In the previous guide, we looked at what to expect from an AI agent and the six categories of use in HR. Now it’s time to put it into practice.
The goal of this content is simple: by the end, you’ll have set up your first agent, ready to go. Not a demo, not a theoretical proof of concept—an HR assistant you can start using tomorrow.
And don't worry: there isn't a single line of code. If you know how to write an email, you know how to set up an agent.
A reminder of the key point: An agent is only as good as the data foundation it’s built on. This guide teaches you how to configure it; that foundation (your HRIS) is what makes it reliable. We’ll come back to this at the end.
You can set up an agent using any large language model (LLM).
The main LLMs are as follows:
| Tool | Publisher | Distinctive Feature |
| Claude | Anthropic | Excellent at writing and logical reasoning, known for his thoroughness and caution, a valuable asset in HR. The Projects feature makes creating agents very easy. |
| ChatGPT | OpenAI | Widespread, large ecosystem (the "GPTs" act as agents). |
| Gemini | Seamlessly integrated with the Google Workspace environment. | |
| Mistral (The Cat) | Mistral AI | French publisher, citing sovereignty and European hosting as key arguments. |
In this guide, we'll use Claude as an example (the same logic applies to others).
Why? In HR, we deal with sensitive text and high-stakes issues: the quality of the writing, the rigor of the reasoning, and the soundness of the model matter more than anything else. That’s exactly Claude’s area of expertise.
The one thing that's non-negotiable: privacy. First and foremost, choose a professional or business plan—not a free consumer account.
Why?
Enterprise plans (Claude Team/Enterprise, ChatGPT Enterprise, etc.) contractually guarantee that your data will not be used to train the model and provide an appropriate security framework. This is non-negotiable when it comes to HR data. We’ll cover compliance (GDPR, AI Act) in more detail in a separate post in this series.
1. Create your account for the selected business plan.
2. Check with your IT department (or in the settings) to make sure that training on your data is disabled; this is the default setting for enterprise plans.
3. That's it. You're ready to create your first agent.
In Claude, an agent takes the form of a Project: a dedicated space that stores its instructions and reference documents, so it doesn't have to explain everything all over again each time.
1. In the side menu, click Projects → New project.
2. Give it a clear name, just as you would name a job title: for example , “Onboarding Assistant” or “Payroll Preparation Assistant.”
3. A brief description of its mission.
One agent = one task. Don't create a "jack-of-all-trades" agent: create one for each recurring task. It's more reliable and easier to manage.
This is the most important step. The instructions ( or the Project's "custom instructions") define who the agent is, what it does, and, most importantly, what it must not do .
A good training program consists of four components: Role · Mission · Framework & Guidelines · Response Format. Here is a template you can copy and paste and adapt—in this case, for an onboarding specialist:
This is the most important step. The instructions ( or the Project's "custom instructions") define who the agent is, what it does, and, most importantly, what it must not do .
A good training program consists of four components: Role · Mission · Framework & Safeguards · Response Format. Here is a template ready to copy, paste, and adapt—in this case, for anonboarding specialist:
ROLE
You are an HR assistant specializing in onboarding new employees
employees at [COMPANY NAME].
MISSION
Based on a new hire's information (position, team, location,
(start date), you can create a personalized onboarding program on
90 days, and you'll be answering his common questions.
FRAMES & RAILINGS
- You must rely SOLELY on the reference materials provided
(welcome booklet, internal policies). If a piece of information is not included there
If you can't find it, say so clearly, and never make up a rule.
- You do not process any sensitive personal data (health, family status
family status, individual compensation).
- You make the recommendation; the final approval always rests with HR or the manager.
- Tone: professional, clear, and friendly. No jargon.
RESPONSE FORMAT
A structured, week-by-week plan, with the following for each step:
the objective, the contacts, and the resources to be shared.
Three golden rules for effective instructions:
- Be specific about the “don’ts.” The safeguards are just as important as the mission.
- Practice humility. “ If you don’t know, say so”—that prevents 90% of delusions.
- Remember who makes the decisions. It's always “you suggest, the human approves.”
An agent without data is like an expert who hasn't been told anything about the company. This is where it all comes down to.
In your Project, add the reference documents that the agent should consult: orientation booklet, internal policies, letter templates, job descriptions, etc. The agent should read these first, before reviewing general knowledge materials.
And that’s where the HRIS becomes crucial. The documents you add manually are useful to get started, but they freeze the information at a specific point in time. The real, dynamic data that an HR professional relies on—headcount, positions, absences, payroll variables, and historical records—resides in your HRIS, where it is structured and constantly updated.
An agent connected to a dedicated HRIS platform (such as quarksUp) does not operate based on outdated information: it reads the current reality and can write its own data back into it. This is the difference between a “gimmicky” agent and a reliable one. The specifics of this connection are covered in the final installment of the series (interfacing, with simplicity and security).
For this first agent: start with a few reference documents. That's more than enough to get the hang of it.
Don't expect perfection on the first try. An agent is fine-tuned the same way you train a new recruit: through iterations.
1. Give her a real-life scenario ( anonymized for this test): “New hire: Léa, Communications Specialist, Marketing Team, Lyon, starting September 2.”
2. Read the response from an HR perspective, not a technical one. Is the tone appropriate? Are the steps correct? Did he make anything up?
3. Revise the instructions accordingly . Too long? Add “be concise.” Is the student making up a rule about leave? Reinforce the guideline: “Use only the provided materials.”
4. Repeat two or three times. After a few loops, the agent is set.
That's normal and healthy: you don't "configure" an agent—you raise one.
Set up an agent = create a Project, write its instructions (role, mission, safeguards, format), provide it with its documents, and then iterate.
No technical skills are required—just a clear understanding of what you're looking for.
Choose a business plan: HR data privacy is non-negotiable.
An agent is only truly reliable when connected to a dynamic, structured database: your HRIS.
Manually added documents are enough to get started; the HRIS makes all the difference in production.
This guide is the second in a series designed to walk you through the process step by step:
1. The Catalog of AI Tools for HR.
2. You are here: How to set up your first Claude agent (or another LLM) from scratch.
3. 10 Use Cases: Claude in Everyday Life, Connected to Your HRIS.
4. Regulation & Compliance: The AI Act and GDPR—What an HR Director Needs to Ensure.
5. Integration: Connect the AI to your HR software—simply and securely.
