AI agents for HR: answering employee questions, handling requests and onboarding new starters

Take a hypothetical manufacturing company with 250 employees, where HR is handled by two people. Every week they answer the same questions. How much leave do I have left? How do I claim travel expenses? When is payday? Who hands out laptops? How do I add my child to the medical plan? Each question takes a few minutes, but together they eat up a good part of their week.

An AI agent is an assistant in Teams that answers such questions based on company policies and procedures. Unlike an ordinary FAQ page, it understands questions asked in people’s own words. It can also start a simple process, such as an equipment request, and pass it on for approval.

The agent does not replace HR staff. It does not settle disputes, conduct difficult conversations or make decisions about people. Its job is to take off HR’s plate the questions whose answers are already written down in policies.

A good test is whether the answer follows directly from a document and is the same for everyone. If so, the agent will cope. If it depends on judging a particular person’s situation, the matter should go to a person. The table below shows a typical split.

New starters have the most questions just when they least know whom to ask. They feel awkward writing to HR for the fifth time in a week, and their manager is in meetings. An agent in Teams answers at any time and does not judge how many times someone asks.

In practice, the agent can take a new starter through the list of things to do in the first days: account, equipment, access card, mandatory training, important policies. It can send a reminder about the date of a medical check-up or training session and give the contact details of the person who deals with a given matter.

However, the agent will not replace a well-planned onboarding process. If the company has no list of steps for a new starter and nobody knows who orders the laptop, the agent will simply say more quickly that it does not know. It is worth sorting out the process itself first, and only then adding an assistant to it.

The simplest agent only answers questions. The next step is requests: an employee writes “I need a second monitor”, the agent asks for details and starts an approval in Power Automate. The manager approves it in Teams, and the employee can ask the agent about the status.

The most work is involved in connecting to the HR and payroll system, for example so that the agent can give the leave balance. This depends on whether the system has an interface for exchanging data (an API), what data it makes available and how it confirms who is asking. For some systems it is a simple configuration, for others a separate project.

That is why it is worth starting without this connection. An agent that gives a link to the employee portal instead of the leave balance still saves a lot of questions. Add the integration once the pilot shows that this is what employees ask about most.

The agent answers based on the documents you point it to – usually policies and procedures stored in SharePoint. If the library holds three versions of the remote work policy, the agent may quote the oldest. It cannot tell the current rule from a draft unless someone tells it.

So before launch, prepare the knowledge using the steps below. This is usually the longest part of the work and the best investment, because tidy policies will help people too.

HR data is among the most sensitive in any company. An agent built in Microsoft 365 sees documents with the permissions of the person talking to it. But if a folder of employment contracts is accidentally open to everyone, the agent can use it too. Reviewing SharePoint permissions is therefore a precondition, not an extra.

The precautionary rule is simple: a general-purpose agent has no access to payroll data, appraisals or personnel files. If it is to provide personal data at all, then only that of the person asking, after their identity has been confirmed through their company account.

Before implementation, talk to your data protection officer, if the company has one. For systems that assess or monitor employees, the GDPR usually requires a data protection impact assessment (Article 35). It is also worth telling employees what the agent is for and what data it can see.

The AI Act, the EU regulation on artificial intelligence, treats employment as a particularly sensitive area. High-risk systems are those that help with recruitment, decide on promotion or dismissal, allocate tasks based on employees’ characteristics or assess their performance. Following the 2026 amendments, the obligations for such systems apply from 2 December 2027.

An agent that answers questions about the leave policy does not fall into this group. It is, however, subject to transparency rules: from 2 August 2026, employees must know they are talking to AI. A clear notice at the start of the conversation and in the assistant’s name is usually enough.

A few rules are worth knowing straight away:

A separate use is help for managers. Microsoft Copilot (with a Microsoft 365 Copilot licence) can draft an announcement about a rota change, summarise notes from a team meeting or suggest a structure for an annual review conversation. This is ordinary writing help, and the decisions remain with the manager.

The risk begins when the tool is supposed to compare employees or indicate whom to promote or dismiss. You then enter the high-risk area under the AI Act and the impact assessment under the GDPR. If the company wants to go in that direction, it needs a separate legal analysis, not an assistant configuration.

Agents for all employees are usually built in Copilot Studio, Microsoft’s tool for creating your own assistants. Conversations are paid for in credits: according to Microsoft’s price list, a pack of 25,000 credits costs USD 200 a month, or you can pay for actual usage. According to Microsoft, an answer generated from documents uses 2 credits, and each agent action, such as creating a request, uses 5 credits. Employees with a Microsoft 365 Copilot licence use such agents without consuming credits. For example, 1,500 questions a month at 2 credits each and 200 requests at 5 credits each come to 4,000 credits, or about one sixth of a pack. Unused credits do not roll over to the next month, so it is worth choosing the pack size after the pilot.

On top of that comes the work: tidying up the policies, configuring the agent, testing against the list of questions and possibly connecting to the HR system. This scope has the greatest effect on the cost of implementation, which is why it is worth starting with a small pilot.

A pilot can last a few weeks and cover one area, such as questions about leave and business travel. Before you start, count how many such questions HR receives each week. After the pilot, compare that number, check the accuracy of the answers against the test list and ask employees whether the agent helped them.

To be honest, not every company needs one. In a company of 40 people where HR is handled by one person who knows everyone, a well-written SharePoint page with the most frequent questions will have a similar effect at a lower cost.

An agent also makes no sense when the policies are outdated or contradictory and nobody wants to fix them. Instead of taking the load off HR, it will start generating questions like “the agent told me something different”. In that case, tidy up the documents first and decide about the agent later.