AI agent training plan template for HR advisers

HR advisers supervising an AI agent need a plan that teaches them to judge its policy answers, drafted letters and case summaries, and that fixes in place the decisions only a person may take: anything affecting someone's employment, pay or health. The plan runs from shared vocabulary to a certified practicum over about six weeks, then renews yearly.

What is this plan for?

HR agents answer employees' policy questions from the handbook and policy library, draft standard letters (contract variations, reference confirmations, absence reviews) at depth 2, summarise case notes, and schedule meetings at depth 3. They should refuse anything outside policy and hold anything sensitive. Health information about employees is special-category data under GDPR, and much of what HR handles is confidential.

The HR adviser becomes the reviewer of that work. The risks are specific: a letter that is correct by policy but wrong for this employee, a summary that loses a detail mattering in a grievance, or a policy answer drawn from an outdated version of the handbook. Time to competence is addressed in how long it takes to train HR advisers to supervise AI agents.

How do you fill it in?

The template is a single form per adviser, in three blocks: learning, practice and authority. Complete the red lines with your HR director; they define what the practicum tests and what the adviser must never approve.

The template

Block 1: learning

1. Adviser name and client groups supported
2. Agent capabilities in scope, with depth
3. Foundation learning: date completed
4. Reviewer learning: modules and dates
5. Written exam: date and result

Block 2: practice

6. Shadow period: dates, certified colleague, items compared
7. Shadow result: matches out of items compared
8. Practicum: dates, assessor, result
9. Calibration attended

Block 3: authority

10. Red lines (list)
11. Internal sign-off: HR director, date
12. Approval rights granted
13. Certificate number and expiry
14. Renewal booked

A worked example

A hypothetical retail group's HR advisory team supporting 4,000 store staff. Values are samples.

1. HR Adviser A; stores in two regions.
2. Policy answers from the policy library, depth 2; standard letters, depth 2; case note summaries, depth 1; meeting scheduling, depth 3.
3. Foundations, week 1.
4. Operator modules: how a governed agent decides, human review, keeping knowledge current, when something goes wrong; weeks 2 to 4.
5. Exam in week 4: 35 of 40.
6. Weeks 3 and 4 with HR Adviser B: A decides each held letter and answer before B releases it.
7. 46 of 50 matched; the four differences were all in absence letters.
8. Weeks 5 and 6, Pryme assessor; at least 95 in every 100 decisions correct; no missed red line.
9. Calibration on absence review letters, week 6.
10. Red lines:
    • Never send a letter that changes an employee's pay, hours or contract without the line manager's and HR's documented decision.
    • Never let a policy answer stand if it cites a policy version other than the current one in the policy library.
    • Never include health details in a letter or summary that goes beyond those who need them.
    • Any message mentioning harassment, discrimination or whistleblowing goes to a person-led route.
11. Signed off by the HR director, week 6.
12. May approve standard letters and policy answers; may not approve dismissals or contract changes.
13. Certified Pryme Operator, valid 12 months.
14. Renewal booked for month 11: update module and short quiz.

The shadow result at item 7 is worth reading. All four differences were absence review letters, where the agent applied the trigger points correctly but missed a note that the absences related to a disability. Adviser B caught it; Adviser A had not yet learned to read the case notes behind the letter. That became the calibration topic in item 9, and the practicum included similar cases.

The policy-version red line matters because HR policy changes often and old versions linger in shared drives. If the agent's sources include a superseded handbook, its answers will be confidently wrong. The adviser checks the citation, and repeated issues go to whoever maintains the agent's sources. Employees in disputes sometimes write messages designed to steer an automated reply; practise with how to train HR advisers to spot prompt injection and pressure. Keep the team consistent with calibration sessions for HR advisers, and see the review method in how HR advisers review an AI agent's decisions. A neighbouring advisory role is covered in the training plan for wealth advisers, and the team-wide order in training teams for AI agents.