AI Claims Cycle Times Drop 75% as STP Rates Near 70%: The Structural Shift in P&C Claims

Straight-through processing in claims sat between 10% and 15% of personal auto and simple property volume until 2024. By early 2026 leading auto carriers report 70% to 90% on basic personal auto, and average cycle time on those books has fallen from roughly 30 days to about 7.5.

The figure that qualifies the headline is Celent's split: 48% of global insurers run generative AI in production, but only 7% operate it at full scale.

Key Takeaways

• 70% to 90% STP on basic personal auto claims at leading carriers, against a 10% to 15% baseline two years ago, which moves the majority of claim count outside human handling.
• 30 days to 7.5 days average cycle time for carriers at production scale, against a J.D. Power industry average of 40.7 days from FNOL to final payment.
• 48% in production, 7% at full scale is the gap that decides whether the reserving consequence is measurable, because processing-method flags exist only where the workflow is integrated.
• 68% outsource AI while 18% monitor vendor risk, per AM Best, and CCC alone serves 27 of the top 30 auto insurers.
• 2.4 percentage points came off the US P&C long-term underwriting expense ratio, split 1.9 points of other acquisition expense and 0.5 points of general expense.

Where the Compression Actually Happens

The 75% reduction is not one improvement. It is three stages compressing by different amounts, and only one of them involves better decisions.

Claims Stage	Traditional Timeline	AI-Enabled Timeline	Reduction
FNOL to Initial Assessment	3-5 days	Minutes	~99%
Investigation and Decision	10-15 days	1-3 days	~80%
Settlement and Payment	5-10 days	Same day to 2 days	~75%
Total Average Cycle	~30 days	~7.5 days	75%

FNOL to initial assessment collapses because photo-based estimating generates line-item repair estimates from smartphone images at intake and language models handle unstructured voice or text descriptions. Travelers' AI Claim Assistant, built on OpenAI's Realtime API, completes the interview, policy verification and triage inside one call, and reported 50% or better STP on qualifying claims with 66% of customers opting into the AI channel.

Investigation compresses for a different reason: parallelization. Fraud scoring runs at FNOL rather than in week two, and document summarization, severity modeling and reserving run simultaneously rather than in sequence. Settlement compresses for a third reason that has nothing to do with AI at all, which is digital payment rails removing 3 to 7 days of mail float.

Scale is still modest at the vendor layer. CCC's Estimate-STP serves 40 insurer clients including seven of the top 10 by direct written premium, and handles roughly 5% of total claims volume across its network, with one large national carrier routing 20%.

The Triangles Split Before the Data Does

The reserving consequence is not that development accelerates. It is that a book now contains two populations with different development patterns, blended in one triangle.

When 70% of simple claims settle within 48 hours of FNOL rather than over 30 days, the reporting and settlement lag for that cohort nearly disappears, while bodily injury and litigated claims develop exactly as before. A chain-ladder selection fitted to the blend produces factors too high for the automated cohort and too low for the manual one. The dollars stay long-tailed even as the counts go short, which is precisely the mix that flatters early development factors without improving ultimate estimates.

Loss adjustment expense moves further and faster than losses do. An STP claim carries near-zero adjusting and other expense, so as automated share grows from 10% toward 70% the blended A&O ratio falls while defense and cost containment concentrates on the complex claims still handled by people. A static LAE factor carried forward through that mix shift overstates the expense provision, and the 2.4 point improvement already visible in the industry expense ratio suggests the shift is under way rather than prospective.

Case reserve behavior changes too. AI-generated estimates built from insurer-specific rules carry less adjuster-to-adjuster variability, which means fewer supplements and less strengthening. The historical relationship between case reserves and ultimate losses was estimated on a human-set population; on a book with growing automated share it is being applied outside the data that produced it.

The Segmentation Requires Data the Fragmentation Prevents

Every one of those adjustments needs the same input: a flag saying which claims were processed automatically. That flag is exactly what a fragmented stack does not produce.

Between 58% and 82% of insurers use AI tools somewhere, only 12% report fully mature capability, and 7% have achieved scalable success. The March 2026 reporting on that gap describes claims data as inconsistent, incomplete or siloed across systems, with different tools and vendors supporting different parts of the process. A carrier in that position can measure its aggregate cycle time and cannot cleanly separate the two cohorts inside its own triangle.

The concentration on the other side of the gap is the second constraint. CCC serves 27 of the top 30 auto insurers, so a model error or data quality issue there reaches most of the personal auto market at once, and AM Best found 68% of insurers outsource AI while 18% actively monitor vendor risk. The processing-method distinction that reserving now depends on is therefore held largely by third parties, under change management the ceding actuary does not observe.

Oversight capacity is the third. Sedgwick's 2026 research found 75% of claims professionals believe AI still needs human oversight, and a carrier handling a million claims a year at 70% STP still reviews 300,000 by hand while supervising 700,000 automated decisions. The NAIC's 12-state AI Systems Evaluation Tool pilot, running January through September 2026, will produce the first standardized read on whether that supervision exists in the form carriers describe.