because Agent Store assets are reusable across users; embedding banking details turns a convenience template into a sensitive-data distribution channel.

Exam caveat: minimization is not the same as refusing all operational context—case IDs, payment status categories, and approved notice language can still be provided when they are needed and appropriately classified. Operational check: before prompting, strip or replace full account and routing numbers with case or voucher IDs, confirm the draft never reprints banking digits, and save only the de-identified prompt pattern for reuse by benefits staff.