AI Agents in HR: What They Are, Where to Start, and Who Decides

The number of active agents in the Microsoft 365 ecosystem grew 15 times in a year, and 18 times in large enterprises. That’s from Microsoft’s 2026 Work Trend Index. Deloitte cites a Gartner forecast that the average Fortune 500 company “is likely to have 150,000 artificial intelligence agents by 2028.”

Neither number says who’s watching any one of those agents, or what happens the first time one gets something wrong.

You’ll get three questions once an agent goes near your people’s data. Your CEO asks what agentic AI in HR could take on. Your CIO asks what it can reach. Your employees ask who sees what, and who makes the call. Here’s an answer to the last one for 122 real HR agents, counted. The same count shows where HR should start with AI agents.

What counts as agentic AI in HR: three traits

Three traits, and you can check any product against them.

An HR agent reads real HR data. That means the HRIS and the applicant tracking system. It also means payroll, calendars and the policy library. It takes more than one step toward a specific HR job, so it might compare two systems and rank what it found. It might then draft a note to the person who owns the problem.

It also starts on a trigger. Something happens, a schedule comes round, or somebody asks. Many agents on our HR Agent Map start from something that just happened: a candidate moves to an interview stage, payroll inputs close, a hiring manager picks someone to offer.

A chatbot answers the question in front of it. An agent works across steps toward a job, and that’s why HR automation built on agents raises the who-decides question so quickly.

Our answer to the who-decides question is a live, counted map of 122 HR agents across 12 HR areas. Every one has a line saying who decides.

Every agent on our HR Agent Map is built around a person, on purpose

We don’t build systems that make autonomous decisions about people. AI may inform or surface. It may rank, summarize or draft. A person makes every call about a person, and the system is built so it can’t do otherwise.

On the HR Agent Map, that rule shows up twice on every agent’s page. The first is a written line saying who decides. The second is a diagram of the data flow, and the diagram puts the person in a specific spot.

Where a person approves or reviews, the diagram draws them on the path from the agent to everything it produces, labelled DECIDES or REVIEWS. Nothing reaches the outputs without passing through them. Where a person sets the rules, the diagram draws them above the agent as a control point, with an arrow into it and a note that they can override any output. The diagram for the one agent with no people decision draws no person at all, and says so.

Microsoft’s Work Trend Index gets to a version of the same question. It says there are “three questions that every Frontier Firm will need to answer: Who reviews agent performance? Who has the authority to update the workflows that agents run? How does a local win get captured and scaled across the organization?”

The first two are about where a person sits. That’s the line we write on every agent’s page.

The four ways a person sits in the flow

Every one of the 122 agents carries one of four values for where the person sits. Here’s the count, with one example for each, and the who-decides line quoted word for word from that agent’s page.

How a person sits in the flow	Agents	What happens	Example	Who decides, from the agent’s page
Approves: on the path, deciding	66	The agent drafts or proposes, and nothing goes out until a person signs off	Offer Drafting Agent	”The hiring manager, compensation and HR approve the numbers and terms; the agent drafts and flags and does not set pay.”
Reviews: on the path, checking	25	Same spot on the path, and the person checks the agent’s work against what they’d expect	Payroll Anomaly Detection Agent	”The payroll manager reviews every flag and decides what is an error, what is legitimate, and what to correct before sign-off; the agent surfaces and explains, it does not change any pay line.”
Sets rules: off the path, controlling	30	A person sets the rules up front, the agent runs inside them without a sign-off on each run, and any output can be overridden	Interview Scheduling Agent	”The recruiter sets who interviews and in what order; the agent proposes and books slots and any human can override.”
None: no people decision	1	The job isn’t a decision about anyone, so the map draws no gate	Total Rewards Statement Agent	”No people decision is made here; total rewards owns the valuation methodology and approves the statement template, and the employee sees only their own data.”

Add the first two rows together and 91 of the 122 agents never produce an output without a person touching it first.

Approves. The Offer Drafting Agent starts when a hiring manager selects a candidate to offer. It assembles the offer letter and approval packet from the approved template, level, location and pay band. It flags three things: pay outside the band, internal-equity gaps against peers in the same role, and clauses that need a check for that jurisdiction. It then stops.

Its guardrails say it never sets or suggests a pay number, and nothing is sent without the approval chain.

Reviews. The Payroll Anomaly Detection Agent runs when payroll inputs close before each pay run. It reconciles the HRIS, time and absence data and the payroll file. It then ranks what looks wrong: pay swings against the last cycle, leavers still on the file, and negative or zero net pay. The payroll manager works through the list. The agent can’t edit a pay line, the payroll file or the HRIS.

This gate has a weak spot, and the HR Agent Map names it. The first pitfall on the agent’s page reads: “Too many low-value flags and the payroll team stops reading.” A review only protects anyone while the reviewer is still reading, so the proof for this agent passes only if “the false-flag list stays short enough that the team keeps reading it.”

Sets rules. The Interview Scheduling Agent starts when a candidate moves to an interview stage. It works through panel calendars and the candidate’s time zone, plus rooms and video links. It sends the invites and prep material. It handles reschedules without a recruiter stepping in.

None. The Total Rewards Statement Agent builds each employee a personal statement. It covers base pay, variable pay and equity. It also covers benefits, pension contributions and paid time off. Each line comes with a plain-language explanation, and nobody’s pay changes because of it. It reads only the employee’s own record.

Why approves and reviews are split, and what sets rules allows

Approves and reviews share one diagram shape. The person sits on the path and nothing ships without them. Only the label on the gate changes, from DECIDES to REVIEWS. So why split them?

Because the person is doing different work. An approves agent delivers a draft, and the person makes the decision it serves. That decision is the pay, the terms or the wording that goes to a candidate. A reviews agent runs a check on a process the person already owns. The payroll manager reads each flag and judges whether it’s an error or a legitimate swing, like a bonus month or back pay.

Twelve agents on the HR Agent Map do jobs where the work itself is a decision about a person. The jobs include screening applications and pay options for one person. They also include a promotion case, a PIP and a restructure. The map’s verdict for those twelve is human decides. All twelve carry approves. None of them reviews, sets rules or runs with no gate.

The second question: if a person only sets the rules, is the agent acting on its own?

Every output stays inside rules a person set, and any human can override any output. The recruiter sets the panel, the order and the hours for Interview Scheduling. The recruiter can cancel any slot, and the log records who overrode what. Its first guardrail reads: “Books only within the panel and order the recruiter set.”

What the growth numbers don’t tell you

Go back to the numbers at the top. Fifteen-fold growth in a year, and a forecast of 150,000 agents in the average Fortune 500 company by 2028. Both count agents. Who decides on each of them is a separate count, and neither source gives it.

Deloitte’s research on how AI agents fit into the human workplace puts two findings next to each other. In a Deloitte study of senior managers and above, two-thirds said disruptions from multi-agent AI could significantly transform how work is performed within the next two to three years. Only about a quarter said their business processes or workforce are prepared for that transformation.

The same piece reports that nearly 70% of respondents to a recent Deloitte survey named “an inability to trust and govern agents” as a major barrier to realizing value from agentic technology.

Here’s our reading, which Deloitte doesn’t state itself. Agents are arriving faster than the answer to who decides on each one. On the employee side, trust tracks with use. Deloitte’s TrustID research finds that workers with high levels of trust in their organizations are 3.8 times as likely to use agentic technology.

Where should HR start with AI agents?

HR should start with AI agents on proven, low-risk work where the data already exists and a wrong answer is cheap to fix. Answering policy questions and booking interviews are two examples from our HR Agent Map. High-sensitivity agents and agents that prepare a decision about a named person come later, once the early ones have earned trust.

That order comes from the HR Agent Map itself. Every agent page says where the agent sits in the build order. The three places are “a first build”, “after a first win” and “once trust is earned”. Our map marks 42 of the 122 agents as a first build. None of those 42 is marked high sensitivity. Thirty-five of them carry our proven verdict. We give that verdict to work that’s widely deployed, well understood and low-risk to build well.

The last group looks different. Eleven of the 12 agents whose verdict is human decides sit there, and so do 30 of the 39 agents marked high sensitivity. The HR Agent Map’s area pages give the reason: “High-sensitivity agents need notice, consultation, or minimum group sizes before they run, which is why they usually come later in the order.”

Each of the HR Agent Map’s 12 area pages also names one first-proof candidate, under the heading “If you did one thing in this area first.” All 12 carry the proven verdict. None of them is high sensitivity.

HR area	First-proof candidate	Why the HR Agent Map puts it first	How a person sits
Workforce Planning & Org Design	Position Data Hygiene Agent	Low sensitivity, the data already exists, and it removes the month-end reconciliation scramble	Approves
Talent Acquisition	Interview Scheduling Agent	The fastest way to show an agent doing real work in the stack without going near a hiring decision	Sets rules
Onboarding	Pre-boarding Paperwork Agent	Provable in one hiring cycle, and later onboarding agents reuse the starter feed it sets up	Sets rules
HR Operations & Shared Services	HR Policy Question Answering Agent	Cheapest to prove and lowest risk, and it shows where policies and the knowledge base are broken	Sets rules
Compensation, Benefits & Payroll	Benefits Q&A Agent	Plan documents are already written, the questions repeat, and the proof needs no individual data	Sets rules
Performance & Talent Management	Self-Review Assistant Agent	The employee controls the data and the draft, with no rating, ranking or manager view involved	Approves
Learning & Development	Compliance Training Tracking Agent	The data already exists, and an audit-ready record is a result the CFO and general counsel both recognize	Sets rules
Manager Enablement & Coaching	Manager Coaching Agent	It needs no HR data, only a chat platform and a privacy agreement, and managers ask for it	Sets rules
Employee Engagement & Listening	Survey Fielding & Response Rate Agent	One survey wave shows reclaimed hours, with no risk to anyone’s anonymity	Sets rules
Employee Relations, Policy & Compliance	Policy Q&A Agent	It touches no case data, and HR partners can grade every answer against the source clause	Sets rules
People Analytics & Reporting	Headcount & Cost Reporting Agent	The pack already exists every month, so two parallel cycles give a clean test	Approves
Offboarding & Alumni	Exit Process Coordination Agent	Every exit runs through it, the tasks are visible, and it makes no judgment about the person	Sets rules

Nine of those 12 run inside rules a person sets, and so do 21 of the 42 first builds. Any of their outputs can be overridden. The Manager Coaching Agent is the one we know best. Culturro built it as Merlin, the AI coach for managers in Risely, our own product. We’ve run it inside enterprises for more than two years.

The HR Agent Map’s order is a starting point, and two things can change it. Some ATS, HRIS or LMS suites already ship a version of a proven agent. If one you run does the job well, switch it on. An agent earns its place when the native feature is missing, rigid or doesn’t respect your rules. The map is also generic. Your own order depends on which HR areas cost you the most and which agents your data can support.

We scored three real HR ideas end to end in our guide to evaluating HR AI use cases. A policy helpdesk goes first, because it’s the cheapest place to be wrong and the fastest place to be seen being right. Our post on AI in HR covers what to prove before an emerging agent goes any wider.

---

If you’ve got a shortlist of agents from a vendor demo or your own team’s ideas, give each one a line. The line names who approves, who reviews or who sets the rules. Where no person decides, it says that no people decision is made. An agent you can’t write that line for is one to hold.

The 10X HR-AI Framework scores one idea in eight questions, and who decides is part of the score. The strategy month does it for your whole list. Each idea is scored and sequenced, and each one gets a named decision-owner, drawn the way the HR Agent Map draws one for each of its 122. The month is a fixed fee. Each use case after it is a fixed fee for an agreed scope, quoted after the proof, and hosting and model usage are billed by your own providers.

Book a 30-minute call and bring the shortlist.

Frequently asked questions