Most insurance underwriting still runs on self-reported data and paper forms. An applicant fills out a questionnaire, maybe attaches a photo, and an underwriter makes a judgment call. The process hasn't changed in decades. The data pipeline is thin, slow, and easy to game.
That's starting to break down. Automated data ingestion from satellite imagery, phone sensors, and internet-facing infrastructure now feeds risk models that price policies in seconds. The shift isn't theoretical. It's measurable in loss ratios, processing times, and customer acquisition costs.
Here's what's actually working.
Computer vision in property insurance
A single aerial image of a residential property contains more underwriting signal than a ten-page application form. Convolutional neural networks trained on labeled roof imagery can detect material type (asphalt shingle, tile, metal, flat membrane), estimate roof age within a two-year window, identify damage patterns from hail or wind, and flag structural features like skylights, solar panels, and swimming pools.
Nearmap captures high-resolution aerial imagery across the U.S. and Australia, refreshing coverage multiple times per year. Cape Analytics, acquired by Nearmap in 2023, built the ML layer on top. Their models classify roof condition on a four-point scale and estimate replacement cost from imagery alone. No physical inspection required.
The accuracy numbers are reasonable. Published validation studies show roof age estimates within two years of ground truth about 80% of the time. Material classification exceeds 90% accuracy on common types. That's not perfect, but it's better than asking a homeowner to guess when their roof was last replaced.
What matters for underwriters: these models run at quote time. A consumer requests a homeowners quote online, the system geocodes the address, pulls the most recent aerial capture, runs inference, and returns property attributes in under three seconds. The old workflow involved scheduling an inspector, waiting days, and manually entering data. The cost difference is roughly $150 per inspection eliminated.
Hazard proximity analysis adds another layer. CV models combined with geospatial data can measure distance to wildfire vegetation, flood zones, and neighboring structures. Insurers writing in California wildfire zones now use vegetation encroachment scoring derived from satellite imagery to adjust premiums or decline coverage entirely.
Telematics in auto insurance
Usage-based insurance (UBI) replaces demographic proxies with actual driving data. Instead of pricing a policy on age, zip code, and credit score, telematics captures what the driver actually does behind the wheel.
The data comes from two sources: OBD-II dongles plugged into the car's diagnostic port, or smartphone sensors. Phone-based telematics use accelerometers, gyroscopes, and GPS to measure hard braking events, rapid acceleration, cornering force, speed relative to posted limits, time-of-day patterns, and total mileage. OBD-II adds engine diagnostics but requires hardware.
Progressive's Snapshot program launched in 2011 and remains one of the largest UBI deployments. It started with a plug-in device and now offers an app-based option. Progressive reports that Snapshot participants who drive less and brake less save an average of $145 per year.
Root Insurance went further. Founded in 2015, Root underwrites primarily on telematics data collected during a test-drive period. The driver installs the app, drives normally for a few weeks, and Root's models score the driving behavior before issuing a quote. Root went public in 2020 and has had a rocky path since, but the underwriting model is directionally correct: price on behavior, not demographics.
The UBI market is growing at roughly 20% annually. By some estimates it will exceed $120 billion in gross written premium globally by 2030. The limiting factor isn't technology. It's consumer willingness to share data and regulatory frameworks that vary state by state.
The predictive ML models behind telematics scoring are gradient-boosted trees in most production systems. Neural networks haven't shown consistent lift over XGBoost-style models for tabular telematics features. The signal is in the features, not the model architecture.
Cyber risk scoring
Cyber insurance barely existed ten years ago. Now it's one of the fastest-growing lines, with U.S. direct written premium exceeding $7 billion in 2023. The underwriting challenge is unique: the risk surface changes weekly as software gets patched (or doesn't), employees click phishing links, and new vulnerabilities are disclosed.
Coalition built its underwriting model around continuous external scanning. Before issuing a policy, Coalition's system scans the applicant's internet-facing infrastructure: open ports, SSL certificate hygiene, email authentication (SPF, DKIM, DMARC), known vulnerable software versions, and presence of credentials in breach databases. This produces a risk score that drives pricing.
Coalition's gross written premium exceeded $520 million, making it one of the largest cyber-focused MGAs. The model works because cyber risk is observable from the outside. You don't need access to internal systems to see that a company is running an unpatched Exchange server or has an open RDP port on the public internet.
The feedback loop is tight. Coalition pushes security alerts to policyholders when new vulnerabilities are found in their stack. This active risk reduction is unusual in insurance. Most insurers don't try to prevent the thing they're insuring against. Coalition's claim is that active monitoring reduces claim frequency, which lets them price more aggressively.
Parametric insurance
Parametric policies pay out automatically when a measurable event crosses a predefined threshold. Wind speed exceeds 130 mph at a specific weather station. Earthquake magnitude exceeds 6.5 within 50 kilometers of the insured location. Rainfall drops below a defined level during a crop growing season.
No claims adjuster. No dispute over damage assessment. The trigger fires, and the payment transfers. The entire claims process collapses to a data lookup.
AI models sit in two places in this chain. First, they set the trigger thresholds and price the coverage. This requires fitting extreme value distributions (generalized Pareto, GEV) to historical weather and seismic data, then simulating payout distributions under various trigger configurations. Second, they validate the event data in real time, cross-referencing multiple sensor sources to prevent false triggers.
The crypto and DeFi world has adopted parametric structures for protocol risk. Risk Harbor built parametric coverage for smart contract exploits and stablecoin de-pegs. If a covered protocol loses funds or a stablecoin drops below a threshold price on a specified oracle, the payout triggers automatically via smart contract. It's a small but interesting application of the parametric model to a risk category that traditional insurers won't touch.
Company profiles
Lemonade launched in 2016 with a direct-to-consumer model for renters and homeowners insurance, later expanding to pet and life. The company runs on a claims bot (AI Jim) that handles first notice of loss and, in straightforward cases, approves payouts in seconds. Lemonade has reported processing claims in as little as three seconds, though that figure represents the best case, not the median.
Lemonade's revenue reached approximately $480 million in 2023. The company serves over two million customers across the U.S. and parts of Europe. Their underwriting model incorporates behavioral economics principles from co-founder Dan Ariely's academic work, though the specific ML techniques behind their risk selection remain proprietary. Loss ratios have improved but the company has not yet reached sustained profitability. The stock trades well below its 2021 IPO peak.
Next Insurance focuses on small and medium business coverage: general liability, professional liability, workers' comp, and commercial auto. The company has raised over $1.1 billion in total funding and covers more than 500,000 businesses. Their pitch is speed. A small business owner can get a quote and bind a policy online in under ten minutes. The underwriting models pull data from public business registrations, review sites, and industry classification codes to assess risk without lengthy applications.
Coalition is covered above in the cyber section. The combination of security monitoring and insurance underwriting in a single product is their differentiator. They've expanded into cyber and technology E&O coverage and now operate across the U.S. and Canada.
You can browse all insurance-focused companies in the AIFI insurance directory. For companies specifically using image analysis and visual data processing, see the computer vision AI type.
What the data shows
The pattern across all four areas is the same: replace self-reported, static data with continuously observed, machine-readable data. Aerial imagery instead of homeowner questionnaires. Phone sensors instead of demographic proxies. Internet scans instead of security questionnaires. Weather stations instead of damage assessments.
Each substitution makes underwriting faster, harder to game, and more granular. The models don't need to be perfect. They need to be better than the status quo, which in most insurance lines is a human reading a form and checking a box. That's a low bar. The companies succeeding here are the ones with proprietary data pipelines, not the ones with the fanciest model architectures.