Nanny Agency Jobs How-to / checklist

Practice approved AI tasks with fictional material, source checks, escalation and records that show who made each decision.

Start with the agency’s allowed tasks

Staff training should start from a written list of approved tools, accounts, tasks and data limits. A generic AI seminar does not tell a recruiter whether they may upload a candidate résumé, draft a status update or connect a database. Name an owner who can answer those questions.

This is a practice plan an agency can adapt to its approved workflow. NIST’s AI RMF Core calls for defined human oversight and practitioner roles.

Practice with original fictional cases

Use invented postings, candidate records and family requests. Ask trainees to draft a neutral process message, verify every factual claim against the fictional source and mark anything unknown. Include a bad output that invents an interview or changes a schedule.

Do not use real candidate, reference or household data as a convenient exercise. If a vendor’s approved production workflow later handles such data, train against that specific permission and documented safeguards, not against a consumer account assumed to behave the same way.

Teach escalation and correction

Give staff a short rule: stop when the tool invents a fact, exposes information, gives legal or pay conclusions, or suggests an employment decision beyond the approved task. Record the output and send it to the responsible reviewer through the agency’s normal incident route.

Fictional example: a model changes ‘family review pending’ into ‘family interested.’ The trainee removes the claim, checks the source record and flags the pattern. The corrected message remains a draft until a person approves and sends it.

Check that the practice transfers

Ask each trainee to complete one observed fictional exercise: identify the approved account, reduce the input, cite the source, explain the output limit and show the handoff. Record completion and unresolved questions, then revisit when a tool or workflow changes.

The measure is fewer unsupported claims and clearer ownership in actual work, not the number of prompts staff wrote. Keep human review, access checks and ordinary agency records in place.

Sources