What should HR advisers never let an AI agent do?

Decisions about a person's job, pay and dignity stay with the HR adviser. Never let an agent decide a disciplinary or grievance outcome, respond alone to a disclosure of harm, change terms or pay, infer someone's health, share one employee's information with another, or send a letter with legal effect. The agent may prepare these; it decides none.

Why does HR need a fixed list?

HR agents spend most of their time on harmless work: policy lookups, holiday balances, standard letters. That success creates the risk. After months of good drafts, an adviser starts approving on sight, and the one message that mattered gets the same glance as the rest.

A short, fixed list of red lines protects against that drift. It tells the agent where to stop, tells advisers where to slow down, and gives the employer a clear answer when an employee or a tribunal asks who made a decision.

Which acts stay with the adviser?

1. Deciding a disciplinary, capability or grievance outcome. These decisions change careers and must be reasoned by a person who heard both sides. The agent may summarise the evidence and draft the outcome letter for the decision-maker.
2. Responding alone to a disclosure. When an employee reports harassment, discrimination, a safety concern or wrongdoing, the first reply sets the tone of everything after it. The agent acknowledges receipt only if the employer has approved that wording, and routes the case at once.
3. Changing pay, hours or contract terms. A change to terms is a contractual act. The agent can prepare the variation letter; an adviser and the authorised manager approve it.
4. Inferring health or other special-category facts. An agent that guesses at a condition from absence patterns creates sensitive data nobody asked for. It works only with what has been formally recorded.
5. Sharing one person's information with another. A manager asking about a colleague's medical history, or one employee asking about another's pay, gets a held reply.
6. Sending anything with legal effect. Dismissal letters, final warnings and settlement offers go out under a named person's authority.

Wealth advisers face a similar structure of client-facing red lines; see what wealth advisers should never let an agent do. Collections teams, who also deal with people under pressure, keep theirs in what collections agents should never let an agent do.

How does the system hold the line?

The red lines live in the employer's own limits, outside the agent. When the agent tries to send a reply on a held topic, the verdict comes back held and names the bound that decided, so the adviser knows why it is waiting. Attempts at acts the employer has ruled out entirely come back refused.

Every one of those verdicts is filed in an append-only record. If an employee later says "the system told me my contract was changed", the employer can show exactly what was drafted, who approved what, and when.

What does a red line look like in practice?

Take a care home group's HR team. An employee writes late on a Friday: "My shift lead keeps commenting on my body, and I don't want to come in on Monday. Can I swap shifts?" The agent recognises the shift-swap request and could process it at depth 3. It also recognises a possible harassment disclosure.

The verdict is held. The agent drafts two things for the adviser: an approved acknowledgement and a shift-swap request marked "pending adviser". The adviser calls the employee that evening, arranges the swap personally and opens a case. Had the agent answered the shift question alone, the employee would have received an efficient reply that ignored what they actually said.

UK and Indian healthcare employers handle these cases under different rules; see HR adviser training in healthcare in the UK and in India. A ready-made structure for preparing a team is in the training plan template for HR advisers, and the cross-role view is in AI agent skills by role.

Where the Academy fits

HR advisers who review an agent's drafts take the Operator track of Pryme Intelligence Academy, where module tests are set in their own industry and role. To earn Certified Pryme Operator, they complete a two-week practicum of real review decisions that a Pryme assessor signs off; missing a single red line, such as approving a reply to an unrouted disclosure, fails it.