What AI can and cannot do in insurance underwriting

Where AI helps property underwriters, where machine learning struggles with insurance data, how to judge a model, and whether AI will replace underwriters.

Updated November 2026

Machine learningProbabilistic modelingHuman intuition1 to 6 claims per 100 policiesThe underwriting decision
In short

In property underwriting, AI mostly means machine learning models that score risks, rank submissions and fill in property data, and newer language models that read applications and documents. It sorts large volumes quickly and consistently. It is weak where claims are rare, outcomes take years to show and regulation limits the inputs, which is why it works best combined with probabilistic models and an underwriter’s judgment.

Where AI is used in underwriting

TaskWhat AI doesExample
Data enrichmentFills in property facts and hazard scores from an address, so the underwriter does not look them upMore than 300 data points returned for an address, a policy number or coordinates
Submission intakeReads applications, schedules and loss runs, and extracts the fields a system needsA statement of values turned into a list of geocoded locations
TriageRanks the queue by fit with appetite and by likelihood to bindOut-of-appetite submissions identified automatically, the rest worked in order
Risk scoringRates a location for a peril, or a risk for expected lossA wildfire risk score built from more than 30 variables
ProspectingFinds properties and segments that match the appetiteMillions of properties scanned for new business, ranked by predicted loss ratio and likelihood to bind

Most of the time saved comes from the first three rows. Underwriters spend up to 30% of their time on data collection, and more than half of submissions are beyond a carrier’s appetite. AI that removes the lookups and the out-of-appetite work gives that time back to the submissions that can be written. UrbanStat’s Risk Selection and Find Better Risks products cover enrichment, triage, risk scoring and prospecting, with AI-based risk scores and a “more likely to bind” score.

Where machine learning struggles with insurance data

  • Claims are rare. Most carriers see a claim frequency of about 1 to 6%, so a model learns from very few examples of what it is trying to predict.
  • Outcomes are incomplete. A policyholder who moves to another carrier leaves no record of whether they stayed claim-free.
  • Success takes years to judge. A risk scored high in year four that claims in year six was scored correctly, but a one-year validation window would count it as an error.
  • Regulation limits the inputs. Underwriting and rating variables must be justified, and some are not allowed at all.
  • Bias carries over. A model trained on past decisions learns the biases in them, so the inputs need the same scrutiny as the outputs.

The first point is an imbalanced dataset problem. With a claim frequency of 1 to 6%, a model that predicts no claim for every policy is 94 to 99% accurate and of no use. Insurance models are judged instead on how well they separate good risks from bad: loss ratio by score band, or precision and recall on the claims, tested on a later period than the one they were trained on.

The three pillars of risk analysis

UrbanStat’s answer to those limits is to combine three things, which it calls the Three Pillars of Risk Analysis:

  1. Machine learning, for patterns across many variables that a rule set would miss, such as a slope that raises wildfire risk in one region and not in another.
  2. Traditional probabilistic modeling, for rare and severe events where history is too short to learn from, such as a 1-in-100-year hurricane.
  3. Human intuition, for the context the data does not hold and for the final decision.

Human judgment has its own failure modes, such as anchoring on a previous score, which practices like blind scoring are designed to counter; see cognitive biases in underwriting. The reasoning behind the three pillars is set out in Is machine learning the silver bullet in underwriting?

How to judge an AI underwriting model

  • Blind, out-of-time testing: was the model scored before the outcomes were known, on a period it was not trained on? UrbanStat publishes its wildfire map before the season and compares it with what burned; see wildfire model performance.
  • A baseline: a model that rates everything high catches every loss and is useless. Ask what share of the book or the area it rates high.
  • Results in business terms: loss ratio and hit ratio by score band, not accuracy alone.
  • Explainability: can an underwriter see which factors drive a score, and can the carrier explain it to a regulator?
  • Inputs: does the model use personal or credit data the carrier would not be allowed, or would not want, to use? UrbanStat never uses personally identifiable information such as gender, age or ethnicity, nor credit scores or financial risk information.
  • Stability: does the score move for reasons that make sense, rather than with last season’s headlines?

What changes for underwriters

AI takes over the lookups, the data entry and the first sort of the queue. The underwriter’s time moves to referrals, accounts at the edge of appetite, broker relationships and the judgment calls models handle badly: unusual occupancies, incomplete information and risks with too little history to learn from.

That shift only works if the tools are built for underwriters. UrbanStat’s need no coding: an underwriter types an address, a policy number or coordinates and works from the Underwriting Desk, a single screen with risk scores, P&L, aggregation, claims analysis and moratoriums.

Common questions

Will AI replace underwriters?

Not on current evidence. AI is taking over data collection, data entry and the first sort of the submission queue, which is a large share of an underwriter’s day. Referrals, unusual risks and accounts with little history still need an underwriter, and regulators expect insurers to stay accountable for decisions their models make.

What is automated underwriting?

Underwriting in which rules and models accept, refer or decline straightforward risks without a person, usually small, standard risks that fit the appetite cleanly. Everything else is referred to an underwriter.

What is the difference between AI and rules-based underwriting?

Rules apply fixed thresholds an underwriter wrote, such as a roof past a set age. Models learn weights from data and can capture interactions that rules miss. Most carriers use both: models to score and rank, rules to enforce the underwriting guidelines.

How do you handle imbalanced data in insurance models?

Judge the model on loss ratio by score band, or precision and recall on the claims, rather than accuracy; weight or resample the rare claims during training; and validate on a later period than the model was trained on.

What data should AI underwriting models avoid?

Personal characteristics such as gender, age or ethnicity, and data that brings their bias in by another route. UrbanStat’s models also leave out credit scores and financial risk information.

Next step

See it on your own portfolio.

Book a 30-minute walkthrough tailored to your portfolio and perils of interest.

Request a Demo →