Every payroll run starts with a mutation overview: hours, leave, sick leave, contract changes. Someone pulls this data from various systems, checks it and gets it ready. The task stops at that complete overview — the run itself is a separate step. The question is whether AI can take over this collecting and preparing.
The answer is partly. Not because the task is too complicated, but because two axes in the profile weigh heavily against automation and this isn't compensated for by the rest.
The task is fairly structured: mutations come from fixed systems, according to fixed categories, in a recurring rhythm per month or four weeks. That argues in favor of automation. The volume is also substantial — for larger workforces, this involves hundreds of lines per run, which makes a task interesting to automate.
But two axes weigh more heavily: cost of errors and compliance are both at the lowest level. A missed sick leave notification, an expired contract that hasn't been processed, or a leave day that's counted twice, directly affects an employee's salary and the payroll tax return. These aren't errors you can easily correct afterward — they affect someone's income and the relationship with the tax authorities. Judgment scope is also low, at 2: there is little room for interpretation, mutations are correct or incorrect, complete or incomplete.
The combination of high structure with low cost of errors and compliance is exactly where an agent model fits: AI collects, matches and flags discrepancies, but a human approves or rejects before the overview goes out the door to the payroll run. Not a full takeover, but a shift from collecting to checking.
A company with one hundred and fifty employees, an HR system and a separate payroll system that aren't linked. Every month, an HR employee manually goes through the leave records, sick leave notifications and contract changes and types them over. This is where an AI agent can make the comparison once the systems are actually linked: automatically retrieving mutations, flagging duplications, marking missing contract end dates. The agent delivers a proposal; the payroll administrator reviews the exceptions and finalizes the overview for the run. The work of searching and retyping disappears, the work of reviewing remains.
At a company without a link between the HR system and payroll system, this task is still largely human work today, simply because the agent has nothing to read from. At a company with many temporary workers, variable hours or shift work, complexity increases — more source data, more chance of conflicting input, so more moments where a human must intervene. And at a company with a collective labor agreement that includes many special allowances, the judgment scope is again somewhat higher than in this profile, which tips the balance slightly more toward human work.
The condition that recurs in all cases: linking the source systems and a review moment before the final run. Without these two, this remains a task that cannot be automated in isolation.
This task doesn't stand on its own. It's connected to other HR tasks further upstream: how schedules are drawn up, how schedule changes are processed, how overtime is calculated and how absence is registered. Where these tasks are already partly done by AI, cleaner, more structured source data automatically becomes available for payroll preparation — which makes the agent approach easier here. Where these tasks are still entirely manual, payroll preparation also continues to rely more heavily on human work, simply because the input is messy.
That is the shift currently underway: not one task suddenly collapsing, but a chain of tasks that each become somewhat more automatable once the systems behind them are in order. At one company, that chain is already largely linked, at another it isn't yet — and that difference, not the technology itself, determines how many hours are already freed up here today.
Anyone wanting to know how many hours that concretely amounts to should calculate in time, not in headcount — and why that distinction matters is explained on the page about calculating in hours instead of people. It's also important to note what does not belong in an AI system: salary data and medical information about illness fall under data with extra protection, and which data your company may not simply have processed by an AI system for this purpose is described on the page about data your company may not use.
A final note: if this analysis seems to point somewhere toward a personnel decision, a different framework applies. For dismissal, job changes or reorganization, separate legal requirements apply, and this page does not provide a basis for that.
To see how this profile relates to the rest of your HR and payroll process, you can fill in the free quickscan: twelve questions, no account needed, resulting in an indication of what portion of the hours in your situation can already be taken over by AI today. The full work scan, which maps the work of your entire company task by task, is still under construction — we deliberately don't offer that here yet.
Vraag maar. Ik ken de kennisbank van deze site; wat ik niet weet, zeg ik erbij.
Answers come from this site’s knowledge base. Not tailored advice, and not a scan of your company.