I have spent a lot of the last year in rooms with HR teams, which is not where I expected a career in offensive security to lead. It turns out HR is a good place to learn what an AI agent is for. The work is full of repetitive, rules‑based, slightly tedious tasks, and the people doing it have a sharp sense of what they are not willing to hand over.
So I started keeping a list. Every use case that came up in a workshop went on it. Then we watched which ones the team was still using a month later. Five survived.
1. Policy questions, with the source attached
"How many days of parental leave do I get if I joined in March?" A large share of an HR inbox is this question in a hundred forms. An agent that reads the policy handbook, answers in plain language and quotes the paragraph it used was the first thing every team wanted and the first thing they trusted.
The detail that made it stick was the citation. Nobody accepted an answer without the source paragraph next to it. Once it was there, people stopped emailing HR and started reading the policy, which was the whole point.
2. Interview scheduling
Coordinating five calendars across two time zones and a candidate who can only do Tuesdays is exactly the kind of job a person resents and an agent does not. The agent proposes slots, chases the people who have not replied, and books the room. A human still sends the offer. Nobody argued about this one.
3. First drafts of job descriptions
Give the agent the role's competencies, the level, the team and two examples of past postings, and it produces a draft in the house style. HR still edits every line, and that is fine. The gain is starting from something instead of a blank page, and the consistency across postings improved because the agent kept reusing the same structure.
The team was clear that the agent never publishes. Draft, yes. Post, no.
4. Onboarding checklists
A new joiner in Singapore in engineering needs a different set of accounts, forms and introductions than a new joiner in Hong Kong in finance. An agent that assembles the right checklist from the role, location and start date, opens the tickets, and nudges the owners saved real hours. It also caught the things people forget, like the laptop that had not been ordered.
5. Leaver access reviews
This one came from me. When someone leaves, their access should go with them, and in most organisations it does not, not completely, not quickly. An agent that takes the leaver list, queries the identity systems, lists what is still open and drafts the removal requests for a human to approve is boring, effective security work. HR liked it because it closed a loop they knew was open. I liked it because I have walked into too many networks through an account that belonged to someone who left two years earlier.
The ones that did not make it
Three ideas came up in almost every workshop and died in almost every one.
Screening CVs. The team did not want a model deciding who gets a phone call, and their legal counsel agreed with them. Even as a "ranking aid" it changed how people read the pile. It went.
Summarising performance reviews. Technically easy. Culturally poisonous. People found out and stopped writing honest reviews.
Sentiment analysis of engagement surveys. Nobody could explain to staff why it was not surveillance, so it was.
What the survivors have in common
Look at the five again. Each one does something people were already doing by hand. Each one shows its work, with a citation, a ticket or a draft. And each one stops one step before a decision that affects a person, and hands that step to a human.
That last part is the design rule I now bring to every agent conversation, in HR or anywhere else. Automate the coordination and draft the document, but never let the agent be the one who decides.
