Among OpenAI's enterprise customers, recruiting is one of the functions where agent use is growing fastest: since February 2026, weekly active users of Codex, OpenAI's agent, grew 41-fold in that area (OpenAI, From assistance to execution 2026). With AI in human resources, HR has a double role: it uses AI in its own work and it usually designs the training for everyone else. Which is why it pays to start with its own work before training the rest.
What is AI used for in human resources?
AI in human resources is used mostly to draft text, summarise information and analyse answers: writing job ads, summarising interviews, building onboarding material, answering questions about internal policy, and grouping the comments from engagement surveys. In all of these the AI writes the first draft and someone on the team reviews it before it reaches a candidate or an employee.
These tasks can be done with the tools your company already has — Claude, Microsoft Copilot or ChatGPT — with no technical project first. AI for recruiting usually starts with job ads and interview summaries, because they are repetitive tasks where a person always reviews the result.
| Task | What the AI does | Who decides |
|---|---|---|
| Job ads | Drafts from the role profile | The hiring lead reviews tone and requirements |
| Interviews | Summarises the notes and sorts answers by competency | The interviewer assesses the candidate |
| Onboarding | Prepares guides, task lists and welcome material | HR validates the content |
| Internal policy | Answers questions citing the source document | HR reviews sensitive answers |
| Engagement surveys | Groups open comments by theme | HR interprets the results and decides what to do |
Which HR tasks can AI do, and which should a person decide?
AI can prepare, sort and summarise, but decisions that affect a person — rejecting a candidate, rating a performance, closing an exit — have to be taken by people on the team. The working rule is simple: the more a result affects someone's working life, the more human control the task needs.
In our AI consulting we grade every task on a three-level scale: AI executes and a person reviews, AI accelerates and a person validates, or a person leads and AI assists. In HR, drafting a job ad sits at the first level, while screening candidates sits, at most, at the third.
Three models of human oversight
For each decision point it pays to choose, and write down, how a person oversees it:
HR · Oversight
Three models, and the task that fits each.
Human in the loop
A person in the process
Steps in on every decision before it takes effect. Fits screening and appraisal.
Human on the loop
A person overseeing
Watches the results and steps in only when needed. Fits answers to common policy questions.
Human in command
A person in command
Sets the rules and can override any result while the AI does the work. Fits job-ad drafts and onboarding material.
Source: Ironhack for Business.
The less direct oversight there is, the more testing and governance the system needs. Which is why it makes no sense to apply the same control to a job-ad draft as to a screening decision.
Pointing at the document instead of answering
When a task comes close to decisions about people, a safer alternative is for the AI not to answer but to say which document and which page holds the information. Asked about leave, for example, the assistant returns the exact clause of the agreement or the internal policy, and someone in HR gives the final answer. Cutting the AI's role from deciding to pointing limits the risk and keeps the time saved on the search.
Reviewing what AI produces costs time too
Using AI in HR does not remove the review work, it moves it. In a survey of workers already using AI at companies with annual revenue above 100 million dollars, 42% of employees in Spain spend up to 1 hour a week reviewing AI-created content, above the global average of 37% (Workday, Beyond productivity: measuring AI's real value 2026).
That review is necessary, but it is also tiring: when one person approves dozens of results a day, attention drops and errors get through. Decide from the start what is always reviewed, what is sampled, and who answers for each kind of result, so the review does not depend on individual goodwill.
Why should HR lead AI training?
HR should lead AI training because it already designs the company's training plans and knows what each role needs. The need is recognised: 66% of Spanish leaders believe employees must update their skills, roles and responsibilities to respond to AI's impact (The Adecco Group, Leading in the AI era 2025).
The training exists too. 67% of organisations in Spain run AI-skills programmes, according to the same report. The challenge is getting that training to people: in Spain, 68% of directors prioritise reinvesting the productivity gained into learning and development, but only 53% of employees see a real increase in those initiatives (Workday, Beyond productivity: measuring AI's real value 2026).
Training that is not applied to the team's real tasks changes little. Which is why we measure tasks, not attendance: the hours a process takes before the training against the hours it takes 60 to 90 days later.
Internal champions in every area
One model that gets training across a whole company is to train internal champions who keep the knowledge alive in their own area. In the AstraZeneca case, an intensive 40-hour programme trained 30 internal champions and every area had its own data reference point. That programme was about data culture, not AI, but the model is the same: HR picks the people, designs the path and measures the impact.
What AI use policy does HR need?
HR needs an AI use policy that sets which tools are approved, what data can be used and who reviews each result. Today, 37% of companies in Spain have no defined policy on the use of artificial intelligence, which leaves it to employees to decide how to work this technology into their jobs (The Adecco Group, Leading in the AI era 2025).
When you apply artificial intelligence to human resources, that missing policy weighs more than elsewhere, because the team works with CVs, personal data and appraisals. If everyone uses their own tool, the company does not know what information leaves or where it ends up — which is what shadow AI means.
The minimum is three pieces:
- a use policy saying which tasks can be done with AI and which cannot,
- a list of approved tools,
- and a record of the team's AI use.
You can check what evidence you are missing with the Article 4 test of the AI Act.
Confidentiality and traceability are the same demands a legal team faces: we explain them in AI for lawyers. And if the tool acts on its own, apply the rules for AI agents.
How to start with AI in human resources: 4 steps
- Take a use census. Before choosing tools, find out what your team already uses and what for. Our workshops start with a 10-minute pre-session survey that tailors the session and gives you the first usage data.
- Train the HR team on its own tasks. Our hands-on workshops on productivity can be run per department, with Claude, Copilot or ChatGPT depending on the tools your company already has.
- Close with concrete commitments. Every workshop ends with commitments to apply what was learned, signed in the room.
- Measure before you scale. Compare the hours each task took before the training with the hours it takes 60 to 90 days later. If the result improves, take the method to the next task or the next team.
When a procedure works, save it as a Claude skill so the whole team repeats it the same way.
Next step
If you want to know which task to start with in your HR team, we will come back with a starting point in under 1 working day.