Fifty-nine North American P&C carriers answered WTW's Advanced Analytics and AI Survey, published March 19, 2026. Over the 2022 to 2024 measurement window, carriers at the leading edge of analytics adoption ran combined ratios six percentage points below slower adopters while growing premium three percentage points faster.

Both halves matter. The margin gap is risk selection and claims handling. The growth gap says the leaders were confident enough to write, not simply selective enough to survive.

Key Takeaways

  • Analytics leaders ran six points better on combined ratio and grew premium three points faster across 2022 to 2024, on responses from 59 carriers.
  • Underwriting analytics is close to saturated: roughly 80% use advanced rating and pricing models with 11% more planning them. The gap is on the claims side.
  • Fraud detection sits at 33% deployment and severity assessment at 29%, both projected to reach 65 to 70 percent within two years.
  • Only 20% of carriers report a defined analytics strategy and only 12% provide regular analytics training, which is what the 42% data quality complaint rate is downstream of.
  • Morgan Stanley projects a 200 basis point expense ratio gap by 2030, 30.5 against 28.5, with operating margins of 15.6% versus 17.4%.

Where the Six Points Come From

The gap is not in pricing analytics, because almost everyone has those.

WTW surveyed senior analytics, actuarial, and strategy decision-makers rather than technology practitioners, so the responses reflect resource allocation rather than tool inventory. Close to 80 percent of respondents rely on advanced rating and pricing models, with another 11 percent planning implementation. Adoption there is broad and mature, which is exactly why it cannot explain a six-point spread.

Claims is where the distribution splits. Fraud detection, the application most claims and actuarial leaders name as the clearest ROI case, sits at 33 percent current deployment. Severity assessment sits at 29 percent. Straight-through processing is at 14 percent with 36 percent planning within two years. Claim triage is at 25 percent, subrogation identification at 20, case reserving analytics at 20, and litigation probability modeling under 15.

Claims Analytics Application Current Deployment Projected 2-Year Adoption
Fraud detection 33% 65-70%
Severity assessment 29% 65-70%
Claim triage 25% Not specified
Case reserving 20% Not specified
Subrogation identification 20% Not specified
Straight-through processing 14% 50% (14% current + 36% planning)
Litigation probability modeling <15% Not specified

That is an industry that has instrumented the price and left the loss dollar largely alone. Underwriting analytics tighten selection and sharpen the rate. Fraud detection stops loss before payment, severity tools compress reserve volatility, and straight-through processing takes out cycle time and loss adjustment expense. The carriers running six points better did both halves.

The model architecture behind it is unglamorous. Gradient boosting machines remain dominant across the 59-carrier sample for both underwriting analytics and fraud detection, which means the leading cohort's advantage is built on interpretable, feature-importance-legible model classes that survive a state rate filing review rather than on opaque architectures.

What Claims Analytics Does to a Development Triangle

The reserving consequence arrives before the combined ratio does, and it arrives as a break in the data rather than as an improvement.

A carrier that adds fraud detection changes which claims reach payment at all. A carrier that adds severity assessment changes the distribution and the timing of settlement on the claims that do. Both alter the underlying claims process that a development triangle summarizes, so loss development factors calibrated on pre-deployment experience describe a process that no longer exists.

The problem is the detection lag. Standard triangle methods will not surface the shift until several accident years have developed far enough to show the new pattern, by which point the selection error is already in the carried reserve. The adjustment has to be made where the process changed, which means segmenting triangles by deployment period and by which application went live on which book, not waiting for the triangle to reveal it.

The scale of what is coming makes this immediate rather than theoretical. Fraud detection and severity assessment are both projected to move from roughly a third of carriers to 65 to 70 percent within two years. That is most of the market changing its claims process inside a single reserve review cycle.

A carrier writing $500 million of commercial auto at a 72 loss ratio, adding severity indicators at FNOL and revising reserve estimates on them, does not close a six-point gap in a year. It does change its development trajectory measurably inside two to three accident years, and the triangle it hands its reserving actuary stops being comparable to last year's.

Human underwriting augmentation is on the same path with a longer fuse: sixteen percent deployment now, 60 percent planning to prioritize it by 2028. GenAI is further along in stated adoption, above 50 percent implementing with another 29 percent planning, but concentrated in document processing and customer communication rather than in the pricing and reserving decisions that move a combined ratio.

The Benchmark the Laggard Prices Against Is the Trap

The widening spread damages the reference points that carriers on the wrong side of it use to check themselves.

An industry benchmark loss ratio averages across carriers whose risk selection now differs materially. A carrier running 65 through analytics-driven underwriting is writing a different distribution of risks than a carrier running 72 by traditional methods in the same line and the same state. Pricing the second carrier to rate adequacy against the industry average overstates adequacy, because its book contains risks the leader has already declined or priced away.

That error compounds on renewal. The laggard prices to the benchmark, attracts the risks the leader rejected, and finishes with a book worse than the benchmark rather than average to it. The rate indication is not wrong arithmetically. Its comparison set is.

The barriers data suggests the spread does not close on its own. Forty-two percent of carriers cite data quality and limited data accessibility, and inadequate IT support runs at a similar share, both descriptions of missing governed data pipelines rather than missing tools. Only 20 percent have a defined analytics strategy and only 12 percent train employees regularly, which is the constraint underneath the low claims deployment numbers. An adjuster who does not understand why a severity model recommends a settlement path overrides it on the high-frequency decisions where the model has the most statistical power.

That is what makes the projected doubling of claims analytics adoption within two years a forecast rather than a plan. Morgan Stanley's 200 basis point 2030 expense differential, 30.5 against 28.5 with operating margins at 15.6 and 17.4 percent, assumes the second half of the market executes a deployment that all but 12 percent of it has no training program behind.

AI in Insurance P&C Insurance Actuarial Technology Insurance Industry

Further Reading on actuary.info

Sources