Somewhere in the insurance industry right now, an underwriter is teaching a machine how to underwrite. An adjuster is teaching it how to settle a claim. An actuary is showing it how to project a loss curve. They are doing this for a contract rate, on their own time, for a company most of their colleagues have never heard of. And when the model gets good enough, the work disappears - which was rather the point of hiring them in the first place.
Mercor, one of Silicon Valley's fastest-growing AI training-data startups, is currently advertising for an "Insurance Expert" role. The posting is unambiguous about the goal: the company is partnering with AI labs to bring in experienced underwriters, actuaries, claims specialists, and agents to sharpen how AI systems handle risk assessment, policy evaluation, and claims processing. Contributors are asked to construct realistic scenarios out of real insurance workflows - underwriting submissions, claims investigations, coverage disputes, rate filings, regulatory exams - and then judge how well a model performs against them. In effect: bring your career's worth of judgment, hand it over piece by piece, get paid by the task.
Mercor isn't unusual in this - it's simply arrived in insurance. Handshake, Surge AI, Scale AI, and a cluster of smaller rivals have built fast-growing businesses on the same model: pay skilled professionals to evaluate and correct AI outputs in their own field, then sell the resulting data to the AI labs. The New York Times reported that Mercor alone pays more than $4 million a day to roughly 30,000 contractors, and that the company was recently in talks with investors over a deal that would value it at twice its October valuation of $10 billion. Handshake, which only pivoted into this business in 2025, told the Times its annualized revenue run rate crossed $1 billion in April, up from $550 million at the start of the year - a jump also confirmed by The Information.
The mechanics are mundane, which is part of what makes the trade unsettling. Reporting from The Wall Street Journal's "The Journal" podcast describes contractors being shown a prompt and a model-generated answer, then asked to judge whether it holds up the way a real expert's answer would - flagging factual slips, tells, or reasoning gaps an outsider would miss. For insurance, that might mean an underwriter grading an AI's read on a commercial submission, or an actuary checking a model's loss projections. Mercor's posting notes contributors won't be given access to any employer's or client's confidential files - a nod to how sensitive that underwriting and claims data actually is, even as the professional's own knowledge is being extracted just as thoroughly.
Insurance Business has spent the past year documenting how AI is hollowing out the profession from within, which makes the training-data trade look less like a side hustle and more like professionals actively supplying the ammunition. A recent analysis on this site, drawing on Harvard Business Review research from Evercore ISI and Visionary Future, found insurance and financial services rank among the occupations most exposed to large language models, precisely because the work is language-heavy, precise, and repetitive - the exact profile these training platforms are built to mine.
That exposure already shows up in the numbers. Insurance Business reported that a Q1 2026 Insurance Labor Market Study from The Jacobson Group and Aon found insurance job openings fell to their lowest monthly level in a decade by December 2025. The same article cited a Brookings Institution and NBER study - not insurance-specific, but flagged by Insurance Business as directly relevant given how much of the industry's administrative workforce fits the exposed profile - showing that workers with high AI exposure and few options for re-employment skew heavily toward the same kind of clerical and processing roles insurance has long relied on. Elsewhere on this site, industry leaders have pointed to the related cost: as AI absorbs the submission-entry and loss-run work that has traditionally trained entry-level underwriters, the industry's apprenticeship pipeline is thinning just as demand for judgment-heavy talent grows.
So the professionals signing up for a few hundred dollars a day to hand their reasoning over to a model aren't operating in a vacuum. They're doing it inside an industry that is already anxious about where its next generation of underwriters and adjusters is supposed to come from - and getting paid, however modestly, to make that problem worse.
"It seems like easy money. It really isn't."
Some of this is straightforwardly exploitative, and the reporting bears that out. The Guardian interviewed ten current and former contractors who rate outputs for Google's Gemini and AI Overviews products through the contractor GlobalLogic. One, a technical writer who took the job expecting ordinary content work, described being pulled into moderating violent and sexually explicit AI outputs with no warning and no mental health support. A separate contractor described her review window shrinking from 30 minutes to 15 for roughly 500-word responses, a pace that left her doubting the quality of her own ratings.
But plenty of contractors went in with their eyes open, chasing the money, and found the trade wasn't what they'd hoped either. One biology professor who picked up Mercor and Handshake gigs over a summer break, hoping for easy extra income, told the Times the work "went downhill pretty fast" - mandatory late-night meetings, vague criticism from managers, flat rates for tasks that dragged on for hours. A Mercor contractor interviewed by The Journal, who'd been brought in to grade an AI's Portuguese-language writing, watched her own corrections become unnecessary within weeks as the model absorbed everything she'd taught it; her rate was cut from $45 an hour to $35, then to a flat $20 per completed task, at which point she quit. Both were, in their own way, paid to accelerate the exact obsolescence they later complained about.
That's the trade being offered, whether or not the person taking it wants to see it clearly: sell your judgment by the hour, and you are not supplementing your income so much as liquidating the asset your career was built on - at a fraction of what it's actually worth.
Danielle Li, a management professor at MIT Sloan, made the underlying economics plain in the Financial Times: professional security has always rested on the scarcity of expertise, and generative AI erodes that scarcity by letting a company absorb one expert's judgment and redistribute it to every future hire in that role, anywhere. Applied to claims and underwriting, that means once a model has learned to replicate how an experienced adjuster reasons through a disputed claim, that judgment stops being scarce - and the adjuster who supplied it isn't compensated for the value it keeps generating after the contract ends. Li's advice to workers amounts to a caution against exactly this kind of short-term deal: think hard about how much reasoning you're handing over, insist on being paid specifically for training contributions rather than treating it as ordinary freelance work, and recognize that taking cheap, uncoordinated data-training gigs undercuts the bargaining position of everyone else who does your job, not just your own.
There's also a legal wrinkle specific to this arrangement: WSJ reporting on Mercor found the company had approached contractors - in one case, visual-effects artists - seeking prior work product the contractors didn't actually own the rights to, since it had been produced under studio NDAs. Mercor says it only licenses material contractors genuinely own and doesn't want anything tied to a current or former employer, but the episode is a reminder that insurance professionals asked to "build realistic scenarios" from their own case files should think carefully about whose information they're actually handing over.
The counterargument - that this demand won't last - has some real support. Anton Korinek, an economist now on leave from the University of Virginia to work at Anthropic, told the Times he expects the need for this kind of human training data to decline "somewhat" as models improve, and pushed back on the idea that most white-collar professionals will end up doing some version of this work. If he's right, the professionals cashing in now are trading their expertise for a payout with an unusually short shelf life - training their own replacement and then discovering there's no third act where the company needs them again.
None of that changes what's happening at the individual level. Every underwriter, adjuster, or actuary who takes one of these contracts is making an individually rational, collectively corrosive choice: pocketing a short-term rate in exchange for helping build the exact tool an employer will eventually use to justify not hiring their replacement. Insurance Business has already reported on how thin the industry's talent pipeline has become and how much weight now falls on judgment-heavy senior roles precisely because the entry-level ones are disappearing. It's worth asking, plainly, whether the professionals feeding that disappearance for a few hundred dollars a day are doing their industry - or themselves - any favors at all.
