AI underwriting tools are moving beyond deciding which risks to accept and starting to intervene on risks already bound, through sensor feeds, real-time alerts and pre-bind mitigation requirements. Carriers running mature pipelines report straight-through underwriting production rising from a 10% to 15% baseline to 70% to 90% (Vantage Point, 2026).

The harder question is whether that automation is lowering frequency or severity, or picking better risks faster. They are not the same claim, and no published metric yet separates them.

Key Takeaways

  • Straight-through production runs 70% to 90% at carriers with mature AI pipelines against a 10% to 15% baseline, with cycle times compressing from roughly three days to three minutes at Hiscox.
  • AIG's assistant reviews more than 500,000 E&S submissions, lifting data collection and accuracy from near 75% to upwards of 90%, with $4 billion of new business premium targeted by 2030.
  • Every published figure is an operational throughput metric. Not one is a loss ratio, a frequency count or a severity trend, which is what a prevention claim would have to move.
  • Full credibility for a frequency estimate needs roughly 1,000 to 1,500 claims, a count a young AI-underwritten commercial segment often has not reached in its first two accident years.
  • Sixfold reports hit-ratio gains of 15% or more across carriers representing $270 billion of gross written premium, a figure consistent with better targeting, genuine prevention, or both.

The Numbers on Offer Are Throughput, Not Loss Cost

The production figures are large enough to matter for loss-cost modeling whichever mechanism drives them. Sixfold's AI Underwriter, launched June 2026 across a customer base representing $270 billion of gross written premium including Zurich, Skyward Specialty, Generali Global Corporate & Commercial and New York Life, reports processing-time improvements of 50% to 97%, hit-ratio gains of 15% or more, and gross written premium per underwriter climbing by as much as 30% (The Insurer, June 2026).

Vantage Point puts straight-through processing on simple claims at 70% to 90% for mature-pipeline carriers against a 10% to 15% baseline. Decerto's claims-side benchmarking puts the industry starting point lower still: average straight-through processing below 10%, with nearly 60% of insurers reporting no meaningful capability and only the strongest personal-lines writers approaching 35%.

MetricPre-AI baselineAI-pipeline carriersSource
Straight-through underwriting production10-15%70-90%Vantage Point, 2026
Industry-wide claims STP<10%up to 35% (top personal lines)Decerto / Aite-Novarica, 2026
Underwriting cycle time~3 days~3 minutesVantage Point, citing Hiscox, 2026
Sixfold customer hit ratiobaseline+15% or moreThe Insurer, June 2026
AIG data accuracy in underwriting pilots~75%90%+Zaffino, CIO Dive, 2026

Note what the table does not contain. Not one figure is a loss ratio, a frequency count or a severity trend. These are throughput metrics, the numbers a vendor can produce inside a single underwriting cycle, and they say nothing yet about whether the risks moving through the pipelines are experiencing fewer or smaller losses.

AIG's assistant, built with Anthropic and Palantir, shows how blurred the line already is in production. It reviews more than 500,000 excess-and-surplus submissions, and CEO Peter Zaffino has credited pilots with lifting data collection and accuracy "from levels near 75% to upwards of 90%, while reducing processing time significantly" (CIO Dive, 2026), with $4 billion of new business premium targeted by 2030. Each figure describes selection speed and data quality. None describes a hazard fixed before it produced a claim.

Three Channels, and Only Two Can Move a Loss Cost

Prevention embedded in underwriting runs through three channels with different actuarial signatures. Pre-bind conditioning flags a hazard during submission review and requires remediation before coverage attaches, the sprinkler-credit logic automated across satellite imagery, permit records and inspection data.

In-force monitoring puts connected sensors on bound risks so alerts trigger a response before a loss crystallizes, documented in commercial property sensor deployment. Dynamic re-underwriting reassesses a risk mid-term and adjusts terms as new data arrives.

Only the first two plausibly change frequency or severity for a risk that would have been written anyway. The third is a faster, more granular selection process. Conflating the three is how a carrier credits a rate decrease for work that never touched the underlying hazard.

That leaves an actuary with a three-way confounding problem on any early loss-ratio improvement. The book may be improving because the system prevented losses on risks that would otherwise have claimed, because it excluded or repriced risks the old process wrote at an inadequate rate, or because a thin and not-yet-credible period ran favorably. Sixfold's 15% hit-ratio gain illustrates the ambiguity precisely: a rising hit ratio on business a carrier wants is consistent with better targeting, with genuine prevention, and with both.

Separating them takes a control comparison, and writing statistically identical risks with and without the intervention is not something a profit-maximizing underwriting operation sets out to do.

So the pricing decision has to be owned explicitly rather than absorbed by a rating algorithm. Limited-fluctuation credibility generally wants 1,000 to 1,500 claims for a full-credibility frequency estimate, which a young commercial segment rarely reaches in one or two accident years, and the earliest development periods carry the widest tail-factor uncertainty.

The defensible treatment is partial credibility on the observed experience blended against the pre-AI loss-cost assumption for the same segment, with a widened contingency load on the AI-underwritten portion. Crediting the effect at full weight after eighteen months of immature data is pricing off a hypothesis, which is the shape of the accelerated-underwriting episode in life insurance, where mortality slippage forced a revisit before seasoned experience existed.

The Reserving Side Is Where the Change Arrives Unflagged

The reserving problem is distinct and harder to see coming. Historical loss development factors are calibrated on the mix of business a carrier actually wrote in each historical accident year. If prevention and re-underwriting shift which risks get bound, which get pre-bind remediation and which get intervened on mid-term, the accident years flowing through the pipeline are not drawn from the population the historical pattern was built on.

The direction of the error depends on the shift. A book increasingly composed of monitored, actively managed risks may develop on a flatter, faster-settling pattern than the legacy triangle assumes. A book where the system is selecting away marginal risks may show a thinner, more homogeneous tail than a blended company factor implies.

Mix shift itself is familiar. Reserving actuaries have adjusted for new state entry, distribution changes and product redesign for decades. What is new is that the operational metrics carriers publish, straight-through rates, hit ratios and cycle times, do not map onto the segmentation variables a triangle would need to isolate the shift, so a carrier without a reserving-side flag separating AI-underwritten cohorts from legacy cohorts cannot see its largest underwriting change in a decade in its own data.

That gap sits downstream of the scrutiny the programs actually attract. The NAIC Model Bulletin, adopted by 23 states and the District of Columbia as of early 2026, requires carriers to document how they test, validate and monitor AI used in underwriting and rating (state adoption).

An examiner can ask what evidence beyond an immature loss ratio supports the loss-cost assumption in current rates. The evidence that would answer it, a matched comparison of similar risks written with and without a specific intervention tracked through three full accident years, is largely absent from what carriers have published, and rate filings reach that question years before reserve reviews do.

Further Reading

Sources