Goldman: AI trims insurance payrolls

Goldman Sachs thinks one of the first clear places where AI is cutting jobs is not Silicon Valley. It is inside insurance back offices, where claims files and underwriting packets have long moved from desk to desk, screen to screen, person to person. In a March 18, 2026 research note on the US labor market, Goldman said AI’s effect is already visible in “particular niches” of employment, and the report circulating this week singled out insurance claims and underwriting as places where payrolls appear to be shrinking by tens of thousands of roles rather than merely being “augmented” by software (goldmansachs.com, x.com). That detail lands because claims and underwriting are built from repetitive office work that computers can now do unusually well. A claim starts with intake: forms, photos, repair estimates, medical records, police reports, emails, call notes. Underwriting looks similar from the other side of the policy: submissions, loss runs, inspections, applications, financials, medical records, and broker correspondence. Much of the job is not the final judgment. It is gathering, reading, indexing, summarizing, and routing all that material so a human can make one (riskandinsurance.com, bcg.com). That is exactly the layer AI attacks first. Risk & Insurance reported in 2025 that carriers were already using large language models for document intake, optical character recognition, entity extraction, indexing, and summarization in both claims and underwriting. Tasks that once took hours of manual review can now be compressed to minutes, especially when the file is mostly paperwork rather than a hard judgment call (riskandinsurance.com). Boston Consulting Group described the same pattern in claims: generative AI is most immediately useful in the administrative parts of the workflow, where insurers expect faster cycle times and, in some cases, 20% to 30% productivity gains on claims-related admin work (bcg.com). So the jobs that get squeezed first are not usually the senior people deciding whether a complex injury claim is fraudulent or whether a messy commercial risk should be bound. They are the support roles around them: intake staff, processors, clerical reviewers, junior analysts, and others whose work consists of turning messy documents into a usable file. PwC wrote in January that underwriting, actuarial, and claims functions are shifting from manual decision-making toward “AI-assisted models,” and warned that automation can hollow out the routine work where newer employees once learned the business (pwc.com). That makes Goldman’s estimate more concrete. The firm is not arguing that AI has remade the whole labor market overnight. Its broader view is still gradual: in its base case, wide adoption unfolds over roughly a decade, with 6% to 7% of workers displaced during that transition and a larger effect if adoption is front-loaded (goldmansachs.com). Insurance matters because it offers a live example of what “front-loaded” looks like. The work is digital, rules-heavy, and document-heavy. The payoff from shaving minutes off every file is immediate. The labor data were already leaning this way before the AI boom became impossible to ignore. The Bureau of Labor Statistics projects employment of claims adjusters, appraisers, examiners, and investigators to fall 5% from 2024 to 2034, and insurance underwriters to fall 3% over the same period (bls.gov, bls.gov). Goldman’s point is that AI may be speeding up a decline that used to be explained more blandly as “automation” or “efficiency.” Insurers are not adopting these tools in a vacuum. State regulators have been building guardrails as AI moves deeper into pricing, underwriting, and claims. The NAIC adopted a model bulletin in December 2023 reminding insurers that AI-driven decisions still have to comply with existing insurance law, and Colorado expanded its own governance rules in August 2025 for insurers using algorithms and predictive models in several lines of business (content.naic.org, doi.colorado.gov). At the same time, consumer fights are already surfacing over how much software should shape coverage decisions (usatoday.com). Inside an insurer, though, the shift can look less dramatic than the headlines suggest. No robot walks in and replaces an underwriter. A document no longer waits in a queue for someone to name it, tag it, summarize it, and forward it. The file arrives half-built. The next person touches fewer pages. Then the next one does too. After enough of those small disappearances, Goldman sees a payroll line moving down. (goldmansachs.com, riskandinsurance.com)