How to train the HR team to review AI agent decisions

HR reviewers should sort the agent's work by how hard it is to undo. Routine answers and changes can be sampled; letters that change someone's employment are read in full. Teach that first, calibrate on real drafted letters and policy answers, shadow live review with every letter checked, and sign off advisers who never let an irreversible step pass unread.

What does the HR team review?

An HR agent typically answers employees' policy questions, processes routine changes such as addresses and working pattern requests within policy at depth 3, drafts letters for contract changes, absence meetings and disciplinary invitations, flags absence patterns against policy triggers, and summarises job applications against criteria. Anything that changes or ends someone's employment is drafted for a person at depth 2 and decided by them.

The HR team's review work divides by consequence. Policy answers: high volume, mostly low consequence, sampled. Routine changes: low consequence, sampled. Drafted letters: high consequence, every one read. Absence triggers and candidate summaries: they shape decisions about people, so they are reviewed before anyone acts on them.

Some HR decisions are effectively irreversible. A dismissal, a withdrawn offer, a disciplinary warning on file: even if later reversed, the effect on the person and the relationship remains. The case is made in irreversibility in HR decisions. Training should make that weight felt.

The head of HR owns the programme. Employee relations sets the sign-off standard for letters. The data protection lead reviews the parts covering personal data.

What are the steps?

1. Learning, week 1: advisers complete lessons on the agent's decision chain, and on which of its outputs are sampled and which are read in full.
2. Agent, depth 1: builds a practice set of past policy answers, letters, absence flags and candidate summaries, with corrections made.
3. Calibration, week 2: advisers review the set alone, then compare. Letters and absence flags come first.
4. Shadow, weeks 3 and 4: advisers review live work; a senior adviser checks every letter before issue and a sample of everything else.
5. Agent, depth 2 and 3 in live work: answers questions, processes routine changes, drafts letters, flags triggers.
6. Sign-off: head of HR and employee relations assess shadow reviews.
7. Ongoing: weekly sampling of policy answers, monthly calibration on letters.

What must advisers learn to check?

Policy answers. Is the answer right for this employee's contract, location and group? Policies often vary by site or grade, and the agent answers from what it is given. Sample answers for those differences.

Letters. Correct facts, correct process stage, correct rights stated (such as the right to be accompanied, where it applies), and a tone that is clear without being cold. A disciplinary invitation that gets the allegation wrong cannot be quietly corrected later.

Absence flags. A trigger has been reached; that does not mean action should follow. Advisers check for disability, pregnancy, caring duties or other circumstances that change what is appropriate.

Candidate summaries. Does the summary reflect the criteria fairly? Did the agent flag anything unusual about the application? Is the same standard applied to every candidate?

Routine changes. Sample to make sure "routine" stayed routine: a working pattern change that affects pay is not.

Anything about bank details. Held, and verified with the employee directly.

An illustrative calibration case shows why letters get full attention. The agent drafts an invitation to a capability meeting for an employee whose absence reached the trigger. Half the advisers approve it. The other half notice, in the employee's file, a recent note from occupational health recommending adjustments that have not yet been made. The agreed standard: where an occupational health recommendation is outstanding, the agent holds the draft and the adviser speaks to the manager first.

Which controls leave evidence?

• Every letter read before issue. Evidence: the adviser's name on each.
• Absence actions reasoned. Evidence: a note on circumstances considered.
• Policy answers sampled. Evidence: weekly sample log.
• Candidate summaries reviewed before decisions. Evidence: reviewer named on each shortlist.
• Sign-off recorded. Evidence: assessment by HR and employee relations.

What should you measure?

• Letter errors found in shadow and after issue.
• Policy answer accuracy in sampling, by site and grade.
• Absence actions later withdrawn.
• Grievances or complaints about HR process.
• Consistency of candidate summary review across recruiters.

Where does it usually go wrong?

All outputs treated alike. Advisers either read everything, and burn out, or sample everything, and miss a bad letter. Teach the distinction.

Policy variations missed. The agent answers from the main policy; a site agreement says otherwise. Make site and grade variations part of calibration.

Triggers drive action. Absence flags lead straight to meetings without considering circumstances.

Managers bypass HR. Managers ask the agent for letter drafts directly. Agree who may request what.

Payroll and support teams follow the same four stages; see how to train the payroll team to review AI agent decisions and how to train the support team to review AI agent decisions. The individual view is in how HR advisers review an AI agent's decisions, and new starters use the reviewer onboarding checklist for the HR team. The full programme is in training teams for AI agents. Pryme Intelligence Academy's Operator track covers this kind of review, with module tests set in the learner's own industry and role.