Article · Agents by Business Function
AI in HR operations: what agents handle and where a person decides
HR is the process where you have to be most careful talking about agents, because it mixes two things that cannot be treated alike: an administrative back office full of repetitive work, and decisions about people’s working lives. An agent fits beautifully in the first and has no business going near the second. Confusing the two is what turns a sensible project into a legal and reputational risk. So I start by drawing the line, then say what sits on each side.
The principle I apply with any board is this: in HR, the agent prepares and the person decides. The agent can gather the information, order the file, draft the summary, coordinate the administrative steps. The decision that affects an employee (hiring, evaluating, promoting, disciplining, dismissing) is made by a person who answers for it. I set out the general frame for what it means for an agent to operate a process in AI agents for business.
The HR back office: where the agent does operate
Most of a people team’s work is not deciding about anyone. It is processing, coordinating and answering questions. That administrative volume is natural territory for an agent:
| Task | What the agent executes |
|---|---|
| Administrative onboarding | Coordinates a new hire’s setup, access, equipment and paperwork, and tracks what is still missing |
| Document management | Extracts data from contracts and documents, keeps files current, and checks nothing is missing |
| Employee queries | Answers policy questions (leave, time off, expenses) with the company’s official information |
| Process coordination | Triggers and follows the steps of recurring processes: periodic reviews, renewals, mandatory training |
| File preparation | Gathers and orders the information a person needs to make a decision, without making it |
The constant is the same as in any agent-run process: volume, known rules and a verifiable outcome. Access is either granted or not. A question about the leave policy has a correct answer written down somewhere. I cover the mechanics of how an agent executes this work with permissions and a trace in the AI agents in operations hub, and for internal employee queries the same pattern fits as in any first-tier support.
The red line: decisions about people
This is where the easy “automate HR” pitch crashes into regulatory reality. Deciding about people (who gets hired, how an employee is evaluated, who gets dismissed) is not just another automation case. It is high-risk territory, and in two senses: the harm to the person if it is done wrong, and the fact that European law explicitly classifies it as such.
The EU AI Regulation places systems used for recruitment, selection and decisions affecting the employment relationship among the high-risk uses, with specific obligations: effective human oversight, transparency, data quality management and traceability. That is not a best-practice suggestion. It is a legal regime. An agent that scores candidates or proposes who not to renew falls squarely inside it, and the company takes on the obligations that come with it.
On top of this sits data protection. An HR file is made of sensitive personal data: health, performance, family situation, sometimes special-category data. Feeding that information into an agent system without a lawful basis, minimization and access control is an infringement waiting to happen. I develop it in AI agents and data privacy. That is why in HR the permission design (which file the agent may see, what it may write, what it must never touch) is not a technical detail, it is the first line of compliance.
The agent prepares, the person decides
The right way to use an agent in the sensitive part of HR is not for it to decide less, it is for it not to decide at all. It can do all the prior work (gather applications, check formal requirements, order the information, draft a summary) and leave the decision, whole and reasoned, in the hands of whoever answers for it. That person should not receive a verdict from the agent that they merely rubber-stamp. They should receive a file to actually decide on, with access to the information and not only to the conclusion.
The difference looks subtle and is enormous. An agent that “recommends” who to hire and a person who approves the recommendation without more is, in practice, delegating the decision to the system, with the legal responsibility intact and human oversight turned into a stamp. The effective oversight the rule demands means the person can reach a different conclusion and act on it, not that they sign what the agent proposes. I have written about governing that boundary between what the system does and what a human decides in AI agent governance.
There is also a technical risk that compounds the legal one: bias. A system that learns from past decisions can reproduce and amplify the biases in those decisions, and in hiring that translates into discrimination. It is not a problem good intentions solve. You solve it by keeping the person as the real decider and auditing what the system proposes. I frame it within the risks of AI for business, where the HR one is among the fastest to materialize.
How I frame it for a board
When a leadership team raises “using AI in HR”, my first question is which side of the line we are talking about. If it is administrative back office (onboarding, document management, employee queries, process coordination), the project is sensible and the return is clear: it frees the people team from repetitive work so their judgment goes where it matters. If what is sought is a system that decides about people, the conversation changes entirely: we enter regulated high risk, and the question stops being “how much do we save?” and becomes “are we willing to take on the oversight, transparency and audit obligations this demands, and to keep a person as the real decider?”.
Most of the time, when the two sides are properly separated, the project that goes ahead is the administrative one, which is where the volume and the savings are without the risk. The decision-about-people part stays human not out of excessive caution, but because it is where the error is most expensive and the law most demanding. A deployment that orders that boundary well is solid. One that erases it to show off automation is a problem that has not gone off yet.
Frequently Asked Questions
What can an AI agent automate in HR?
The administrative back office: onboarding and setup coordination, document management of files, answering employee policy queries, and tracking recurring processes. It is work with volume, known rules and a verifiable outcome. What it should not automate are decisions about people, which stay human.
Can an agent decide who to hire or dismiss?
It should not. Hiring, evaluation and dismissal are decisions that affect people’s working lives and that European rules classify as high-risk uses, with obligations of effective human oversight and transparency. The agent can prepare the file. The reasoned decision is made by a person who answers for it.
Why is HR high-risk territory for AI?
For two compounding reasons. The harm to the person if a decision about their employment is made wrong, and the legal frame: the EU AI Regulation places selection and employment-management systems among the high-risk uses with specific obligations. On top of that, an HR file is full of sensitive personal data subject to data protection.
How do you control bias in a recruitment agent?
By keeping the person as the real decider, not a validator, and auditing what the system proposes against outcomes. A system that learns from past decisions can reproduce their biases, so the defense is not trusting the model but preserving effective oversight and traceability of every proposal so it can be reviewed.