Starting point
The company, around 220 employees, production with shift work and an administration, had an HR team of two. Payroll had been run by the tax advisor's office on DATEV for years, payslips reached employees through Arbeitnehmer online, and there was nothing wrong with that part. The problem sat next to it. The day after payroll, employees stood in the doorway asking why the night shift premium was lower than the month before. A bank wanted a salary certificate, HR opened the last one and changed name and figures by hand. And on the 25th one of the two spent half a day on the Excel sheet with the hours from time tracking, which then went to the firm by email, with a note about which line still needed correcting.
The moment it tipped was a transfer error. A row in the Excel sheet had slipped, and one employee's night hours ended up next to the name of the person below him in the list. Nobody saw it before sending, the firm had no reason to see it, and so it landed in the payroll run. Two employees received wrong amounts, a correction the following month, a week of unrest on the shift floor. The head of HR told me later that the error itself was not the worst part; it was that she could not say which version of the sheet had actually gone to the firm.
At the first meeting the head of HR, the works council chair and, by video, the payroll clerk from the tax advisor's office sat at the table. The firm said it would gladly receive the hours as a file through one of the routes DATEV provides; that was less work for them than an email with addenda. The works council chair said payroll data was the most sensitive kind of personnel data for him, and he wanted to see what goes in and what comes out of every routine before it runs. Both were the foundation the project was built on.
Approach
Week one belonged to the stopwatch. Both people in HR noted every pay question, every certificate and every minute on the time tracking sheet for a week. The list was unambiguous: after the bonus payment the month before, more than thirty pay questions had come in, almost all on the same four topics. Certificates were derived from the last document five to eight times a month. And the monthly close of the hours cost half a day, plus the addenda. None of these three tasks needed a new payroll system or a change of firm. DATEV stayed where it was.
In week two the first routine began, the answers to recurring pay questions. The works council chair sat in the room, and we first set out on paper what the AI tool gets and what it does not. It gets the premium policy, the bonus policy and the works agreement on working time, documents every employee may read anyway. It gets the question without name, without personnel number, without amounts. It never gets the payslip. The answer cites the section in the policy, points to the payslip in Arbeitnehmer online for the individual amount and makes no commitment on the individual case. HR reads every draft before it goes out.
The second routine was certificates and employer letters. Here the AI tool has the smallest role, and that is deliberate. For every kind of letter, salary certificate for the bank, confirmation for the parental benefit office, there has been a template with fixed fields ever since. Name, period and start date come from the personnel record, amounts come only from the payslip, which HR reads itself. The AI tool phrases the cover note, nothing more. Before signing, HR checks four values against the source: name, period, amount, date. Since then no letter starts from the last similar document.
The snag came with the first routine, and it came early. The first answer template did exactly what it was not supposed to do: an employee asked, anonymised, how high the night premium would be for twelve night shifts, and the draft calculated an amount, neatly derived from the policy, but without knowing the payslip and therefore without any guarantee. HR caught it while reading; the draft did not go out. We then wrote the rule firmly into the process, not just into the prompt: no figures in the answer, every amount is a pointer to the payslip in Arbeitnehmer online, and every answer with a number in it goes back for revision. The works council chair took that rule into the agreement.
In parallel ran the third routine, the handover of time data, and in parallel ran the framework. With the firm we fixed the handover route DATEV provides, in this case an import file via the ASCII interface in the format the firm specified; the Lohnimportdatenservice through a marketplace partner stayed on the list as the next step. Before every handover the file goes through a checklist: missing days, duplicate entries, premium types, totals against the previous month. The AI tool only helps spot anomalies in the anonymised column structure; it does not see names or personnel numbers. The file gets a version and a date, the firm gets one version instead of three emails. Alongside ran the works agreement, the data processing agreement with the AI vendor, the review of the existing agreement with the firm, and alignment with the data protection officer. After three months the agreement was signed, and the three routines ran for one payroll cycle without me.
Outcome
A pay question now arrives by email, is anonymised, and HR sends an answer with the section in the policy and the pointer to the payslip in Arbeitnehmer online. No amount is calculated, no commitment is made. A certificate is created from the template and the personnel record, the four values are checked against the source, then comes the signature. At month end the hours go to the firm as one checked file with version and date, on the route DATEV provides for it.
The AI tool did not see a payslip at any step; that was the condition, and it held. The works agreement names the three routines, the documents the tool receives as context, the rule against figures in answers and the procedure for every new routine. At the review after the second payroll cycle the team had added two items to the checklist for the hours and asked the firm whether it wants to connect the Lohnimportdatenservice. That the two of them discussed that without me was the goal.