Starting point
The company, around 280 employees, services with two sites, had done its homework. Personio had been live for two years, and the system sat: joiners, leavers, absences, master data, all maintained and current. The HR team consisted of three people, a head of HR and two HR officers, thinly staffed for a company that size. It was proud of its data quality, rightly so.
Only the writing work carried on as if Personio did not exist. Whoever needed a reference letter opened the last similar one in Word, swapped name and period and hoped the job description still fit. Candidate status emails were typed freely, by whoever had time, in the tone that person had that day. And the monthly report for management came together on Sunday evening: an export from Personio, pulled into Excel, three sentences of explanation underneath, sent shortly before midnight.
It tipped in a management meeting. The report showed high turnover in one department, the managing director asked for the reason, and nobody at the table could explain the figure, not even the head of HR who had sent it. The same week a candidate received two emails from two colleagues, one with an invitation, one with a rejection. The head of HR called me and said the system was clean, the problem was everything that comes after it. At the first meeting she, the data protection officer and the works council chair sat at the table. The chair had negotiated the Personio rollout and said straight away that the agreement from back then did not contain the word AI. That was the starting point.
Approach
Week one belonged to the stopwatch, as always. For one week all three noted which task started with a look into Personio and ended in Word or the mail client, how often it came and which fields it really needed. Three routines ended up at the top: the reference letter, the candidate status email and the monthly report. The same day we wrote down a principle that still sits above the agreement today: Personio is the source, the company's AI tool is never the storage. Which AI tool the company uses did not matter for the principle, and it stayed interchangeable throughout the project.
The first routine was the reference letter. Instead of copying the last document, HR exports from the Personio profile only what a letter needs: position, period, department, job description, manager. Added to that are the manager's bullet points. Both go into the house reference-letter prompt, which knows the grading levels and the wording the company uses. The draft comes back, HR checks every grade, verifies names and dates against Personio and signs. The works council chair sat next to us during the first run and wanted to see what the prompt receives and what it does not.
This is where the snag came. The first export we pulled from Personio contained salary fields, the date of the last pay adjustment and the sick-day counter, because the default template simply took everything in the profile. Nobody had wanted that, nobody had noticed, until the data protection officer opened the file. Together with him we cut the export template down to five fields and saved it in Personio as a dedicated template, so that nobody picks the default template again by accident. The incident was later written into the works agreement word for word, as an example of why export templates belong in the contract.
The second routine was the candidate status emails. For each status in Personio Recruiting the team collected three good emails from recent months, removed the names, and the company's AI tool built one template per status in the house tone. The templates have lived in Recruiting since then; when sending, HR adds one personal sentence. Two contradictory emails to one candidate are no longer possible, because the status in Personio determines the text and not the person who happens to be writing. The third routine was the monthly report: HR condenses the report in Personio to figures per department and month, without a single person in the export, and the AI explains the table in five sentences. HR checks every statement against the table and sends report and explanation together. The turnover in that one department, by the way, had a simple reason that was in the export: three fixed-term contracts had ended in the same month.
The works agreement ran in parallel. The existing agreement on Personio stayed in place and received an addendum: the three routines, the export templates with their fields, the review step by HR before anything is sent, the works council's control rights, and that every new routine is shown to the works council first. Alongside that we tightened roles and permissions in Personio, two accounts from the rollout period could still access areas nobody needed any more. The works council met every three weeks, the data protection officer reviewed the data processing agreement with the AI vendor. After three months the addendum was signed, the three routines were running, and the team ran them without me.
Outcome
A reference letter now starts with an export of five fields and the manager's bullet points, not with the last document. A candidate receives the email that belongs to their status in Personio, no matter who sends it. The monthly report goes to management with an explanation HR has checked against the figures, and the questions in the meeting are about the matter rather than the table. Sunday evening belongs to the head of HR again.
Personio has remained the only source, no record has changed where it lives, the company's AI tool has kept nothing. The works agreement names the three routines, the export fields, the review step and the procedure for new routines. At the six-month review the team proposed a fourth routine, the reminder for the probation review based on the start date in Personio, and agreed it with the works council on its own. That it needed nobody from outside for that was the goal.