AI Insights Jul 17 2026

Jul 17, 2026

By James W. Moore

Someone Is Getting Paid Right Now to Teach an AI How to Do Your Job

Mercor, one of Silicon Valley’s fastest-growing AI training data startups, is currently advertising for an “Insurance Expert” role. The posting asks experienced underwriters, actuaries, claims specialists, and agents to build realistic scenarios out of real insurance workflows, underwriting submissions, claims investigations, coverage disputes, rate filings, and regulatory exams, and then judge how well an AI model performs against them. The pitch is straightforward. Bring a career’s worth of judgment, hand it over piece by piece, get paid by the task.

Mercor is not unusual in this. It has simply arrived in insurance. Handshake, Surge AI, Scale AI, and a cluster of smaller rivals have built fast-growing businesses on the same model, paying skilled professionals to evaluate and correct AI outputs in their own field, then selling 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 was recently in talks with investors over a deal that would value it at twice its October valuation of $10 billion. Handshake 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. One contractor brought in to grade an AI’s Portuguese language writing watched her own corrections become unnecessary within weeks as the model absorbed everything she had 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. The Guardian separately interviewed contractors rating outputs for a major AI product who described being pulled into moderating disturbing content with no warning and a review window that shrank from 30 minutes to 15 for 500-word responses.

Danielle Li, a management professor at MIT Sloan, put the underlying economics plainly 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, 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 is not compensated for the value it keeps generating after the contract ends. Not everyone agrees the demand will last. 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.

Why This Matters for Insurance:

The industry does not need to speculate about whether this touches insurance specifically. It already has. 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, and separate research citing Harvard Business Review analysis from Evercore ISI and Visionary Future ranks insurance and financial services among the occupations most exposed to large language models, precisely because the work is language-heavy, precise, and repetitive. That is the exact profile these training platforms are built to mine. 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 tool an employer may eventually use to justify not hiring their replacement. For carriers and agencies, the more useful question is not whether your people should take these gigs. It is whether you know if they already have, and what proprietary reasoning may have left the building in the process.

Sources:

An FT Report Backed by Anthropic and OpenAI Researchers Says Insurers Don’t Know What They’re Already Covering

A Financial Times report published July 14 covers findings from the Artificial Intelligence Underwriting Company, a specialty insurer, co-authored with researchers from Anthropic and OpenAI. The core finding is stark. More than 90% of insurers’ exposure to AI currently sits inside “silent” cover embedded in conventional policies rather than in any AI-specific product, which means the risk is largely unpriced and, in many cases, unnoticed by the carriers holding it. The researchers point to a shift already underway in litigation, from claims about chatbots generating bogus information toward claims involving AI agents that take action on a company’s behalf, exposure that can reach into professional negligence and wrongful death territory depending on what the agent was empowered to do.

Not everyone in the industry accepts the framing at face value. Some figures quoted in the coverage argue the warnings are overblown and amount to a marketing push for new AI-specific insurance products, a fair challenge given that AIUC sells exactly that kind of coverage. AIUC co-founder Rajiv Dattani countered that the legal uncertainty itself, not the marketing, is what has held back broader enterprise AI adoption at larger companies. Whatever the motive behind the report, the underlying claim, that AI risk is accumulating inside general liability, E&O, and tech E&O books without anyone having separately identified or priced it, is a different and more concrete assertion than the general “AI is risky” warnings the industry has grown used to tuning out.

Why This Matters for Insurance:

This is the thesis this newsletter and the site have been building toward for months, now showing up with an institutional co-author list attached. If AI agent exposure is already sitting inside existing books unpriced, that is not a future underwriting problem to solve before the next renewal cycle. It is a current reserving and portfolio question, and one that will not announce itself the way a new peril usually does, with a filing requirement or a rating bureau bulletin. It will announce itself with a claim nobody expected to be covered. Carriers that wait for a standalone AI liability product to formalize the category before taking inventory of what they are already carrying are choosing to find out the hard way.

Sources:

Every Data Center Is Also an Insurance Policy. Swiss Re Just Put a Number on It.

Swiss Re Institute’s July 8 sigma report, “World Insurance in 2026: Shock Absorbers in a Fragmenting World,” estimates that hyperscaler capital expenditure on AI infrastructure will reach $750 billion in 2026, contributing roughly 0.2 to 0.3 percentage points to US GDP growth even as broader global premium growth softens to 0.6% for non-life business this year. The report frames this as a genuine offset. While supply chain shocks and geopolitical fragmentation are dragging on the wider economy, the buildout of AI data centers, energy systems, and semiconductor manufacturing is creating entirely new pools of insurable risk across property, engineering, cyber, liability, and business interruption lines.

The scale involved is what makes this more than a routine commercial property story. New AI data centers can cost more than $20 billion to build before equipment goes in, and lenders are increasingly requiring insurance limits sized to the full construction cost even though maximum probable loss scenarios run far lower. Swiss Re estimates aggregate demand from this buildout could reach $90 billion in premium by 2030, with capital and capacity needs in the hundreds of billions behind that. The re/insurance industry, per the report, can currently support only a fraction of the limits these projects require at competitive rates through traditional construction risk policies, which is already forcing structural experimentation in how these placements get built.

Why This Matters for Insurance:

Read alongside this week’s FT report on unpriced silent AI exposure, Swiss Re’s numbers complete the picture from the other direction. AI is simultaneously creating a genuinely new, quantifiable, and growing category of insurable risk in the physical infrastructure sense, and quietly accumulating as unpriced exposure inside carriers’ existing books in the operational sense. Those are two different conversations happening on two different balance sheets, but they are the same underlying phenomenon: insurance absorbing a technology transition faster than its own pricing models can formally recognize it. The carriers building capability now to underwrite data center construction risk are, in a sense, further ahead on AI than the ones still treating agentic exposure as a someday problem. Both muscles need to be built at the same time.

Sources:

The Agent Isn’t Being Replaced. The Job Underneath the Agent Is.

Insurance Journal’s July 13 issue profiled three veteran account managers and CSRs on how AI has changed their day-to-day work, and the pattern across all three was consistent. AI now handles policy checking, premium comparisons, loss summaries, and initial proposal drafting, tasks that used to consume hours, while the professionals themselves have shifted toward verification and client relationship work. “I think there is still enough of a need for client touch that AI isn’t eliminating many positions,” said Michelle Salow, an executive account manager at Heffernan. Kasey Conners of the Big I’s Agents Council for Technology drew a sharper line: certificates, endorsements, coverage changes, renewal follow-ups, and policy reconciliation are all automatable, but the client advisor function is not going away, even as nobody in the industry has fully mapped where that boundary sits.

A separate piece from InsuranceNewsNet, written by two Troutman Pepper Locke attorneys, put numbers behind the regulatory side of that same shift. As of mid-2026, 25 states have adopted the NAIC Model Bulletin on the Use of AI Systems by Insurers, with roughly 29 states now carrying some form of AI insurance guidance once New York, California, Colorado, and Texas’s own frameworks are included. The NAIC’s 12-state pilot of its AI Systems Evaluation Tool, launched in March, runs through September and is expected to be formally adopted at the 2026 NAIC Fall National Meeting, giving regulators a structured framework for market conduct exams covering AI adoption breadth, governance structure, and high-risk system review. On the distribution side, the same piece cites research suggesting roughly half of US consumers will use AI tools to shop for insurance in 2026, and notes that some carriers are already revising agent commission agreements downward on AI-originated policies while adding new contract provisions restricting how agents may use third-party AI tools at all.

Why This Matters for Insurance:

The optimistic reading from the account managers interviewed, that client relationships are AI-proof, is probably right as far as it goes. But it sidesteps the harder question the regulatory data points toward: what happens to the pipeline that used to turn entry-level administrative work into senior client-facing expertise, when the administrative work is the part getting automated first. An industry veteran with 38 years of relationship capital is not replaceable by AI. A hypothetical version of that same person five years into their career, who never got to spend two years doing certificates and endorsements long enough to learn the coverage logic underneath them, is a different and much less certain proposition. The commission compression and AI-use contract restrictions showing up in agent agreements right now are the earliest visible symptoms of carriers and distributors renegotiating who captures the value AI creates, well before anyone has answered the talent pipeline question.

Sources:

From the AI World: IBM Just Had Its Worst Day in 115 Years, and Anthropic Was Part of the Story.

IBM shares fell roughly 25% on July 14, wiping out about $67 billion in market value in a single session, the steepest one-day decline in the company’s history, dating back further than the 1987 Black Monday crash. CEO Arvind Krishna issued a rare pre-earnings warning that second-quarter revenue of $17.2 billion would miss the $17.85 billion analyst consensus and offered an unusually blunt explanation in his letter to investors. “We did not adapt and move quickly enough, and numerous large deals failed to close on the timelines we expected,” he wrote. Krishna pointed to two forces hitting at once: enterprise customers redirecting capital toward AI datacenter hardware ahead of expected price increases on supply-constrained components, and a wave of deals frozen while clients reassessed cybersecurity exposure following Anthropic’s Mythos launch, whose vulnerability discovery capabilities reportedly prompted some customers to pause large software contracts pending review.

The sell-off spread quickly across the software sector, with Workday down 10%, ServiceNow down 8%, and Salesforce down 6.2% within hours, while cybersecurity stocks were the notable exception, rallying through the session as the same clients who froze other software spending kept paying for security. IBM has since committed $5 billion to Project Lightwell, an open-source security effort backed by major banks, and will deliver its full Q2 earnings report on July 22.

Why This Matters for Insurance:

Set the stock move aside. The mechanism behind it is the more interesting data point for carriers underwriting technology risk. A single frontier AI capability release triggered a real-dollar, real-time repricing of enterprise cybersecurity exposure across an entire software sector in the space of a trading session. That is a volatility pattern tech E&O and cyber underwriters have not historically had to model, capability-driven, correlated, and fast enough that traditional annual policy cycles cannot react to it as it happens. It is also a preview of the kind of correlated model exposure this newsletter flagged two weeks ago in connection with the FLARE-AI vulnerability reporting infrastructure. When the flaw or the fix is in the model layer rather than any single vendor’s code, an entire portfolio of insureds can move together, and carriers writing tech risk on annual terms are, in effect, underwriting a risk that now sometimes reprices in an afternoon.

Sources:

If You Found This Useful

This newsletter exists because insurance executives and independent agents need sourced, direct analysis on AI, not vendor marketing dressed up as news. If a colleague would benefit from receiving it, forward this issue or share the link.

And if you have not yet downloaded The Independent Insurance Agent’s AI Playbook, it is available for free at insuranceindustry.ai/ai-playbook. More than 33,000 words of practical guidance, no cost, no catch.

Sources

Subscribe to AI Insights Newsletter

AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.

Related Posts

  • AI Insights Jul 3 2026

    Salesforce ties its AI pricing to actual resolutions, not activity. AWS puts $1 billion behind…

  • AI Insights Jul 10 2026

    A new industry AI ranking crowns Allianz the leader, but the real finding is buried…

  • AI Insights Apr 17 2026

    Carriers are quietly rewriting coverage for AI exposures as new ISO exclusions take hold. Munich…

468