AI Agents in HR: What They Can Actually Do Today

An AI agent in HR is software that chooses its own next step and acts on your systems without a person approving each action. That is the whole definition, and it is worth holding onto this quarter, because most of what is being sold to you as an HR agent does not meet it. In June 2025 Gartner estimated that of the thousands of vendors claiming agentic AI, only about 130 are genuinely agentic. The rest are doing what Gartner named agent washing: “the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities.”

That is not automatically a scandal. A well-built assistant with a workflow attached to it can save your team real hours. The problem is paying agent prices and accepting agent risk for assistant capability, and then discovering the gap in month four when something goes wrong and nobody can reconstruct why.

So the useful question in a demo is never “is this an agent.” It is: what is this allowed to do without asking me, and what happens when it is wrong? Everything below is organised around that question.

Here is a test you can apply in a meeting without a technical background. Ask what the software does between the moment you give it a goal and the moment it comes back. If the answer is that it decides which steps to take, in what order, and acts on live systems to do it, you are looking at an agent. If the answer is that it presents you options and waits for a click at each stage, you are looking at an assistant with good user experience.

Neither is a bad thing to buy. They are different things to buy, and the label changes three practical matters: the price, the autonomy you are being asked to grant, and where liability sits when the output is wrong. A chatbot that gives a wrong parental leave answer creates a ticket. An agent with write access that acts on a wrong parental leave answer across forty employees creates a project.

The ai agent vs chatbot hr distinction gets muddied deliberately, and Gartner’s language gives you cover to push back in a sales conversation without sounding hostile. The firm has now named the same pattern twice in different markets. In May 2026, at its supply chain event in Barcelona, it warned about agent washing in supply chain planning software — a different market from HR, and I flag that because the finding does not transfer as evidence about HR products. What transfers is the shape of the problem. Gartner’s assessment there was that most current agentic capability improves the user experience through query interpretation, recommendations and conversational support, rather than changing how decisions actually get made. If you have sat through four HR agent demos this year, that sentence will feel familiar.

Every capability list on page one of this search organises itself by HR function: recruiting agents, onboarding agents, HR service desk agents. That ordering is useless to you, because it tells you what the vendor built rather than what you can safely allow.

Organise by autonomy instead. The question is not which department the thing sits in, it is how much blast radius the task carries when the software gets it wrong.

Tier	What it means	Test
Runs unattended	Reversible, high volume, cheap to get wrong	If this is wrong 5% of the time, does anyone get hurt?
Human confirms each consequential step	Real-world side effects, recoverable with effort	Would you want your name on this if it went out wrong?
Never alone	Affects pay, status, record, or a protected characteristic	Could this be defended to a regulator or a court?

These are the agentic ai hr use cases worth your attention, in the order you should consider them.

This tier works now, genuinely, and it is where the boring value lives. First-line policy answers, where the software retrieves from your actual documents and cites which one it used. Drafting: job descriptions, offer letters, interview guides, meeting summaries, all handed to a human for editing rather than sent. Scheduling and calendar coordination. Data-quality sweeps that scan your HRIS for missing fields, duplicate records and impossible dates, and flag them rather than fixing them. Case and meeting summarisation.

What these share is that a wrong answer is visible and cheap. A badly drafted job description gets edited. A flagged record that turns out to be fine gets dismissed in two seconds. These autonomous hr workflows are unglamorous and they are the ones actually shipping.

Candidate sourcing and outreach sits here, and the most persuasive evidence for that placement comes from a vendor rather than a critic. When ADP launched a set of partner AI agents on its marketplace in March 2026, its own description of the talent acquisition agents was that they identify qualified candidates against job requirements, make contact and keep them engaged — while keeping hiring decisions in the recruiter’s hands. ADP also requires partners on that marketplace to meet its responsible AI principles, covering human oversight, bias mitigation, explainability and ongoing monitoring.

Read that concession carefully. The company selling the agent drew the line at the decision. That tells you more than any critique I could write.

The same tier holds interview coordination, onboarding task orchestration, HR service desk triage and resolution, and benefits change requests. The pattern is that the software proposes and a person signs. This is the copilot shape, and it is what our own Vera AI capability is built around: the agent assembles the work and a human commits it. That is a design position rather than a technical limit, and I would argue for it even in products I have nothing to do with.

Multi-week processes with branching judgement. Anything requiring a stable picture of who reports to whom when your systems disagree about it, which is most mid-market companies. Investigations. Performance calibration. Redundancy selection.

And the hard line: anything that writes to pay. A model may propose a number, but the number that reaches a payslip has to come from rules a person can re-run and get the same answer from. That is its own subject and it deserves the space [pending: T-07].

Gartner’s own framing of why this tier stays stubborn is worth quoting, because it is the analyst firm rather than a sceptic: “Most agentic AI propositions lack significant value or return on investment (ROI), as current models don’t have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time.” The same June 2025 research predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls. Note what is not in that list: the technology not working. Projects are dying of cost and governance, not capability.

For balance, Gartner also expects real movement — at least 15% of day-to-day work decisions made autonomously by 2028, up from zero in 2024, and 33% of enterprise software applications including agentic AI by then, up from under 1%. This is a category that arrives. It arrives later than the demo implies, in narrower places.

Having watched a few of these go into real HR workflows, I can tell you the failures are almost never the ones the risk register anticipates. Nobody’s agent goes rogue. Here is what actually happens.

Permissions. An agent that can read everything will answer questions it should not. Point one at your HR drive and it will cheerfully surface the compensation review spreadsheet to whoever asks the right-shaped question. The control is scoping read access to a curated set before launch, not after the first incident.

Stale source documents. This is the quiet one. Your agent answers from the policy library, and your policy library has a 2023 parental leave document that nobody deleted. A human answering that question would hesitate. The software will not, and it will be wrong confidently, at volume, in writing. The control is that nothing enters the agent’s source set without a named owner and a review date.

No audit trail of why. Plenty of tools log what happened. Far fewer record what the software was working from when it decided. If you cannot reconstruct the reasoning six months later, you cannot defend the outcome, and HR is a function where you get asked to defend outcomes.

The handover problem. The agent completes 80% of a task and stops, and a person inherits a state nobody documented. Half-finished onboarding is worse than not-started onboarding, because not-started is visible.

The supervision gap. You have given people a tool that proposes actions and no guidance about when to override it. Left alone, they either rubber-stamp everything or ignore it entirely, and you will not know which until you audit.

The unglamorous truth is that the work of shipping one of these is mostly permissions, document hygiene and deciding what the thing may do without asking. The model is the easy part. It was always going to be the easy part.

Take these into your next demo. Each one has a bad answer worth listening for.

“What does it do without asking me?” A specific list is a good answer. “It’s fully autonomous, but you stay in control” is a non-answer, and it is the single most common thing you will hear.

“Show me the log from a run that failed.” Watch what happens in the room. A team that has run this in production has a failure to show you and is mildly bored by the request. A team that has not will offer you a successful run instead.

“What happens when the source document is out of date?” You want to hear about ingestion controls, review dates and citation. If you hear that the model is very good at reasoning over documents, you have your answer.

“What permissions does it hold, and who granted them?” The right answer is a scoped list and a named approver. A vague answer here is the one that should end the meeting, because it is the failure that hits you hardest and soonest.

Notice that none of the four is about accuracy. Accuracy is what the demo is designed to show you. These are about what the software does on a bad Tuesday, which is the thing the demo is designed to avoid.

Start where the task is high volume, well defined, and cheap to get wrong. That is almost the exact opposite of where HR teams instinctively want to start, which is the painful, complicated, judgement-heavy process that eats everyone’s week — precisely the work that has the least tolerance for a confident wrong answer.

Policy retrieval and drafting are the honest first candidates. They are dull, they touch every team, and their failure mode is an edit rather than an incident. Sequencing the rest of it across a function is a bigger question than this article can answer properly [pending: T-05].

What is the difference between an AI agent and a chatbot in HR? An agent chooses its own next step and acts on your systems; a chatbot responds and waits for you to act. The practical difference is autonomy: ask what the software does without a human clicking approve. If the answer is nothing, it is an assistant, whatever the pricing page calls it. Both can be useful, but they carry different risk and should carry different prices.

What can AI agents in HR do without human approval? Work that is reversible, high volume and cheap to get wrong: first-line policy answers with a citation to the source document, drafting for human editing, scheduling, meeting and case summarisation, and data-quality sweeps that flag problems rather than fixing them. Anything with a real-world side effect on a person should have a human confirming the step.

Are AI agents safe to use for hiring decisions? Not for the decision itself. Agents are reasonable at sourcing, outreach and coordination, and vendors generally say so themselves — ADP’s own marketplace framing keeps hiring decisions with the recruiter. Selection decisions carry legal exposure, require explanation, and in several jurisdictions are regulated specifically. Use agents to prepare the decision, never to make it.

Why do so many agentic AI projects get cancelled? Gartner predicted in June 2025 that over 40% would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The technology failing is not on that list. Most cancelled projects were experiments started from hype without a defined problem, and they died at the point someone asked what the return was.

Do we need new HR software to use AI agents? Usually not immediately. Most major HR platforms are adding agent capability to what you already run — Workday, for one, launched developer tooling in June 2026 for building, connecting and verifying agents across HR, finance and IT, with much of it in early access. Check what your existing stack ships before buying a separate tool, and ask what governance comes with it.