An HR team's competency matrix for AI agents sets out what each person must be able to do once agents answer policy questions, draft letters and sort employee requests. The core skills are judging a draft against policy, protecting health and other sensitive data, and keeping every decision about a person's job with a person.
What is this matrix for?
HR agents tend to start at low depth with high sensitivity. An agent that reads the policy library at depth 1 and drafts replies or letters at depth 2 touches pay, absence, health and conduct. None of its work moves money directly, yet a careless draft can still cause harm, breach a duty of confidence or prejudice a later dispute.
The matrix makes sure the people around the agent can catch that. It also records a boundary: agents in HR assemble and draft; they do not decide outcomes for employees. Where data protection law applies, decisions based solely on automated processing with significant effects on a person are restricted, which in the GDPR is Article 22. HR teams should treat that as a line, not a setting.
Use the matrix before the agent shadows any HR inbox, and refresh it whenever the agent gains a new kind of letter or a new source. It fits under the broader guide to AI agent skills by role.
How should HR fill it in?
Take each task the agent supports, then ask two questions: what could go wrong in the draft, and who would notice? The skill is whatever the noticing takes. Set levels as:
- A: understands the vocabulary and can explain to an employee why a request is waiting for a person.
- B: reviews and sends agent drafts in a live queue, with reasons when editing or refusing.
- C: approves held items with legal or pay effect and coaches others.
- D: owns the agent, sets its sources and bounds, and answers for it.
The template
| HR task the agent supports | Skill the person needs | HR adviser | HR business partner | HR operations lead | Agent owner | Evidence |
|---|---|---|---|---|---|---|
| Answers policy questions | Checks the answer is grounded in the current policy version | |||||
| Drafts absence and leave letters | Spots health information that should not be repeated | |||||
| Drafts contract change letters | Confirms the change was approved by a person | |||||
| Sorts grievance and conduct emails | Routes to a person without the agent replying on substance | |||||
| Prepares reference requests | Checks the source of every factual statement | |||||
| Responds to access requests | Recognises a data subject request and its deadline | |||||
| All tasks | Reads why an item was held or refused | |||||
| All tasks | Reconstructs a sent letter from the record |
A worked example
Take a hypothetical employer of 2,000 staff with an HR team of six advisers, two business partners, an operations lead and an agent owner from HR systems.
| HR task | Skill | Adviser | Partner | Ops lead | Owner | Evidence |
|---|---|---|---|---|---|---|
| Policy questions | Grounded in current version | B | B | C | D | Sample of 30 answers re-marked |
| Absence letters | Health data handling | B | C | C | B | Scenario test |
| Contract changes | Change approved by a person | B | C | C | A | Approval trail check |
| Grievance emails | Routes without replying | B | B | C | B | Live queue review |
| References | Source of each fact | B | C | C | A | Sample check |
| Access requests | Recognises the request | B | B | C | B | Drill |
| All | Reads why held or refused | B | B | C | D | Foundations test |
| All | Reconstructs a letter | A | B | C | D | Audit walkthrough |
The agent owner sits at A for contract changes and references on purpose. They answer for the agent but should not be the person approving its letters about named employees. The approver rule for the HR team keeps those approvals with the operations lead and business partners.
The specific HR risk is the absence letter. An employee may email that they are off with a named condition. The agent, drafting a return-to-work letter, may repeat it. Advisers at level B must catch that, because health information is special-category data and its spread inside the organisation should be limited to people who need it. That check should be in the reviewer rota for the HR team as a named duty, not left to chance.
Public bodies carry extra duties of fairness and transparency, so a public-sector HR team may set higher targets. The brief on AI agent training for HR advisers in the public sector in Canada shows how one market shapes that.
Where the Academy fits
Pryme Intelligence Academy suits this matrix level by level. Foundations gives the whole HR team level A in about four hours, free. Advisers and partners reviewing live drafts take the Operator track, whose module tests are set in their own industry and role. The agent owner adds the Builder track to reach level D.
Questions people ask
Can an HR agent ever decide anything about an employee?
It can make routine, reversible changes within bounds the business sets, such as booking an approved meeting. Decisions with real effect on someone's employment, pay or record should stay with a person who can explain them.
Who should own the HR agent?
Usually the HR operations lead or the owner of the HR process it works in. HR systems staff can build it, but ownership should sit with someone who answers for HR outcomes.
How is this different from the lending team's matrix?
The lending team's competency matrix is about credit decisions with money at stake. HR's risks are confidentiality and fairness to employees, so its matrix weights data handling and routing over financial approval.