MUNICH — One of Germany’s largest insurers and its works council signed an agreement on Monday that spells out, clause by clause, what an A.I. assistant may and may not do with the 2,400 people who process its claims, a document that labor lawyers say is the most detailed of its kind in the country.
The 11-page accord, negotiated over nine months, allows the company to deploy a system that drafts claim decisions, summarizes medical reports, and flags likely fraud. But it forbids the software from producing individual productivity scores, requires that every automated recommendation be explainable in plain German, and gives each employee the right to overrule the machine without justification. Overrides may not be counted in performance reviews.
“We did not want to stop the technology. We wanted to stop the surveillance that usually comes bundled with it,” said Birgit Lohmeier, chairwoman of the works council, who led the employee side of the talks. “The system can be very good at reading a report. It should not be reading the person who reads the report.”
German law requires companies to win the consent of elected works councils before introducing any technology capable of monitoring staff, a provision dating to 1972 that has become a battleground as A.I. tools arrive in offices. Unions have complained that many employers roll out assistants first and negotiate later. The Munich agreement reverses that order and, unusually, will be published in full so that other councils can copy it.
The company’s chief operating officer, Stefan Gruber, said a pilot in the motor claims unit had cut average processing time by 38 percent and reduced the backlog of files older than 30 days from 14,000 to under 3,000. “The economics are not in question,” he said in an interview. “What was in question was whether 2,400 people would trust the tool, and you cannot order trust by memo.”
Annika Weidenfeld, a professor of labor law at Ludwig Maximilian University, said the agreement went further than the workplace provisions of the E.U.’s A.I. Act, whose obligations for high-risk systems began phasing in this year. “The act tells you what an employer must document. This tells you what an employer may not do,” she said. “That is a different kind of rule, and it comes from the shop floor rather than Brussels.”
Not everyone on the floor is persuaded. Several claims handlers, speaking on condition that they not be named because they were not authorized to talk to reporters, said the ban on individual scoring would be hard to enforce if team-level statistics could still be broken down by shift. Ms. Lohmeier acknowledged the concern and said a joint committee would review query logs every quarter.
The agreement runs for three years. By then, Mr. Gruber said, the company expects to have retrained roughly a third of its claims staff for roles in which, as he put it, “the machine does the reading and the human does the deciding.” Ms. Lohmeier’s version was shorter: “We will see who is still here to decide.”
A version of this article appears in print on Oct. 12, 2026, Section A, Page 11 of the European edition with the headline: In Munich, an Insurer and Its Works Council Strike a Deal on How A.I. May Watch Employees.
