The evidence: sources, scope, and corrections

Larger accounts give the underwriter more judgment work and less admin

Non-core and administrative work falls from 47% of underwriter time on books averaging under $10,000 in premium to 35% above $250,000, while risk analysis and pricing climbs from 22% to 33%. The separating variable is account size rather than the specialty label: specialty and commercial respondents report near-identical splits. That suggests a use case worth testing on large accounts, but the survey does not measure the value of AI or multi-agent systems.

Until August 2026 this page and the insurance and specialty page claimed specialty underwriters spend far more time per account than commercial ones. The Accenture survey below does not support that comparison: the specialty and commercial splits are within a point of each other. It supports the account-size comparison instead, and no public source we could find gives absolute hours per account. Both pages were corrected.

"SME underwriting is characterized by maximum automation and speed, with algorithms handling the bulk of simple risks and humans intervening regarding exceptions. Large corporate underwriting sits at the opposite end, marked by bespoke, slow and expert-driven processes to capture unique risks."

"However, even today 30–40% of underwriter time is often administrative, creating significant automation opportunities."

"We chose this line of business because it involves a considerable amount of manual data extraction and analysis ... AI technology, deployed in the right way, has the potential to remove manual tasks from our specialist underwriters, freeing them up to focus on more complex risks where human expertise and analysis are a must."

"At Arch, the intake-to-decisioning cycle time was running two to three days. That stretched quote turnaround times to three to four days."

"P&C underwriters spend on average ~40% of their time on non-core and administrative tasks, while risk analysis/pricing, and negotiation and sales support take ~30% each."

"Automation is lower the larger the account, and is lower in specialty lines and reinsurance."

Capco, "Bionic underwriting" (Jun 2026) · link · Hiscox and Google Cloud lead-underwriting announcement (Dec 2023) · link · Ivans, Arch submission-intake case (May 2026) · link · Accenture, 2021 P&C Underwriting Survey, slides 13–15 · link

The account-size and line-of-business percentages are read from the survey's two stacked-bar charts on slide 15 ("What percentage of your time do you spend each month on the following activities?"). Series were assigned by reconciling each bar against the overall averages on slide 14 (30% risk analysis/pricing, 31% negotiation and sales, 39% non-core), weighted by the published bases: total 434, of which personal lines 76, commercial 295, specialty 50, reinsurance 13; by average account premium, <$10K 111, $10–49K 104, $50–249K 128, +$250K 91. All three series reconcile to within 0.5 points. Fielded by computer-assisted web interviewing among members of The Institutes plus a Risk & Insurance sample.

AIG: Lexington submission volume and a stated ambition

AIG’s 2025 annual report states that Lexington processed more than 370,000 submissions in 2025, up 26% year over year, against an ambition of 500,000 by 2030. It separately reports a 35% increase in the submit-to-bind ratio for Lexington Middle Market Property.

Removed a bundled expense-ratio figure that referred to General Insurance overall. It did not measure Lexington’s AI impact. Submission growth alone does not isolate AI’s contribution or establish a cost saving.

AIG, 2025 annual report (company-reported results) · Read the annual report

Markel: 113% increase in GWP per full-time equivalent

Cytora’s 2023 Markel case reports a 113% increase in gross written premium per full-time equivalent (GWP/FTE). It also describes a strategic-partner quote service commitment of two hours, compared with 24 hours previously. These measures do not establish a general time-saving percentage.

Replaced an unqualified productivity label with the measured denominator and distinguished a service commitment from an observed average turnaround time.

Cytora, Markel case study (2023, vendor-reported) · Read the case study

Zurich: 5 countries in 90 days, 80% less manual processing

"Cytora submission intake across commercial lines; 5 countries in 90 days, 20+ markets in 16 months; 95%+ extraction accuracy; 80% less manual submission processing." (integration phases, competitor table; corrected in the Aug 2026 evidence pass)

"Zurich is scaling AI-powered risk digitization across 20+ markets within the first 16 months, reducing manual processing time by over 80% ..."

"Proven at scale across seven countries and two major lines of business, processing complex, multi-language submissions with 95%+ accuracy."

Cytora customer blog (May 2026) · link

Chubb: 150bps combined-ratio target, 85% process automation

"Targeting 150bps combined-ratio savings over 3–4 yrs; ~85% automation of major UW/claims processes." (integration phases, competitor table)

"Chubb told investors in April 2026 that nine to ten AI and digital transformation projects would deliver 150 basis points of run-rate combined ratio savings over three to four years (Chubb Q1 2026 earnings call, April 22, 2026) ..."

"The company targets 85% automation of its major underwriting and claims processes, expects roughly 70% of the organization to be touched by the transformation within three years ..."

actuary.info, "Chubb's 150-Basis-Point AI Savings Promise" (Jul 2026) · link

CFC: Lane Assist, agentic underwriting pilot

"'Lane Assist', billed by CFC as a world-first agentic underwriting pilot in specialty; email → quote recommendation in seconds for low-complexity cyber." (integration phases and timeline)

"...the launch of Lane Assist, a world first pilot of agentic underwriting in specialty insurance that takes a submission from email through to quote recommendation in seconds."

"Believed to be a world first in specialty insurance, Lane Assist is designed to increase the proportion of submissions handled on a low touch basis ..."

CFC press release (Apr 2026) · link

Hiscox: a terrorism-risk quoting proof of concept

Hiscox’s Q1 2024 statement reports that its 2023 proof of concept reduced the time to quote a terrorism risk from three days to three minutes. The statement describes further development; this figure is not a production average for all renewals.

The earlier secondary passage described this as production performance and explicitly denied it was a proof of concept. Replaced that passage with the carrier’s own account and aligned the related pages.

Hiscox, Q1 2024 trading statement · Read the primary statement

Agents can accelerate pricing analysis, code, and documentation

Agents can support end-to-end pricing and costing model development by extracting filing logic, profiling data, building independent challengers, generating governed rater code, testing, and maintaining documentation. Actuaries retain method, assumption, validation, rate, and release authority.

"Traditionally, this required actuaries to manually parse hundreds of rates over several weeks. Now, agentic artificial intelligence (AI) can autonomously reverse-engineer regulatory rate filings in a fraction of the time."

"The Actuarial Agent will support a range of targeted use cases designed to augment the expertise of Aviva's underwriters and brokers, such as enhancing pricing tools with new insights and optimising existing model code."

"Draft or update pricing models in natural language, and the agent generates schemas, rating logic, user interface components, and parameter tables with domain accuracy."

"This research project will examine how autonomous, goal-driven AI agents can transform traditional actuarial processes including data extraction, financial modeling, reserve analysis, pricing, valuation, regulatory compliance, and risk assessment."

Deloitte, "Agentic AI: Transforming pricing analysis in insurance" · link · Aviva/hyperexponential London Market pricing pilot · link · hyperexponential Actuarial Agent demonstration (vendor claims) · link · SOA Research Institute, agentic actuarial workflow scope · link

Sixfold: reported processing-time reductions of 50–97%

Sixfold reports customer processing-time reductions of 50–97%, hit-ratio increases of at least 15%, and GWP per underwriter growth of up to 30%, as reported by Reinsurance News. These are vendor-reported outcomes across deployments, not a guaranteed result for one workflow.

"...customers using its technology have recorded substantial operational improvements, including processing times reduced by between 50% and 97%, increases in hit ratios of at least 15% and growth in gross written premium per underwriter of up to 30%."

Reinsurance News on Sixfold's AI Underwriter launch (Jun 2026) · link

CLARA: average indemnity and attorney involvement

CLARA reports average indemnity of $77,807 for attorney-involved casualty claims versus $15,936 for unrepresented claims, about 4.9 times as much. This association does not establish savings caused by AI triage.

"Our research found that for casualty claims with attorney involvement, the average indemnity costs are 390% higher than for unrepresented claims ($77,807 vs. $15,936). Claim duration is 295% higher."

CLARA Analytics product page (vendor research) · link

Skills made procedures reusable; harnesses made long loops governable

Agent Skills separated reusable procedural knowledge from prompts and models. Long-running harness patterns then made goals, progress evidence, handoffs, tests, and stopping conditions explicit, improving consistency across agent loops.

"This led us to create Agent Skills: organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks."

"Instead of building fragmented, custom-designed agents for each use case, anyone can now specialize their agents with composable capabilities by capturing and sharing their procedural knowledge."

"However, compaction isn't sufficient. Out of the box, even a frontier coding model ... running ... in a loop across multiple context windows will fall short ... if it's only given a high-level prompt."

"The task often terminates upon completion, but it's also common to include stopping conditions (such as a maximum number of iterations) to maintain control."

Anthropic, "Equipping agents for the real world with Agent Skills" (Oct 16, 2025; open-standard update Dec 18) · link · "Effective harnesses for long-running agents" (Nov 26, 2025) · link · "Building effective agents" · link

METR: a benchmark trend at 50% task success

METR’s March 2025 study found that the human-task duration associated with 50% success on its software and reasoning tasks had doubled roughly every seven months over the preceding six years. This benchmark trend does not establish the reliability of an insurance workflow.

Added the benchmark scope, historical period, and success threshold; removed the implication that this supplies a dependable business-planning clock.

METR, March 19, 2025 · Read the study and its limitations

Correction: unsupported scenario probabilities removed

Earlier versions quoted 12% and 64% estimates for alternative AI futures from secondary commentary. We could not establish a primary basis for treating these as calibrated probabilities. They have been removed from the timeline’s planning guidance.

See the revised planning checkpoint

Ivans: real-time appetite becomes the #1 carrier-selection factor

"Agencies now rank real-time appetite information as the #1 factor in carrier selection (29% of respondents, up from 12% in 2024)." (integration phases, §2 implications)

"For the first time, real-time appetite information within the agency's preferred rating solution has emerged as the leading deciding factor agencies consider when choosing carrier and MGA partners — cited by 29% of respondents, jumping up from 12% in 2024."

Ivans, 2025 Agency-Carrier Connectivity Trends survey (Dec 2025) · link

AI 2027: a scenario with uncertain timing

AI 2027 presents a possible sequence of rapid progress. Its authors clarify that they do not know when AGI will be built and distinguish their modal year from longer median estimates. The cited page does not support our earlier assertion of updated 2028–2030 medians.

Removed the unsupported median range and replaced the timeline’s implied delivery schedule with questions for scenario planning.

AI 2027, including the authors’ November 2025 clarification · Read the scenario