will AI replace insurance underwriters?

Six of the seven tasks in your job have zero AI penetration. That's not a rounding error. That's the structure of the work. Examining an applicant's full financial picture, health history, and property condition to form a risk judgment requires you to weigh factors that don't come with clean labels. A flood-prone property with a recently renovated foundation, owned by a business with thin margins but 20 years of clean claims history, isn't a spreadsheet problem. It's a judgment call.

The correspondence tasks are the same way. Writing to field representatives, medical personnel, or legal contacts to get the information you actually need requires you to know what's missing, why it matters, and how to ask for it without triggering a refusal. AI can draft a letter. It can't decide which question to ask or read the subtext in a claims adjuster's terse reply.

The high-stakes decisions, authorising reinsurance when risk is elevated, applying rating adjustments for substandard risks, evaluating catastrophe exposure across a portfolio, all require accountability. Someone has to sign off. Someone has to stand behind the number. When a hurricane wipes out a coastal portfolio, the insurer doesn't go back to the algorithm. They go back to the underwriter. That professional accountability is yours, and no tool on the market today comes close to taking it.

The one task AI handles well in underwriting is declining obvious excessive risks, essentially flagging applications that fail basic thresholds. But even there, the failure modes are significant. AI models trained on historical claims data inherit the biases in that data. A neighbourhood that was redlined 40 years ago still shows up as high-risk in the numbers. The model declines. The underwriter never looks. That's a fair housing liability waiting to happen.

AI also can't read documents the way underwriters do. It can extract fields from a standardised form. It struggles with a 60-page property inspection report where the important detail is buried in a surveyor's footnote, or a physician's letter that technically clears the applicant but uses hedging language that an experienced underwriter would flag. Context like that requires professional experience, not pattern matching.

Reinsurance decisions are a particular problem for AI. Authorising reinsurance involves negotiation, relationship history with specific carriers, and judgments about how much risk your book can absorb across correlated scenarios. The Anthropic Economic Index places tasks like this at near-zero automation penetration precisely because they're judgments made under uncertainty, with real financial consequences, by someone who's accountable for the outcome.

The one place AI genuinely works in underwriting today is automated risk scoring at the intake stage. Tools like Zelros and Shift Technology connect to external data sources, credit bureaus, property databases, claims history, and generate a preliminary risk score before a human underwriter touches the file. For straightforward personal lines, auto or standard homeowners, this speeds up the initial triage and means a low-complexity file doesn't sit in a queue for a week.

Document processing tools like FRISS and Guidewire's AI features can scan incoming applications and flag missing information, extract key fields, and match data against your company's internal records. That cuts the time spent manually cross-referencing a new application against existing policies to check for concentration risk on a single address or related-party exposure. It's not glamorous, but it removes a step that used to take 20 minutes per file.

For research-heavy decisions, some underwriters are using general AI tools like ChatGPT or Claude to pull together background on unfamiliar industries or property types before writing a commercial policy. If you're underwriting a cannabis dispensary for the first time, a quick AI-generated summary of the regulatory environment and common loss drivers gives you a starting framework. You'd still verify everything, but it beats starting from scratch. The honest picture here is that AI handles the front end of the process, the intake, the data gathering, the initial screening. The actual underwriting decision stays with you.

The biggest shift is at the start of the file. A file that used to arrive as a stack of documents you'd manually sort through now arrives pre-screened. The obviously incomplete applications get kicked back automatically. The basic risk flags are already annotated. You spend less time doing clerical cross-checks and more time on the files that actually need judgment.

What hasn't changed is the middle of the process. Calling a field rep to ask why the applicant didn't disclose that prior claim. Reading through a commercial property inspection and deciding whether the electrical system concern is a hard decline or a rating adjustment with an endorsement. Talking through a reinsurance arrangement with a broker who's pushing for terms you're not comfortable with. That's still 80% of the job, and it still takes the same amount of time it always did.

The end of the day looks different in one specific way. Less time spent writing routine correspondence from scratch. The documentation tools covered above generate a first draft of standard letters faster. But you're still editing, still deciding what the letter says, still responsible for what goes out. The volume of decisions you're expected to handle has gone up slightly. That's the real change. Not that the work disappeared, but that the expectation around throughput has shifted.

The BLS projects 2.6% job losses for insurance underwriters through 2034, which translates to a meaningful reduction in a workforce of 127,000 people. That's not a catastrophic collapse, but it's a real headwind. And the cause isn't AI replacing the judgment work. It's AI handling the intake and screening work that used to justify hiring an extra underwriter when volume went up. The easy files are getting processed with fewer hands.

The 8,200 annual openings sound reassuring until you account for what those represent. Most of those are replacement openings from retirement and attrition, not new positions. The number of seats isn't growing. That means you're competing for replacements in a field where the number of complex, high-value files is holding steady but the entry-level pipeline is narrowing.

The roles that are most at risk within underwriting are high-volume personal lines positions, standard auto, standard homeowners, where the files are predictable enough that automated scoring handles most of the decision. Commercial lines, specialty lines, surplus and excess, reinsurance, these areas are considerably more protected because the complexity of each file is genuinely high. According to O*NET task data, the concentration of judgment-intensive tasks in underwriting is among the highest of any financial services role. That's why the exposure score sits at a modest 8%, even as the overall job count shrinks.

The 8% AI exposure score for underwriters is likely to creep up over the next five years, but not dramatically. The limiting factor isn't AI capability in general. It's the specific nature of underwriting decisions. Authorising reinsurance, evaluating catastrophe exposure, adjusting a policy for a substandard risk, these involve correlated judgments across multiple inputs where the error cost is high and the accountability is legal and financial. For AI to meaningfully penetrate those tasks, you'd need models that can explain their decisions in regulatory-compliant terms and take on liability for bad outcomes. Neither of those things exists today.

The scenario where this job is genuinely threatened looks like this: AI builds a strong enough track record on complex commercial decisions over a decade that insurers start accepting model-generated risk assessments for regulatory purposes. That's a 10-year story at minimum, and it requires regulatory change, not just better models. The more immediate pressure over the next five years isn't replacement. It's consolidation. Fewer underwriters handling larger portfolios, supported by better tools. That's a workload story, not an extinction story.

The task list makes the strategy obvious. The six irreplaceable tasks all involve judgment, correspondence, and complex risk evaluation. Double down on commercial lines and specialty lines experience. These are the areas where file complexity is highest, automated scoring is least reliable, and the human decision carries the most weight. If you're currently in personal lines auto or standard homeowners, consider moving. Those books are where the volume-reduction pressure will show up first.

Build your ability to evaluate risks that don't fit standard models. Cannabis operations, emerging technology companies, climate-exposed coastal properties, anything that sits outside the historical data the automated tools were trained on is a place where your judgment is genuinely irreplaceable. An underwriter who's comfortable writing a bespoke policy for an unfamiliar industry is much harder to replace than one who processes 40 standard homeowners files a day.

The correspondence and relationship skills matter more than they used to. Field representatives, brokers, and medical consultants are your information network. Knowing how to get a complete picture of a risk from sources that don't send clean data is a skill AI can't replicate. Invest in those relationships. Also, get comfortable with the documentation and screening tools your company uses, not to become a power user, but to understand where the automated judgment ends and yours begins. That boundary is where your value sits, and knowing it precisely makes you better at the job.