HR advisers need four pieces of understanding before an agent answers employees alongside them: how depth limits what the agent may send, how to check which version of a policy it relied on, what a held reply is asking of them, and how to review a draft that touches someone's pay, health or job. The rest is practice.
Why is HR different from other agent deployments?
Most employee questions are routine. How much holiday do I have left? What is the parental leave policy? Where is the expenses form? An agent can answer these well, and HR advisers welcome the time back.
The difficulty is that a routine question can carry a non-routine fact. "How much sick leave do I have?" may come from someone about to disclose a serious illness. "Can I see my file?" may be the start of a grievance. HR advisers already read for this. What they must learn is how the agent reads for it, where it stops, and what it hands to them.
What should HR advisers understand about the agent?
- Depth is set per task. A typical HR agent reads the policy library at depth 1, drafts replies and letters at depth 2 for an adviser to send, and may answer narrow factual questions, such as a holiday balance from the HR system, at depth 3 once it has earned that on real cases. Advisers should know which tasks sit at which depth.
- Sources carry a stamp. The HR system's leave balance is sealed. A manager's note about an absence is signed. An employee's account of what happened is declared. A policy pasted into an email is external. Advisers should check which one the agent leaned on, especially after a policy change.
- The decision chain explains every hold. Five bounds apply at once: the role scope, the agent's certificate, its sources, the employer's own limits and the platform guardrails. The lowest decides, and the verdict names it. "Held: own limits, topic is health" tells the adviser why the draft is waiting.
- Certificates lapse. An agent's certificate lasts 90 days, and a run that clears no depth drops it to depth 1 the same day. After a major policy rewrite, expect re-certification and a temporary rise in held work.
- The record is permanent. Every draft, approval and edit is filed in an append-only record. Advisers should write review notes they would be comfortable seeing in a tribunal bundle.
What does the agent handle, and what stays with the adviser?
| The agent | The adviser |
|---|---|
| Finds and quotes the right policy clause | Interprets a policy where it is unclear or silent |
| Drafts replies to routine questions | Sends anything about health, conduct, pay disputes or exits |
| Prepares standard letters from templates | Signs letters that change terms or carry consequences |
| Flags messages that read as a grievance or disclosure | Handles the grievance or disclosure |
| Summarises a case file for the adviser | Advises the manager on the case |
The same split of reading, drafting and deciding appears in other advisory roles; compare what wealth advisers should learn first and what underwriters should learn first.
What does a careful review look like?
Take a hospital trust's HR team. A nurse asks the agent how sick pay works after six months' absence. The agent drafts a clear reply quoting the sick pay policy. The verdict is held, because the employer's limits route any question about long-term absence to an adviser.
The adviser checks the source and finds the agent quoted the policy version that applied before a recent revision of the long-term absence clause. The draft is accurate to the old text and wrong for this nurse. The adviser declines it, notes "superseded policy version; source library holds both", and writes the reply personally, adding an offer to talk. The note leads the team to retire the old version from the agent's sources, so the same mistake cannot recur.
Healthcare HR adds its own weight: staff health data is special-category data, and the questions often come from people under strain. See HR adviser training in healthcare in Germany and in India for market detail.
Which controls make this safe?
Three matter most. First, health, conduct and exit topics are held by the employer's own limits, not left to the agent's judgement. Second, advisers approve under their own names, never a shared mailbox identity, so the record shows who sent what. Third, the team keeps its policy library current, because the agent can only be as right as its sources. Where decisions about employees could have significant effects, GDPR Article 22 and its UK equivalent shape how much human involvement is required.
Advisers are the reason this works, a point developed in why HR advisers are the real control. The skills every role shares are set out in AI agent skills by role.
Questions people ask
Do HR advisers need technical training?
No. They need to understand depth, sources, verdicts and the record, which are concepts, not code. Most of the skill is the review judgement they already have, applied to a new kind of draft.
Should employees be told an agent drafted the reply?
Yes. People should know when they are dealing with an AI system, and in HR the trust cost of finding out later is high.
Can the agent answer questions from managers about their staff?
Only within the access the employer grants. A manager's question about one employee's health or conduct should be held for an adviser.