HR Chatbot vs AI Agent: What's Different, and Which Do You Need?

A good answer, then a dead end

Here’s a moment that plays out in a lot of HR teams. It’s a composite made up for this guide, so the person and the details are invented.

Sam’s partner is due in March, and Sam wants to know how much parental leave they can take. At nine in the evening they open the HR assistant in Microsoft Teams and type the question. The answer comes back in seconds. It quotes the right policy, gives the number of weeks and explains how the weeks can be split. Sam is impressed.

So Sam types the obvious next thing: “Can you book the first two weeks for me?” The assistant replies with a link to the leave form in the HR system. Now Sam has to log in and check their balance, then fill in the dates and pick an approver. The assistant was useful right up to the point where Sam needed something done, but then it handed the work back.

That gap is what this guide is about. The tool Sam used is a chatbot. A tool that could have checked Sam’s balance, filed the request and sent it to their manager would be an AI agent. Vendors use both words loosely, and some sell the same product under either name, so the label won’t help you decide what to buy. What helps is knowing two things: what each kind of tool does when it gets something wrong, and who has to own the systems it touches.

HR chatbot vs AI agent: the differences that decide cost and risk

A chatbot reads and replies. It takes a question, finds the answer in your policies and writes it back. An agent can also change things, because it connects to your HR systems and takes steps there on the employee’s behalf. The most important of those systems is usually the HRIS, which is the core system that holds employee records, leave balances and job details.

Most comparisons stop at that definition, so the table below looks at the things that decide whether either one works for you. The entries are our judgment from building these systems, not findings from a study.

	HR chatbot	HR AI agent
When it’s wrong	It gives a wrong or out-of-date answer. The employee may act on it, but nothing in your systems has changed.	It takes a wrong action. A request gets filed for the wrong dates, a record changes, or someone sees data they shouldn’t.
What it needs from IT	Access to the documents it answers from, and a place to live, such as Slack, Teams or the intranet.	Connections into each system it acts in, a login for the agent itself, and a permission for every action it’s allowed to take.
What HR owns	The content. Every policy it quotes has to be current, and someone has to keep it that way.	The content, plus the rules for each action: who can ask for it, when it needs a human to approve, and what gets logged.
Time to first value	Mostly set by how clean your policies are. The software is rarely the slow part.	Mostly set by how easy your HR systems are to connect to, and how long it takes to agree the permission rules.
A sensible first job	Answering policy and benefits questions that have the same answer for everyone.	One high-volume request with a clear approval path, such as booking leave.

Chatbot: rules-based and generative

There are two kinds of HR chatbot, and they fail in different ways.

A rules-based chatbot follows a script someone wrote in advance. It offers buttons or menus, and each choice leads down a fixed path. It never invents an answer because it can only say what’s in the script, but the trouble is that employees don’t ask questions the way the script expects. If someone types “my dad’s in hospital, can I take time off?”, a rules-based bot may not connect that to the compassionate leave policy. As a result, people hit a dead end and give up.

A generative chatbot uses a large language model, which is the same kind of technology behind ChatGPT. It reads your policy documents and writes an answer in plain language. It copes with odd phrasing far better, but it has the opposite weakness: it can sound completely sure of itself while being wrong. That’s why the better tools use grounding, which means every answer is tied to a named source document and the bot shows the employee where the answer came from. If the source isn’t there, a well-built bot says so and hands over to a person.

Chatbots are where a lot of HR teams start, and the survey figures point the same way. IBM reports that a Gartner survey from February 2024 found 38% of HR leaders were piloting, planning or using generative AI. That was up from 19% in mid-2023. Employee-facing chatbots for HR services were the top use case, at 43%.

Agent

An agent does what the chatbot does, and then it acts. An agent would look up Sam’s leave balance in the HRIS, check the dates against the policy and file the request. It would then send the request to Sam’s manager for approval and tell Sam when it’s done. So Sam would ask once, and the agent would do the steps that Sam had to do by hand.

The agent can only do that if it knows who is asking. A chatbot can give the same answer to everyone. An agent has to give Sam’s answer to Sam and nobody else, so it needs a secure way to confirm identity, and it needs permission to read and change that person’s records.

IBM is one of the best-known examples of this at scale. IBM says its internal assistant, AskHR, handles more than 2.1 million employee conversations a year and has automated more than 80 HR tasks. IBM also says employees can submit vacation requests through it, which is exactly the step Sam’s chatbot couldn’t take. Those are IBM’s own figures about its own system. IBM also sells AI software, and in 2025 it added its own watsonx Orchestrate platform to AskHR. That shows what’s possible, but it isn’t a benchmark for a first project.

Why the line between them is blurry

You’ll meet tools that don’t fit neatly into either box. Many generative chatbots can now trigger actions, so a vendor can add a “submit request” step to a chatbot and start calling it an agent. Likewise, some products sold as agents only answer questions in practice.

So don’t decide based on the name, because what a tool can actually do comes from three things. The first is its integrations, meaning which of your systems it’s connected to. The second is its permissions, meaning which actions it’s allowed to take and for whom. The third is governance, meaning the approvals, logs and limits that stop it doing something it shouldn’t. Two products with the same label can differ completely on all three.

My advice is to ask any vendor the same plain question. Which actions can it take, in which of our systems, under whose permission? If the answer is “none yet”, you’re buying a chatbot, whatever the brochure says. A chatbot may be exactly what you need, but you should know that’s what you’re paying for.

How to choose

Start with your own requests, not with the products. Pull the last hundred questions that came into your HR inbox, ticket queue or helpdesk. Answer four questions about them.

1. How many of them end with someone changing a record? Mark each request as “answer only” or “needs something done in a system”. A policy question is answer only. A change of bank details needs something done. If most of your volume is answer only, a chatbot covers most of the value. But if a large share needs a change made, an agent starts to earn its extra cost.

2. Is the answer the same for everyone, or specific to the person asking? “How many days of carry-over do we allow?” has one answer. “How many days do I have left?” has a different answer for every employee. The second kind needs the tool to know who’s asking and to read their record, which is a big step up in what IT has to provide even before the tool changes anything.

3. Who owns the knowledge base? A chatbot is only as good as the policies it reads. If your handbook lives in five places and nobody is responsible for keeping it current, the chatbot will repeat the mess back to employees. Fix that before you buy anything, because an agent makes this worse by acting on what it reads.

4. What does a wrong action cost? Booking leave on the wrong day is annoying, but it’s easy to spot and easy to undo. Changing someone’s pay is a different matter, and so is showing one person another person’s salary. The higher the cost of a mistake, the more approval steps and checks the agent needs. At some point those checks make the agent slower than the form it replaced.

Once you’ve answered those, the choice usually falls into one of three places.

A chatbot is the right answer when most requests are answer only and the answers are the same for everyone. It’s also right when your HR systems are hard to connect to. Many teams should start here and get the content right, and then add actions one at a time later.

An agent is the right answer when a large share of requests ends in a record change and those changes follow clear rules. It also helps if there is one HR system that IT can connect to without a long project. A routine request, like booking a day of annual leave, is a common place to start. Leave cases that involve health or family data usually come later.

Neither is ready when your policies are out of date or scattered, or when nobody in HR can own the content. It’s the same when employees don’t use the chat tools you’d put it in. A Pew Research Center survey from February 2025 found that 55% of US workers rarely or never use AI chatbots at work. Only about one in ten use them every day or a few times a week. That covers workplace chatbots in general and not HR bots, but it’s still a good reason to check that people will use a simple assistant before you give it the power to act.

Running the choice through the 10X Framework

When we decide what to build, we score each idea with our 10X Framework. It uses three lenses, each scored from 1 to 5. Impact is how much the work matters if it succeeds. Trust to earn is how much trust the system has to earn before people will rely on it, so a low score is good here. Readiness is whether the data, systems and owners are in place today. We don’t add the three into one number, because a single score hides which lens is the problem.

Each idea then lands in one of four tiers: Build now, Proof first, Build with guardrails and Not yet. Here’s how that tends to play out for this choice. A policy Q&A chatbot has little trust to earn. If it takes on a real share of the questions your team answers every week and the content is clean, it often lands in Build now. If the impact is smaller, it lands in Proof first. An agent that books leave has more impact and more trust to earn, which often puts it in Build with guardrails, with an approval step and a log. An agent that changes pay or bank details has a lot of trust to earn. At best it lands in Build with guardrails. If the data isn’t clean or nobody owns the result yet, it’s Not yet. These placements are illustrations, and your own scores depend on your systems and content. If you’re sorting several ideas at once, our guide on how to prioritize HR AI use cases goes further.

What each one gets wrong

Every tool makes mistakes. The useful question is what kind of mistake, and what stops it doing harm. A chatbot’s mistakes are about words, while an agent’s mistakes are about actions, and that difference decides which controls you need.

	HR chatbot	HR AI agent
Typical mistake	Gives an answer that’s wrong, out of date, or made up	Takes an action that’s wrong, or takes it for the wrong person
Who notices	Often nobody, until the employee acts on it	Often the employee or manager, but sometimes nobody until payroll runs
How far it spreads	One conversation at a time	Into the HR system, and from there into anything that reads from it
How hard to undo	Easy, once you correct the source and tell the employee	Depends on the action. Some are a click, some need a payroll correction

Chatbot failures and their controls

Failure	Control
Quotes a policy that has since changed	One owner per policy, a review date on every document, and old versions removed from what the bot can read
Invents an answer when the policy is silent	Grounding: every answer cites its source, and the bot says “I don’t know” when there isn’t one
Gives a general answer to a personal question	Clear wording that tells the employee when the answer may differ for them, and who to ask
Never hands over to a person	An easy route to a human for sensitive topics such as health, grievances or anything legal
Nobody uses it	Put it where employees already work, and track which questions it fails on

Agent failures and their controls

Failure	Control
Acts for the wrong person	Strong identity checks, so the agent only acts for the person signed in
Takes the right action with the wrong details	A confirmation step that shows the employee exactly what will be filed before it’s sent
Shows someone data they shouldn’t see	Scoped permissions: the agent can only read what the person asking is allowed to read
Makes a change that needs a human decision	Approval thresholds that route anything above a set level to a person
Makes a change nobody can trace	An audit log of every action, who asked, and what changed
Makes a change that’s hard to reverse	A tested way to undo each action before it goes live

Most of those agent controls aren’t new ideas. HR already uses approvals, segregation of duties and audit trails for people. An agent needs the same controls, written down clearly enough for a system to follow. Our post on AI governance in HR covers how to set them up.

What drives cost and effort

We won’t put a price on either option here, because prices vary widely by vendor and by how much of your setup is already in place. Any figure without a dated vendor source would be a guess. What we can do is show where the money and the time go, so you can ask for quotes that cover the right things.

A chatbot’s costs come from four places:

• Cleaning up the knowledge base, which is usually the biggest job and almost always underestimated.
• Setting it up in the chat tool your employees use.
• Writing the hand-over rules for when the bot can’t answer.
• Someone in HR reviewing failed questions every week and fixing the content behind them.

An agent adds these:

• Connecting to each HR system it acts in. Every extra system makes this harder.
• Designing permissions for each action and agreeing them with IT, security and sometimes legal.
• Building the confirmation, approval and undo steps.
• Reviewing a sample of actions on a regular basis and not only the failures.

The pattern is that a chatbot’s cost is mostly content, while an agent’s cost is mostly connections and controls. If your vendor’s quote covers the software but not these, the real cost is somewhere else in your budget.

It also helps to agree early who owns what. Here’s how it usually splits.

Task	HR	IT	Vendor
Policy content and its accuracy	Owns
Deciding which actions the agent may take	Owns	Advises
Approval rules and thresholds	Owns	Advises	Builds them in
Connections to the HRIS and other systems		Owns	Builds or supplies
Identity checks and access to data		Owns	Supports
The AI software itself and where it runs		Reviews	Owns
Reviewing actions and failed answers	Owns	Supports

Build or buy

Buying makes sense when the assistant comes from a vendor you already use and your requests fit what it does. If your HR system’s own assistant can answer from your policies and book leave in that same system, that’s often the shortest path. Building makes sense when requests cross several systems, or when your rules are specific enough that a packaged tool would need heavy changes anyway. Building on your own systems also means the logic and the content stay yours if you change vendors. Our post on whether to build or buy an AI chatbot goes through that choice in more detail.

Whichever route you take, ask the vendor or builder these seven questions.

• Which actions can it take today, and in which systems?
• How does it confirm who is asking before it reads or changes a record?
• What happens when it doesn’t know the answer?
• Can we see a log of every action? Who can see that log?
• How do we undo an action it got wrong?
• Who updates the content? How quickly does a change show up in answers?
• Where is employee data stored? Is it used to train the AI?

If you want to see which employee requests are good candidates for a chatbot or an agent, the HR Operations & Shared Services area of the HR Agent Map sets them out. It lists agents for answering HR questions, routing cases and producing letters. It also covers employee records and leave. It names a policy question agent as the first one to prove, because that’s the cheapest and lowest-risk place to start. The rest of the HR Agent Map covers the other areas of HR as well.

HR chatbots in Slack and Teams

The best place for an HR assistant is usually the chat tool your employees already have open. People don’t want to visit an HR portal to ask a quick question. If the assistant sits inside Slack or Teams, asking it feels like messaging a colleague.

We’ve built this way before. Risely is our own product, which we built and run. It has an AI coach called Merlin that works natively in Slack and Teams. Merlin is a coaching product and not an HR helpdesk, so it says nothing about HR chatbot results. What it does show is the practical side of living inside a chat tool. Risely keeps what a person does on their own private. When coaching is assigned, the company sees summaries only. An assistant in a chat tool has to know who it’s talking to, and it has to behave differently in a direct message than in a shared channel.

Those same points matter for an HR chatbot or agent. An employee asking about sick leave in a public channel is a very different case from the same question in a private message. So decide early where the assistant is allowed to answer personal questions, and keep anything about a named person out of shared channels.

Next step

If you’re weighing a chatbot against an agent, the most useful thing you can do is look at your own requests and your own systems. That’s what we do in a 30-minute call, where we look at the requests coming into HR and the systems they touch and then tell you whether we’d start with a chatbot, an agent or neither yet. Anything we build runs on your own systems, so the content and the rules stay yours. Book a 30-minute call to talk it through, or explore the HR Agent Map to see where an assistant could help first.

FAQ

What is the difference between an HR chatbot and an AI agent?

An HR chatbot answers questions, usually from your policies and handbook. An AI agent can also take actions in your HR systems, such as filing a leave request or updating an address. Because of that, an agent needs to know who is asking and needs permission for each action. A chatbot’s mistakes are wrong answers, while an agent’s mistakes are wrong actions.

Do I need an AI agent or a chatbot?

Look at your last hundred HR requests. If most of them only need an answer that is the same for everyone, a chatbot covers most of the value. If many of them end with someone changing a record, an agent may be worth it. Both need current policies and an owner in HR. If you don’t have those yet, fix that first.

Can a chatbot become an agent?

Yes, and that’s often the best route. You start with a chatbot that answers well from clean content, and after that you add one action at a time. Each one gets its own permissions, approval rules and log. That way you learn how people use it before it can change anything important.

How much does an HR chatbot cost?

It depends far more on your setup than on the software. A chatbot’s biggest cost is usually cleaning up and maintaining your policy content. An agent’s biggest costs are connecting to your HR systems and designing the permissions. Ask any vendor to price those parts separately from the licence.

Is an HR chatbot safe for employee data?

It can be, if it’s set up with care. A chatbot that only answers general policy questions doesn’t need to see personal data at all. Once it answers personal questions or takes actions, it needs strong identity checks and permissions limited to what the asker is allowed to see. It also needs a log of what it did. Ask where the data is stored and whether it’s used to train the AI.