How to teach HR advisers to bank corrections that improve an AI agent

HR advisers bank a correction by writing down, in general terms, why the agent's answer or letter was wrong, and passing that reason to the agent owner as a change to policy sources, brief or bounds. The HR-specific discipline is stripping out the employee's details first, so a fix never carries personal data into the agent.

What gets corrected in an HR team?

An HR agent typically answers employee questions on policy, drafts letters for absence, flexible working or probation reviews, and prepares case notes for advisers. It works at depth 2: it drafts, and an adviser sends. The corrections advisers make fall into a few recognisable kinds.

The agent quotes the superseded family leave policy because the new version was published as an attachment the agent cannot read. It applies the head office rule on overtime to a site with its own local agreement. It drafts an absence letter that is accurate but cold, in a case where the employee has disclosed a serious illness. Or it answers a question that should have gone to an adviser, such as whether a grievance can be raised anonymously against a manager.

Each of these is fixable, but in different places. The adviser's job is to tell which. The general picture of who learns what is in AI agent skills by role.

How does the banking process run?

1. Agent, depth 2: drafts the answer or letter, citing the policy section and version it used.
2. Person (HR adviser): corrects and sends, then records a reason: outdated policy, wrong local rule, tone, should have escalated, or one-off.
3. Person (adviser): writes the reason without names. "Applied head office overtime rule to a site covered by a local agreement", not the employee's name, site and pay.
4. Agent, depth 1: groups reasons weekly so the HR lead sees which policies and sites generate repeat corrections.
5. Person (HR lead): decides what to bank. A single tone correction may be a one-off; four in a month on illness-related absence is a pattern.
6. Person (agent owner): makes the change. A new policy version replaces the old one as a sealed source from the policy library; a rule about local agreements goes into the brief; questions about grievances against managers become a hard escalation.
7. Attest replays the agent against past employee questions and letters, checking that the escalation and refusal gates still clear.

What do HR advisers need to learn?

The first skill is writing a correction reason that is useful and contains no personal data. Advisers are used to detailed case notes; a correction reason is the opposite, a general statement of what the agent did wrong. This is a habit, and it takes practice. A reason that names the employee or describes their health is a data protection problem waiting in the record.

The second is seeing that most HR corrections are source problems. Policies change often, local agreements multiply, and the agent can only be as current as the library it reads. An adviser who notices that the agent quoted a superseded policy should check whether the new version is in the library at all before blaming the brief.

The third is knowing what should never be banked as a rule. Tone in a bereavement case, a judgement to allow extra leave outside policy, a decision to treat a complaint informally: these belong to the adviser. Banking them would turn discretion into a default. The everyday checks that sit underneath this are covered in how HR advisers review an AI agent's decisions.

Which controls and evidence should be in place?

• No personal data in reasons. The HR lead spot-checks reasons each month. Evidence: the reviewed sample and any redactions.
• Policy library as the only source. The agent reads policies only from the sealed library, never from emailed copies. Evidence: the provenance class on every policy citation.
• Escalation rules for sensitive topics. Grievances, disciplinary matters and health disclosures route to an adviser. Evidence: held and escalated verdicts in the record.
• Re-certification before go-live. Evidence: a dated certificate showing each gate.

What should the team measure?

Measure repeat corrections by policy: if the family leave policy keeps generating corrections after a banked change, the source fix did not take. Measure how long a new policy takes to reach the agent after publication; this is often the largest delay and is usually an ownership gap between HR policy and the agent owner. Measure the share of correction reasons that needed redaction, which should fall as advisers learn the habit. And watch escalations: if they drop sharply after a change, check that sensitive questions are not now being answered by the agent.

Where does it usually go wrong?

Policy updates that bypass the agent are the most common failure. HR publishes a new policy on the intranet, but nobody tells the agent owner, so the agent keeps quoting the old one until an adviser notices. Tie policy publication to a source update in the same step.

The second failure is treating local variation as error. A large employer can have a dozen site agreements; advisers correct the agent site by site and never bank the rule that it must check which agreement applies. Wealth advisers face the same pattern with client-specific mandates, as teaching wealth advisers to bank corrections shows, and IT teams face a version of it with group mappings, covered in teaching IT administrators to bank corrections.

The third is over-banking tone. Advisers soften the agent's letters, and someone asks for the brief to make all letters warmer. The result is warmth in the wrong places, including formal warnings. Tone rules should be specific to letter type.