The mechanism is mundane and expensive. Ahmad Jackson's employer put AI on initial loss reporting, expecting simple claims to resolve and complex ones to reach a person already organised [3]. What arrived instead, he says, was a wave of misclassified claims that adjusters rerouted to the right department by hand [4]. He moved to another carrier and told WIRED the tool was getting things wrong and adding work for adjusters [6]. An initial loss report is a record other parties act on, which is why its errors do not stay inside the company.
Automation is supposed to absorb the volume so staff handle only the hard cases. Lemonade's spokesperson Paul Staats put the promise plainly, saying automation lets employees spend their care and expertise on the most complex claims [12]. The company's own published numbers describe something different. By the end of last year its chatbot AI Jim was taking the first report 96 percent of the time, while automation resolved roughly 55 percent of all claims [11]. Every claim has a first report, so about 41 points of the book begins with a bot and finishes with a person [17]. Whether that handoff lands as clean intake or as cleanup is the exception rate, and no insurer publishes it.
The 98 percent figure comes with a caveat from Glassdoor's own economist. Chris Martin says reviews turn more anti-AI when workers sense layoffs coming [13], so part of that number is the job market rather than a verdict on any tool [1]. The market justifies the dread. BLS projected in 2024 that adjuster employment would fall by 18,900, or 5 percent, over the coming decade, naming technology as a major force [8]. A 5 percent decline of 18,900 implies a base near 378,000 jobs [15], which makes the 21 percent single-year drop WIRED cites from BLS data for May 2025 to May 2026 worth roughly 79,000 positions, about four times the whole decade's projection [9][16]. It is also not only junior staff complaining: Geoffrey Conrad, a claims executive in Mobile, Alabama, told WIRED there is an AI fatigue [7].
For anyone rolling this out on Monday, benchmark accuracy is the wrong screen. Two properties of the step you are automating decide who pays for its errors. First, whether the output leaves the building, meaning a claimant, an attorney or a regulator relies on it. Second, whether the person who catches the error can reverse it without escalating. When the output leaves and the catcher has no authority, the work moves to the least powerful person in the chain, who then explains it to a lawyer. State Farm's position that claims need a mix of human and digital expertise [14] reads less like caution than like an accounting of where exceptions land. The instrumentation that would settle this is dull and cheap: exceptions per hundred automated intakes, and minutes of rerouting per exception. Skip that measurement, and Glassdoor reviews become the record of what happened instead.
