Which HR Operations Should I Hand to AI Agents First?
Hand AI agents the HR operations that run on written rules and high volume first: answering policy questions from the handbook, onboarding paperwork and account requests, absence and leave records, and checking payroll inputs before the cut off. Keep every decision about a person, hiring, pay, performance and dismissal, with a named human, and build the four agents one at a time with 30 days in shadow each.
Why companies start HR automation in the wrong place
Most CEOs I talk to begin with recruitment. It's the loudest cost and the vendors pitch it hardest. So the first agent ranks CVs, and within a month someone asks why it dropped every applicant with a gap in their history. That's the one place in HR where the law has already drawn a line. The EU AI Act puts AI that filters job applications in its high risk class, and UK GDPR gives a candidate the right not to face a solely automated decision with a significant effect on them.
Meanwhile the real cost sits in the HR inbox. In a 100 to 300 person company I usually find 2 or 3 days a week of HR time going on questions the handbook already answers, joiner paperwork, sick day records and chasing payroll changes before the 20th. None of it needs judgement. All of it needs a rule and a record. That's where the first agent goes.
The order I build HR agents in
1. Sort every HR task by volume and by rule
Take one month of the HR inbox and the HR system's change log. List every task, how many times it happened, and whether a written rule decides the outcome. A task that happens 50 times a month and has a rule in the handbook is an agent job. A task that happens twice and needs someone to read the room isn't. I've never seen this list come back with fewer than 30 tasks, and the top 4 by volume nearly always carry over half the hours.
2. Start with the policy inbox
The first agent answers questions about leave, expenses, benefits, notice periods and working from abroad. It reads only the handbook and the employee's own contract terms, and every answer quotes the clause it came from. Anything the handbook doesn't cover goes to a person, with the question and the 2 closest clauses attached. In shadow for 30 days, the HR lead reads each draft before it goes. I want 9 of 10 drafts sent unchanged before the agent answers alone. When it fails that bar, the fix is usually the handbook, not the agent.
3. Then onboarding, in the joiner's language
The second agent starts the moment a contract is signed. It opens the accounts request with IT, books the laptop, sends the right to work checklist, chases the manager for the first week's plan and builds a day one pack. It writes that pack in the joiner's own language if English isn't their first. The measure is simple: on day one, does the person have a working login, a laptop and a named buddy by 10am? Most companies I walk into hit that for about half of joiners. The target is every joiner, within 2 hiring cycles.
4. Absence records and the payroll input check
Payroll rarely goes wrong in the calculation. It goes wrong in the inputs: a leaver confirmed after the cut off, a bonus agreed in a corridor, unpaid leave never logged. The third agent keeps absence records straight from the request and the manager's approval. The fourth runs 3 working days before cut off. It compares starters, leavers, pay changes and absence in the HR system against the payroll file and lists every difference. A person clears the list and signs the run. The agent never changes a figure in payroll.
5. Draw the red line on paper before you build
Write down what no agent decides, and have the CEO sign it. Mine is 5 lines long.
| HR task | Agent does | Human decides |
|---|---|---|
| Policy questions | Answers from the handbook, quotes the clause | Anything the handbook doesn't cover |
| Onboarding | Requests, chasing, day one pack | The manager owns the first week's plan |
| Absence and leave | Records it, flags patterns to the manager | Any welfare or capability conversation |
| Payroll inputs | Lists every difference 3 days before cut off | Clears the list and signs the run |
| Hiring, pay, performance, dismissal, grievance | Paperwork only, after the decision | All of it, every time |
Each agent gets its own login, read access only to the fields it needs, and a log of every record it opens, kept for 12 months. HR data is the most personal data a company holds. I cover the access register in data governance for AI agents.
From my operating seat
I'm Hayat Amin, an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. I spent twenty years in the C-suite, with three exits and three FT100 listings. Now I build and run AI agent systems inside companies, across finance, sales pipeline, onboarding and compliance evidence, and I coach CFOs one to one.
I did payroll off a spreadsheet in the early days and owned it later, once headcount made that impossible. The calculation was never the problem. The inputs arriving late were, and every delay landed on the same 2 days each month. Inside one client today, the onboarding agent sends each joiner their day one pack in their own language before they start. That's the agent I'd copy first if I were you, because every new hire feels it in the first hour.
How can AI be used in HR without putting the company at risk?
Use it on the paperwork around people, never on the judgement about them. An agent can answer a leave question from the handbook, build a joiner's day one checklist and flag a missing bank detail before payroll. It shouldn't rank candidates, set pay or score performance. Give it a separate login, read access only to the records it needs, and a log of every record it opens.
Can AI agents screen CVs?
They can sort and summarise, and I don't let them reject anyone. Under the EU AI Act, AI used to filter job applications is classed as high risk, and UK GDPR gives people the right not to face a solely automated decision with a significant effect. My rule: the agent checks the hard requirements you wrote down, like a right to work or a licence, and a recruiter reads every CV it would have dropped.
Will AI agents replace the HR team?
No, they take the admin off it. In a 100 to 300 person company I usually find HR spending 2 or 3 days a week on questions and paperwork a rulebook could answer. Move that to agents and the same team spends the time on hiring managers, retention and the hard conversations. I haven't yet cut an HR headcount because of an agent, and I wouldn't plan a business case on it.
Where I come in
This is the work I build and run inside companies as an AI agent operator. We start with one month of your HR inbox and the change log, sort it in the first week, and the policy agent is in shadow by week 3. If you haven't written your processes down yet, start with mapping a process for an agent, and the wider picture is in running a back office on AI agents. Finance leaders who own HR as well will find the rest at AI for CFOs. How many HR questions landed in your inbox last month? Start at meethayat.com and tell me.