What problem does it solve?
Once a submission is in appetite, the real work starts. A middle market or specialty underwriter reads hundreds of pages of operations descriptions, loss runs and supplementals, searches the web and internal systems for anything the broker did not mention, checks the risk against underwriting guidelines and writes a narrative that justifies the decision to a referral authority, an auditor or a reinsurer.
Much of that time goes into finding and restating information rather than judging it, and the quality depends on how thorough each underwriter is on a busy day. Exposures buried in a website or a court filing are easy to miss, narratives vary in structure, and experienced underwriters spend hours on write ups instead of broker relationships and complex risks.
- A Cytora case study on Markel UK reports that, before the project, Markel's underwriters spent more than 30% of their time on low skill, low value tasks such as rekeying risk data into different systems and pulling third party data by hand.Markel uses Cytora and achieves +100% productivity uplift to fuel growth (2023)
How does it work?
- Assemble the file. The copilot gathers the extracted submission, prior policies and claims, internal notes and approved external data for the insured into one view.
- Research the insured. An agent searches approved sources (company website, news, court and regulatory records) and summarizes what is relevant to the line of business, with links.
- Check against guidelines. Retrieval over the insurer's underwriting guidelines and referral rules flags where the risk falls outside authority, where information is missing and which questions to ask the broker.
- Draft the narrative. The copilot writes a first draft of the underwriting narrative or referral note in the insurer's format, citing the document and page behind each statement.
- Underwriter decides. The underwriter edits the draft, sets terms and price, and signs the decision; edits and outcomes are logged to improve prompts, retrieval and guidelines.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cycle time reduction | Too few to pool | 50% | 1 | 1 vendor |
| Employee adoption | Too few to pool | at least 90% | 1 | 1 vendor |
| Time saved per task | Too few to pool | 120 minutes | 1 | 1 vendor |
Value drivers: Employee productivity, Risk and loss reduction, Speed and cycle time, Compliance quality.
Indicative value
A commercial insurer with 100 underwriters in middle market and specialty lines
USD 525,000 to USD 4.1 million
Underwriter time released per year
How this is calculated
Formula: underwriters * filesPerUnderwriter * hoursSaved * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Underwriters using the copilot underwriters, underwriters | 100 | 100 | The reference insurer. |
| Submissions fully assessed per underwriter per year filesPerUnderwriter, submissions per underwriter per year | 150 | 250 | Editorial assumption for middle market and specialty lines. Replace with your own volumes. |
| Hours saved per assessed submission hoursSaved, hours per submission | 0.5 | 1.5 | Conservative against the benchmark on this page (Zurich North America underwriters report an average of 2 hours saved per submission). |
| Fully loaded cost of an underwriter hour costPerHour, USD per hour | 70 | 110 | Editorial assumption. Replace with your own fully loaded cost. |
What it leaves out: Time released, not cash saved. It leaves out the value of better risk selection (exposures found that change a decision), more quotes per underwriter, platform and data costs, and the time underwriters spend checking drafts.
Who already uses it?
8 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Accelerant Holdings
Global · Insurance · 2025
Accelerant runs a risk exchange that connects specialty MGAs (its Members) with risk capital. Its 2025 annual report says incoming data, from Member bordereaux to third party sources, is validated, transformed and governed using AI agents, that internally developed AI tools and models assist Members' underwriting, and that its risk evaluation tools help Members identify, classify, validate, research and price underwriting opportunities. Members also get AI supported claims insights, actuarial analysis and portfolio management to manage rate adequacy. Engineers, data scientists, product managers and designers made up 34% of its workforce at the end of 2025. No outcome figures are disclosed.
No outcome disclosed.
American International Group
United States · Insurance · 2025
At its March 2025 investor day, AIG presented AIG Underwriter Assistance, a generative AI solution in production in Financial Lines that extracts data from broker and agent submissions, augments it with AIG and approved third party data, summarizes each submission and ranks submissions by appetite and propensity to bind, after which the underwriter analyzes the output and quotes. AIG describes the prior submission to quote process as taking about three to four weeks, with underwriters unable to review every submission, and says the assistant prepares submissions for review within one day. AIG frames the build around a human in the loop principle and was extending the same components to claims.
- Cycle time: up to 1 days, submission to underwriter ready file, in production lines
"AIG Underwriter Assistance Synthesizes and Prepares Submissions for Underwriter Review Within One Day"
Claimed by: organization
Skyward Specialty Insurance Group
United States · Insurance · 2025
Skyward Specialty announced in December 2025 that its partnership with Sixfold was entering its second year. The platform preprocesses submissions and generates recommendations on prioritization, appetite alignment and risk summarization and assessment, while underwriters stay in the loop to apply their judgment. The platform is live across six business units and more than 10 product lines, with an average deployment timeline of 8 to 10 weeks. The company presents the partnership as a step toward fully AI powered underwriting across its US property and casualty lines; no outcome figures were disclosed.
No outcome disclosed.
Arch Capital Group
Bermuda · Insurance · 2024
Arch's 2024 annual report says it uses AI for catastrophe modelling and predictive analytics and, in its insurance operations, to provide more information about past experiences and submissions so that its professionals can make more data driven underwriting decisions. Every new generative AI technology proposed for use in its operations requires approval and is monitored closely. No outcome figures are disclosed.
No outcome disclosed.
Hiscox
United Kingdom · Insurance · 2024
Hiscox London Market combined its own Hiscox AI Laboratories (Hailo) with Google Cloud's Gemini model to automate lead underwriting from email submission to quote in its sabotage and terrorism line. In scope risks are assessed by the model and the process generates an email to the broker with pricing and other data completed, ready for underwriter review. After a December 2023 proof of concept, in which Hiscox said the manual extraction step can take up to three days and quotes could be produced within three minutes, the model went live in August 2024. It initially covers renewals of existing US and Canadian sabotage and terrorism risks, excluding the New York and Chicago metro areas.
No outcome disclosed.
Bowhead Specialty
United States · Insurance · 2025
Bowhead's casualty underwriters use Kalepa's AI underwriting platform to bring research, data and underwriting guidelines into one workspace, surface information missing from broker submissions and apply appetite consistently. Bowhead's head of casualty says the book profile has improved because the tool helps avoid heightened risk profiles that were not visible in the broker's submission; no figures are given. Bowhead's own 2025 annual report states that it does not currently use generative AI tools, so the platform should be read as data and analytics support rather than a generative assistant.
No outcome disclosed.
Generali Global Corporate & Commercial
Italy · Insurance · 2025
As its cyber book grew, Generali GC&C chose Sixfold as its first external AI partner and connected it to its cyber data sources and scoring system, so that all available risk information is structured in a dashboard for the underwriter against Generali's own guidelines. The vendor reports that over 90% of underwriters adopted the platform, that most cyber submissions now go through it with turnaround times for distribution cut by 50%, and that risk engineering reports take a few hours instead of about two days. Its Global Head of Operations and IT, Matthew Richardson, is quoted as saying that Sixfold's input is now required for every quote.
- Cycle time reduction: 50%, turnaround for distribution channels on cyber submissions
"Today, most Cyber submissions are accelerated through the Sixfold solution, cutting turnaround times for our distribution channels by 50%."
Claimed by: vendor - Employee adoption: at least 90%, cyber underwriters
"The results were immediate: over 90% of underwriters actively adopted the platform, reporting consistently high accuracy scores."
Claimed by: vendor
Zurich North America
United States · Insurance · 2024
Zurich North America's U.S. Middle Market underwriters had to comb through hundreds of pages of exposures, operations, loss runs and supplemental forms per submission and then write a compliant underwriting narrative. After Sixfold was selected as one of nine winners of the Zurich Innovation Championship, the team rolled out a tool that gives underwriters an AI generated first draft of the narrative in Zurich's appetite, format and tone, with accuracy tracking and feedback sessions with underwriters. The vendor reports an average of 2 hours saved per submission and expansion from four offices to dozens within six months; an underwriter describes the tool surfacing a major exposure that changed a risk assessment.
- Time saved per task: 120 minutes, per submission, as reported by underwriters
"Underwriters reported saving an average of 2 hours per submission"
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Current underwriting guidelines, referral rules and authority levels per line, in retrievable form
- Extracted submission data (from an intake process) and policy and claims history per insured
- A list of approved external sources and data vendors, with use conditions
- Examples of good underwriting narratives to set the format and tone
Systems to integrate
- Underwriting workbench or CRM
- Policy administration and claims systems (read only)
- Third party data providers and web research tools
- Document management for submissions and guidelines
Complexity: Medium
The models are capable; the hard parts are clean underwriting guidelines to retrieve from, approved data sources, a narrative format underwriters accept, and a way to measure whether the drafts are right. Adoption depends on underwriters trusting the output, so accuracy tracking matters as much as the build.
- 1
Choose a line where narratives are long and consistent
Middle market property and casualty, cyber and professional lines work well: the files are big, the narrative format is standard and underwriters feel the pain.
- 2
Clean the guidelines first
Put underwriting guidelines and referral rules into one current, owned source. The copilot can only be as right as the guidelines it retrieves.
- 3
Define what a good draft is
Agree a checklist with senior underwriters (exposures covered, guideline breaches flagged, sources cited) and score a sample of drafts against it before rollout.
- 4
Pilot in a few offices with feedback loops
Start with a small group, hold regular feedback sessions and publish accuracy results to the users. Zurich North America combined accuracy tracking with feedback sessions and, according to its vendor, expanded from four offices to dozens within six months.
- 5
Instrument adoption and edits
Track how often drafts are used, how much they are edited and whether flagged exposures change decisions; low edit rates on bad drafts are a warning sign, not a success.
Guardrails
- The underwriter signs every decision; the copilot never binds, prices or declines on its own
- Every statement in a draft links to its source document or web page
- Research limited to an approved list of external sources
- Guideline retrieval restricted to the current approved version, with owners and review dates
- No use of protected characteristics or proxies for them in any scoring the copilot surfaces
KPIs to instrument
- Time from assignment to underwriting decision, per line
- Share of drafts used and the average edit distance
- Exposures flagged by the copilot that changed the decision
- Accuracy of drafts on a scored monthly sample
- Weekly active underwriters as a share of licensed users
Human in the loop
Underwriters review, edit and own every narrative and decision. Underwriting management reviews a monthly sample of drafts against the quality checklist, and referral authorities see the draft and the underwriter's edits together.
Common failure modes
- Automation bias
- Underwriters accept a fluent draft without checking it. Show sources inline, sample drafts for quality and make the underwriter confirm key facts.
- Stale guidelines
- The copilot confidently applies a guideline that changed last quarter. Give each guideline an owner and a review date and retrieve only the current version.
- Research that drifts beyond approved sources
- Web research pulls in unreliable or personal information about individuals. Restrict sources and log every page used.
- Time saved but decisions unchanged
- If the copilot only speeds up writing, risk selection does not improve. Track flagged exposures and outcomes, not just minutes.
What are the risks and rules?
EU AI Act
Depends on design
For commercial property and casualty lines the copilot is not listed in Annex III. Used for risk assessment of natural persons in life or health insurance it falls under Annex III point 5(c) and is high risk, with risk management, data governance, logging and human oversight duties, and deployers must carry out a fundamental rights impact assessment under Article 27.
Guidance
- Opinion on Artificial Intelligence governance and risk management (European Insurance and Occupational Pensions Authority, Europe). Published in August 2025 and addressed to national supervisors, it sets out how insurance sector legislation applies to AI systems that are not prohibited or high risk under the AI Act, including governance, fairness, explainability and human oversight.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(c) makes life and health insurance risk assessment and pricing of natural persons high risk.
- SB21-169, Protecting Consumers from Unfair Discrimination in Insurance Practices (Colorado Division of Insurance, North America). Requires insurers to test external data, algorithms and predictive models for unfair discrimination; Regulation 10-1-1 sets governance requirements for life, private passenger auto and health benefit plan insurers.
Controls to put in place
- AI inventory entry per line of business with an accountable underwriting owner
- Documented approved source list and data use conditions
- Retention of drafts, edits and final narratives for audit
- Periodic bias and fairness testing where personal lines or individuals are in scope
- Model provider risk assessment and exit plan under DORA
Frequently asked questions
- How much time does an underwriting copilot save?
- Zurich North America's middle market underwriters report saving an average of 2 hours per submission with AI drafted narratives. For Generali GC&C, turnaround times for its distribution channels on cyber submissions were cut by 50%. Both figures come from the vendor's case studies, and the Zurich figure is self reported by underwriters rather than a time study.
- Does it replace underwriting judgment?
- No. The deployments on this page keep the underwriter as the decision maker: AIG shows the underwriter analyzing the AI output before quoting, Skyward Specialty keeps underwriters in the loop to apply their judgment, and at Zurich North America underwriters start from an AI first draft. The value is in reading more, missing less and writing faster.
- Is an underwriting copilot high risk under the EU AI Act?
- It depends on the line. Commercial lines are outside Annex III. Risk assessment of individuals for life or health insurance is high risk under point 5(c), which brings conformity, logging and human oversight obligations.
How to cite this page
Blits.ai AI Use Case Library, "AI copilot for underwriting risk assessment", last verified 26 September 2026, https://www.blits.ai/ai-use-cases/underwriting-risk-assessment-copilot. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published