Pick a narrow, reversible task

Start by asking the system to suggest a label for a message, such as order status, return request or product question. Keep normal inbox ownership and leave sending, refunds and account changes with staff. Agree on a label called Needs review for mixed or unfamiliar requests. Write a one-line definition and a representative example for every label.

Build a test set before touching the inbox

Collect examples that staff can label consistently. Remove identifying details or write fictional messages that preserve the wording patterns. Include messages with incomplete order details, more than one question and wording that is hard to classify. Keep a small second group for the final test so the same examples aren’t used for adjustments and evaluation.

Make mistakes visible

Track the label suggested, the label chosen by the reviewer and the time to review. One invented test message says a parcel arrived damaged and asks for an address correction. A reviewer might choose Needs review because it combines two requests. Count mistakes separately for messages that could delay a customer’s next step. An average accuracy figure can mask those cases.

Test a daily operating procedure

Have staff run the triage step alongside their usual work for a few shifts. Each morning, sample a handful of ordinary suggestions and review all low-confidence or escalated cases. If review time rises or important messages land in the wrong queue, pause the pilot and adjust the labels or workflow. Use only approved tools and data handling procedures.

Write down the handover

Document who checks the queue, how they correct suggestions and who can pause the pilot. A useful worksheet has columns for example ID, expected label, suggested label, review time, outcome and correction reason. Review the log weekly while the label set is still changing.

Related guides

Further reading. NIST AI Risk Management Framework.