GUIDE · AI FOR DEPARTMENTS

AI for human resources: an assistant for policies and onboarding

After this guide you will know how to prepare an assistant for questions about policies and onboarding, which documents to approve first, which topics to keep out of scope, and how to check that answers stay correct over time, without ever using it for decisions about people.

Updated october 2026
7 min read
Altovar Technical Editorial Team

1. Why HR can use it for repetitive questions

Every new hire asks the same questions in their first weeks: what time they start, how to request leave, which policy applies to business travel, whom to contact for access. The same questions reach the HR office several times a month, with answers that vary depending on who responds. An assistant that answers from official policies reduces this load and makes answers more consistent. It also gives HR more time for interviews and support.

The benefit is not only about time. Consistent information is also a matter of fairness: two people asking the same question should receive the same answer. For this reason the use case suits a first pilot, provided the assistant is limited to general information and does not enter individual assessments. The owner should verify that this limit is respected in every new version.

2. Prerequisites: up-to-date policies and a clear scope

The premise is that policies exist, are written down and are approved. Leave, working hours, expense reimbursement, safety, code of conduct, onboarding procedures: each document must have a date, an owner and a single version. If rules change by word of mouth or in scattered emails, the assistant cannot know this and will reproduce the confusion that already exists. It is therefore wise to start with an inventory, not with the tool.

You also need to define the scope. Which questions are permitted and which must always go to a person: for example anything concerning individual pay, disciplinary proceedings, sickness, health data or sensitive personal circumstances. Write this down before choosing the tool, because this list is the real scope of use for the assistant. Whoever approves this list should be the HR director, with legal support.

3. The operational steps

The project can start small and expand only if the trial holds up. You do not need to connect the assistant to staff databases. A closed set of approved documents, with clear access rules and an owner for the content, is enough. The first version must be useful to whoever runs onboarding, not to someone who wants to showcase the technology. Start from the questions that come up every week, not the exceptional ones.

Each step ends with a quality check. Ask an owner to read the answers produced and to flag any that are wrong or incomplete. This work seems simple, but it is what makes the assistant reliable over time. If nobody has time for this check, it is better to postpone the project. The check must be built into the service calendar, not left to goodwill.

  1. List the most frequent questions from new hires and from HR staff, with the policy that answers each one.
  2. Verify that each policy is in its approved version, with date and owner, and remove superseded versions from the archive.
  3. Write the list of topics excluded from the assistant, with an indication of who handles them.
  4. Load only approved documents and require every answer to cite the document and section it comes from.
  5. Have the system answer a set of trial questions and compare the answers with those of an HR manager.
  6. Correct the documents where the answer was ambiguous, not only the system's behaviour.
  7. Open the service to new hires with a clear channel to speak to a person, and record the questions the assistant could not answer.

4. Common mistakes to avoid

The first mistake is including employees' personal documents in the assistant, such as appraisals, individual contracts or interview notes. Even if the system is secure, the processing must be assessed separately and the purpose must be clear. The second is uploading draft policies or several versions, leaving the system to choose the one that seems most recent. An unapproved draft must never appear in the archive.

The third mistake is presenting the assistant as an authority that decides. If an answer concerns a worker's right, the company must be able to say who confirmed it. The fourth is failing to plan for what happens when a question is sensitive: the system must recognise it and refer it to a person, without improvising. The standard reply should be written in advance, together with a reference to who can intervene.

5. How to check that it works

Verification is done with a set of trial questions prepared before launch, including the sensitive ones. For each, write the expected answer, taken from the official policy. Then compare the system's answer. Classify the differences: wrong information, incomplete information, or an answer that should have referred the matter to a person. Record each wrong answer together with the document that caused it.

Verification must be repeated whenever policies change and at regular intervals. A document changed without updating the archive produces outdated answers, and this happens more often than people think. Assign an owner to the verification and keep a simple log of corrections. Corrections should be made before the next trial, and the owner should review the log every time.

6. Limits: data protection and decisions about people

An assistant that answers general questions about policies processes limited data and carries a lower risk profile. The picture changes when the system is used to evaluate employees or candidates, to assign tasks, to decide promotions or to analyse behaviour. The EU Artificial Intelligence Act and the GDPR impose stricter rules for these uses, and some practices are prohibited. Check with an adviser before extending the scope.

For employees, transparency is also an issue. Anyone using the assistant must know what it does, what data it processes and how to ask for human review. Inform staff with a clear document and, where internal policy or legislation requires it, involve the employee representatives provided for. The data protection officer should approve the scope before launch. The adviser can indicate which steps require formal review.

7. How to start: the first list of questions

This week, ask whoever manages staff to write down the questions they hear most often from new hires, with the policy that answers each one. Check whether that policy exists, is up to date and has an owner. If an answer is not written down anywhere, you have already found the first piece of work to do, before any tool. The owner should also be able to say who updates it.

Then decide with the data protection officer which topics remain excluded. Only with a clear scope and approved documents does it make sense to evaluate an assistant. If the policies are not ready, the project should wait: an assistant answering with old rules does more harm than no assistant. A postponed service is better than one that gets things wrong confidently and repeatedly.

Frequently Asked Questions

Can the assistant answer questions about a specific employee?

No, it should not. Questions about salary, leave already taken, appraisals, sickness or personal circumstances concern individual data and require an answer from an authorised person. An assistant can explain the general rule, for example how the business travel reimbursement works, but it must not access an individual's data or give an answer on a specific case.

Do I need to tell employees who use the assistant?

Yes. Transparency is good practice and, in many cases, a requirement under the GDPR. Anyone using the tool must know what data is processed, for what purpose and whom to ask for a review. The information notice should be written with the data protection officer and, where internal policy or legislation requires it, shared with staff representatives.

Can the assistant be used to select candidates?

This is a very different use and must be assessed separately. Selection directly affects people and may fall within the high-risk systems covered by the AI Act. Some forms of analysis are prohibited. Before considering it, check with an employment law adviser and the data protection officer, and do not use the tool as an automatic filter.

How do I keep the answers up to date?

You need an owner for each policy, a review date and a process that links any change to the document with an update to the archive. Every time a rule changes, the archive must be updated and the verification repeated. It helps to keep a list of trial questions, so the check is quick and does not depend on one person's memory.

How important is the provider?

Very, for the technical side and for security. Ask where data is stored, for how long it is kept, whether it is used to train models and who can access it lawfully. Also ask how the provider handles changes and support. Decisions about content and scope, however, remain with the company, not the provider.

About this content

General informational guide. Regulations and technical requirements change over time: always check the official sources and assess your specific case with a qualified technical or legal advisor.

GET IN TOUCH

Talk to our team.

We'll show how Altovar integrates into your infrastructure. No commitment, no sales pitch. Just a technical conversation.

BOOK A DEMO →