The management decision is which workflow deserves investment, what evidence a pilot must produce, and when to expand or stop. This roadmap offers a sequence for those decisions. Its schedules and budgets are planning proposals, not a forecast of your return.
The examples focus on specialty insurance and the preparation of complex submissions. The plan starts with triage and intake so that teams can measure results while preserving underwriting authority. For another industry, retain the gates and replace the use cases, economics, and regulatory requirements. The case itself is set out under Insurance & specialty.
- Adoption ≠ scale. 76% of US insurance executives surveyed reported GenAI use in at least one function (Deloitte, June 20241); a separate BCG study found 7% of insurers had scaled AI, including predictive AI3, and <20% of carriers are at scale in any single business domain (McKinsey/LIMRA 20254). McKinsey reported an association between AI leadership and 6.1x the total shareholder return of laggards over five years (McKinsey7); advanced-analytics carriers ran combined ratios ~6 points lower (WTW, 2022–248). Neither comparison isolates a return caused by AI adoption.
- Specialty offers concrete examples to evaluate. Published examples include AIG Lexington17, Kinsale21, Hiscox London Market23, CFC20. These show possible workflow designs; they do not establish an industry-wide lead or guaranteed return.
- Map the regulatory requirements first. The August 6, 2026 NAIC map lists 24 states + D.C. with adopted bulletins13; a 12-state pilot of the NAIC's AI examination tool runs through September 2026.14 Confirm local applicability and complete controls before the affected use begins.
- The proposed plan: four gated phases over ~36 months: Foundation & assistive AI (0–6), Core workflow integration (6–18), Agentic & portfolio-level (18–36), AI-native operations (36+). Budget ~2–3% of IT spend in year one, scaling to 10–15%; start with a proposed even split between technology and adoption, then validate it against the work.
1 · Where the industry actually is
Survey evidence
Widespread use does not mean scaled adoption
Two surveys describe different aspects of adoption. They use different samples and definitions; the 7% is not a subset of the 76%.
Share of respondents within each survey. Deloitte includes life/annuity and P&C insurers; its P&C result alone was 70%. BCG includes predictive and generative AI.13
These are reported associations, not estimates of the return caused by adopting AI. Sources: Deloitte (published 2025), BCG (published 2025), McKinsey, WTW.
| Metric | Value | Source |
|---|---|---|
| US insurance executives reporting GenAI use in ≥1 function | 76% overall; 70% P&C | Deloitte, June 2024 survey (published 2025)1 |
| Insurers successfully scaled AI | ~7% | BCG 20243 |
| Carriers at scale in any business domain | <20% | McKinsey/LIMRA 20254 |
| P&C insurers generating value at scale in core workflows | 38% | BCG 20265 |
| Insurers still in pilot / proof-of-concept stage | ~60–66% | Capgemini 20266; BCG3 |
| Insurers tracking no AI metrics at all | 42% | Capgemini 20266 |
| AI leaders vs. laggards, 5-yr total shareholder return | 6.1x | McKinsey 20257 |
| Combined-ratio gap, advanced-analytics users vs. laggards | ~6 pts lower | WTW 20268 |
| Share of H1 2026 insurtech funding going to AI startups | 95.2% | Gallagher Re data9 |
Alongside model quality, the cited studies identify data readiness (a barrier cited by 78% of respondents in a life-insurance study11), change management (~two-thirds of the challenge, per BCG38), and governance fragmentation (68% of insurers say controls exist but are fragmented; only 24% are fully confident in them, per Grant Thornton 202610). That is the orientation's rule at organizational scale: context in, quality out. Data readiness decides what a model can be handed, workflow redesign decides where it is handed over, and governance decides who answers for what comes back.
This plan moves from experimentation to a working deployment, then to wider integration where the evidence supports it. There is no required destination or universal timetable. The deployment ladder helps define the authority a system needs at each step.
2 · What competitors have already done
| Competitor | Move | Reported result | Reported |
|---|---|---|---|
| AIG (Lexington) | AIG Assist and AI-supported underwriting | 370,000+ Lexington submissions processed in 2025, up 26% year over year; Lexington Middle Market Property submit-to-bind ratio up 35%, as reported by AIG17 | 2025 results; annual report published 2026 |
| Markel | AI Centre of Enablement; Cytora risk flows; AI-underwritten casualty unit (Cortex, with Bain) | 113% GWP/FTE uplift; strategic-partner quote SLA 24h → 2h (Cytora case, 2023)18 | 2023 case study; 2026 announcements |
| Chubb | Publicly committed AI transformation | Targeting 150bps combined-ratio savings over 3–4 yrs; ~85% automation of major UW/claims processes19 | Apr 2026 |
| CFC | "Lane Assist", billed by CFC as a world-first agentic underwriting pilot in specialty | Email → quote recommendation in seconds for low-complexity cyber20 | Apr 2026 |
| Kinsale | Enterprise AI license for every employee; merged Analytics + Technology under one chief | Dozens of internal bots for UW/analytics productivity21 | 2025–26 |
| Zurich | Cytora submission intake across commercial lines; 5 countries in 90 days, 20+ markets in 16 months | 95%+ extraction accuracy; 80% less manual submission processing; triage path to 15 minutes22 | May 2026 |
| Hiscox (London Market) | Gemini-based quote automation | Terrorism-risk quote proof of concept: 3 days → ~3 minutes23 | 2023 proof of concept; reported 2024 |
| Tokio Marine HCC | Cytora partnership in cyber & professional lines | Intake/triage automation; risk judgment stays with underwriters24 | Dec 2025 |
| Allianz Commercial | hyperexponential pricing transformation | 13 pricing tools shipped in 13 weeks25 | 2026 |
| Ryan Specialty | AI submission processing; internal ChatGPT for all staff | Reports internal AI access for employees and the start of submission-intake and analysis automation in 2025; quantitative impact not established here26 | 2025–26 |
| Brokers: Amwins, CRC, WTW | REDY INTEL; Neuron placement platform; $625M AI plans | Quotes in minutes; AI-driven placement analytics27 | 2026 |
Broker responsiveness is a useful outcome to measure: the Ivans 2025 survey ranked real-time appetite information first among carrier-selection factors for 29% of respondents, up from 12% in 2024.28 That supports attention to turnaround; it does not prove that faster responses alone win profitable business.
The table combines vendor case studies, carrier statements, pilots, and targets. Treat each as evidence about its stated scope. Before procurement, request results on comparable files, a definition of the metric, and the cost of exceptions and human review.
This table is about outcomes: reported results and plans, with the date and status needed to interpret them. For where these deployments sit against model capability and regulatory pressure, and which of them mark turning points rather than announcements, read the adoption track of the timeline, which is where the chronology lives.
Figure
Reported speed-to-quote, before and after
Text comparisons, with no length encoding. Markel reports a service-level commitment; Hiscox reports a proof of concept. They are not comparable production averages.
Sources checked: Cytora's Markel case study (2023) and Hiscox's Q1 2024 statement.
3 · The four gated phases
Each phase has explicit exit gates: governance, data, and measurement criteria that must be met before advancing. Treat the timings as planning assumptions and the gates as conditions to demonstrate before expanding a deployment.
Proposed programme · months from start
Four phases, with a decision at each gate
Illustrative programme months, starting at zero. Phases 1–3 have proposed time windows; Phase 4 has no fixed end date. Bar lengths follow the month scale. Select a phase to read its detailed plan.
Gate 1 · month 6. Governance program adopted, inventory complete, two pilots on metric
Gate 2 · month 18. Outcomes-testing methodology documented before AI touches selection or pricing
Gate 3 · month 36. Disclosure, appeal with human review, and drift monitoring before automated decisions
Continues beyond the range shown; no fixed end date.
Target return windows Planning targets to validate in pilots, not measured forecasts
Regulatory obligations follow jurisdiction and calendar dates, independently of this programme clock. See Governance and the regulatory timeline when setting gates.
Phase 1Foundation & assistive AI
Months 0–6Give everyone safe, useful tools; build the governance spine; prove value on document work.
What gets built- Approved AI access for trained staff (reviewed training, retention, and data terms); acceptable-use policy; consumer tools banned for company data
- Written AIS governance program mapped to applicable state requirements, AI inventory incl. vendor-embedded AI, cross-functional governance group
- Assistive pilots (AI drafts, human decides): submission document summarization (loss runs, SOVs, financials), claims file summarization, RAG knowledge assistant over guidelines/appetite
- Data readiness assessment: document pipelines, core-system API posture, data-quality baseline
- Governance program written and adopted; inventory complete
- ≥2 pilots meeting pre-agreed quality and value thresholds, including human review and rework
- All-staff training wave 1 done
Start with bounded document work whose output a reviewer can check. Confidentiality, accuracy, and downstream impact still determine the controls; a summarization task is not automatically low risk.
Phase 2Core workflow integration
Months 6–18Integrate demonstrated capability into underwriting and claims workflows, using a vendor, an internal build, or a combination justified by the requirements.
What gets built- Submission intake & triage in production (vendor): extraction, clearance, appetite scoring, third-party enrichment: a candidate supported by reported carrier and vendor results (including faster intake and improved throughput2917)
- Claims triage & document intelligence: severity/litigation prediction at FNOL (average indemnity for attorney-involved casualty claims is about 4.9 times as much as for unrepresented claims in CLARA’s vendor dataset30), reserve recommendation support (CLARA reports 12.8x ROI for one workers’ compensation deployment of its full suite; this does not establish the return from reserve support alone31)
- Bordereaux ingestion/validation if delegated authority (85–94% time savings reported32)
- Pricing and costing model development: agents profile and reconcile data, fan out independent challengers, generate rater code and tests, reverse-engineer filings, and maintain documentation; actuaries retain assumptions, validation, rate, and release authority.484950 See the practical workflow
- Intake AI live for ≥1 business unit with measured turnaround / hit-ratio lift
- Outcomes-testing methodology documented for anything touching selection or pricing
- Model-risk framework (NIST AI RMF-aligned42) operational; hub-and-spoke model with business-unit owners
Buy vs. build: compare product fit, integration effort, data rights, support, exit costs, and evaluation results on the same cases. An external survey's success rate is not a forecast for this insurer. Review the existing core-system roadmap before funding overlapping capability.
Phase 3Agentic & portfolio-level AI
Months 18–36From assisting tasks to orchestrating workflows, under explicit human authority.
What gets built- Agentic workflows for bounded, low-complexity segments: email → clearance → enrichment → pricing indication → quote recommendation, with underwriter approval (the CFC Lane Assist pattern)
- Portfolio management AI: continuous monitoring, appetite steering, accumulation insight (BCG estimates potential +1–3% GPW growth and −1–2.5 pts combined ratio34)
- Claims leakage and subrogation pilots: test duplicate-payment checks and recovery referrals on reviewed historical files; measure confirmed recoveries, false positives, and operating cost
- Broker-facing speed: real-time appetite APIs, integration where brokers are building AI intake layers
- Agentic workflow live in ≥1 line with human-override logs and drift monitoring
- Bias and outcomes testing meets the documented requirements for each relevant jurisdiction and line
- A Finance-reviewed value case in ≥1 domain; report cash impact, capacity, and service metrics separately
Phase 4AI-native operations
36+ monthsThe operating-model redesign: processes built around AI execution, with humans on judgment, exceptions, relationships, and governance.
What it looks like- Workforce redesign; new roles (AI product owners, model risk officer); capacity shifted to growth
- Multi-agent "virtual coworker" underwriting for routine segments (the McKinsey trajectory)
- Chubb's 150bps combined-ratio target19 and BCG's 15–25% operating-cost reduction5 live at this phase
Reached by compounding Phases 1–3, not by a separate program.
Illustrative decision walkthrough
A gate can expand, revise, or stop a pilot
Follow a fictional loss-run summarization pilot through its review. The calendar sets a meeting date; the evidence determines the outcome.
- Set the standardAgree what success and unacceptable failure mean.
- Review the evidenceAssess quality, complete cost, and control performance together.
- Choose an outcomeExpand, revise, or stop, with a recorded reason.
- Control the next releaseKeep ownership, scope, and monitoring explicit.
1 · Set the standard
Before running the pilot, the business owner agrees the cases to test, required source links, acceptable handling time, spending limit, and escalation rules. A reviewer still checks every brief in this example.
Example acceptance rule: each material loss figure must match its source. A missing valuation date must be flagged, and the tool may not send a broker message or make an underwriting decision.
2 · Review the evidence
Example finding: drafts arrive faster, but reviewers repeatedly discover an unflagged missing date. The time saving does not resolve that quality gap.
3 · Choose an outcome
Example decision: revise. Add an explicit valuation-date check and retest on held-out cases before proposing wider use.
4 · Control the next release
Record the decision, accountable owner, permitted use, remaining limits, and next review trigger. Changes to the model, source material, or workflow should prompt the relevant checks again.
Expanding volume does not automatically expand authority. A larger summarization pilot can still require the same human approval for every consequential decision.
Illustrative application of the gates above. The harness controls support evaluation and permission limits; the pilot decision sheet supports the value assessment.
4 · The economics
- Proposed budget shape: ~2–3% of IT budget in year one, scaling to 10–15% by year three, as an illustrative funding envelope. Replace these ranges with a bottom-up budget for software, usage, data, integration, evaluation, and adoption. Context: carriers' IT spend averages ~4.5% of GWP44; two-thirds of insurance CEOs plan to allocate 10–20% of budget to AI (KPMG39).
- Proposed 1:1 allocation: use an even technology/adoption budget as a starting point, then adjust to the workflow. McKinsey describes change management as about half the effort, not a measured dollar-for-dollar budget requirement.7 Leaders invest in change management at ~3x the average rate (Capgemini6); the industry currently spends 72% on tech vs. 28% on adoption, an inversion to avoid.6
- Potential sources of value (reported outcomes and estimates, specialty-relevant43): underwriter capacity (30–40% of commercial underwriter time is admin, per McKinsey7; Cytora reports Markel +113% GWP per full-time equivalent in its 2023 case study18), speed (reported improvements in specific quote workflows1722), claims (20–30% LAE reduction potential, per BCG5; early litigation triage; leakage prevention35), and portfolio steering (1–2.5 pts combined ratio, per BCG34).
- Planning targets, to validate: 6–18 months to first production ROI; 18–36 months to enterprise-level P&L impact. 67% of insurance CEOs now expect returns in 1–3 years (KPMG 202539).
Planning assumptions and survey context
A proposed budget, with adoption funded alongside technology
The budget ranges and even split below are this site's planning proposal. The 72/28 allocation is a separate survey finding.
The 72/28 split is reported by Capgemini's 2026 World Property & Casualty Insurance Report release. McKinsey's change-management advice concerns effort; it does not establish the proposed budget ratios.
5 · Why most insurers fail, and the countermeasures
| Failure mode | Evidence | Countermeasure |
|---|---|---|
| Pilots that do not progress | BCG reported 7% of insurance respondents had scaled AI in its 2024 study3; definitions and samples vary | Phase gates tied to production metrics, not demos; kill/scale decision at each gate |
| Data not ready | 78% of life-insurance study respondents cite data readiness11 | Data readiness assessment and document pipeline in Phase 1, before scaling |
| Adoption failure | ~2/3 of the challenge is people (BCG38) | 1:1 adoption budget; underwriters co-design tools; AI responsibilities in job descriptions |
| Poor sourcing decision | A vendor product may need substantial integration; an internal build needs ongoing ownership | Compare both on total cost, fit, support, and results on representative cases |
| No measurement | 42% of insurers track no AI metrics (Capgemini6) | Every pilot has a named owner and a metric tied to expense ratio, hit ratio, or cycle time |
| Governance as afterthought | 56% cite regulatory uncertainty as top scaling barrier10 | Governance program is Phase 1 deliverable #1, not a later retrofit |
6 · Regulatory non-negotiables, sequenced with the phases
- Before a pilot starts: map the entity, jurisdictions, data, and intended decisions to applicable requirements. Assign accountability, approve tools, and record controls and evaluations. The NAIC Model Bulletin describes supervisory expectations; adoption and scope depend on state action.1213
- Before selection or pricing use: document validation and discrimination testing, retain human-override authority, and confirm applicable professional standards. New York DFS Circular Letter 7 and Colorado's external-data rules have distinct scopes; satisfying one does not establish compliance with the other.1516
- Before consumer-facing use or greater automation: have Legal/Compliance confirm notice, explanation, review, and appeal requirements. Test monitoring, escalation, incident response, and the ability to stop the system before release.
- E&S: have counsel assess the domicile, licensing status, and placement jurisdictions. Rate/form flexibility does not by itself determine which AI or consumer-protection requirements apply. See Insurance & specialty.
This section sequences the obligations against the phases; it does not restate them. What is required, why, and a policy template to adapt are on Governance, which is the canonical statement for the site.
Sources & references
Superscript numbers in the text point here. Links verified August 2026; a few publishers (BCG, McKinsey, WTW) block automated checks but open normally in a browser. Carrier-reported figures are as reported by the companies or their vendors: not independently audited or guaranteed to transfer to another workflow. For the load-bearing claims, verbatim source passages are on the evidence page.
- Deloitte, "Scaling generative AI in insurance" (April 2025; survey June 2024). deloitte.com
- Bain & Company GenAI adoption figure, as reported by actuary.info, "The AI-proof gap in insurance governance" (2026). actuary.info
- BCG, "Insurance Leads AI Adoption. Now It's Time to Scale" (2024/25). bcg.com
- McKinsey/LIMRA, "Insurance 360: Industry trends" webinar deck (Nov 2025). limra.com (PDF)
- BCG, "The AI-First Property and Casualty Insurer" (2026). bcg.com
- Capgemini, World Property & Casualty Insurance Report 2026 (press release, May 2026). capgemini.com (PDF)
- McKinsey, "The future of AI in the insurance industry" (2025). mckinsey.com
- WTW, "Insurers using advanced analytics and AI report strong returns on investment and premium growth" (Mar 2026). wtwco.com
- actuary.info, "Insurtech H1 2026: AI funding concentration" (Gallagher Re data, Jul 2026). actuary.info
- Grant Thornton, "Insurance Insights 2026: AI Impact Survey" (2026). grantthornton.com
- LIMRA/Equisoft, "Assessing data readiness for AI in the life insurance industry" (Jan 2025). equisoft.com
- NAIC, Model Bulletin "Use of Artificial Intelligence Systems by Insurers" (adopted Dec 2023). content.naic.org (PDF)
- NAIC, AI Model Bulletin state adoption map (accessed Aug 2026). content.naic.org (PDF)
- NAIC, AI Systems Evaluation Tool pilot project summary (12 states, Mar–Sep 2026). content.naic.org (PDF)
- New York DFS, Insurance Circular Letter No. 7 (2024): AI in underwriting and pricing. dfs.ny.gov
- Colorado DOI, SB 21-169 algorithm and external-data governance regime. doi.colorado.gov
- AIG, 2025 Annual Report: Lexington submission volume and Middle Market Property submit-to-bind results. aig.com
- Markel: "AI Centre of Enablement" press release (Mar 2026), markel.com; Cytora case study (2023: +113% GWP/FTE, strategic-partner quote SLA 24h → 2h), cytora.com; Cortex unit launch, reinsurancene.ws
- Chubb AI combined-ratio target, as reported by actuary.info (Q2 2026 earnings coverage). actuary.info
- CFC, "CFC pilots agentic underwriting with launch of Lane Assist" (Apr 2026). cfc.com
- Kinsale Q4 2025 earnings call (enterprise AI licensing, internal bots). fool.com
- Cytora, "Zurich scales agentic AI to 5 countries in 90 days" (2026). cytora.com
- Hiscox Q1 2024 trading statement: 2023 terrorism-risk quote proof of concept, 3 days → 3 minutes; implementation into the live environment underway. hiscoxgroup.com
- Tokio Marine HCC, "Strategic collaboration with Cytora" (Dec 2025). tmhcc.com
- hyperexponential, "Powering Allianz Commercial pricing transformation" (Jun 2026). hyperexponential.com
- Ryan Specialty, 2025 annual report, technology and operations discussion. ir.ryanspecialty.com
- CRC Group, "REDY INTEL" (Mar 2026), crcgroup.com; Insurance Business, "Behind WTW's AI number: Neuron" (2026), insurancebusinessmag.com
- Ivans, "2025 Insurance Agency-Carrier Connectivity Trends Survey" (2025). ivans.com
- Reinsurance News, "Sixfold introduces AI Underwriter" (customer results: reported processing-time reductions of 50–97%, +15% hit ratios, +30% GWP/underwriter). reinsurancene.ws
- CLARA Analytics, litigation cost data ($77,807 vs $15,936 attorney-involved vs unrepresented). claraanalytics.com
- CLARA Analytics, "Solution to systemic over-reserving" case study (12.8x ROI). claraanalytics.com
- Verodat, bordereaux management (85–94% processing time savings). verodat.com
- Akur8 Discover (rate-filing research automation). akur8.com
- BCG, "Agentic AI for P&C insurance portfolio management" (2026). bcg.com
- Peakflo, "Insurance claims leakage prevention with AI" (leakage rates and pre-payment prevention). peakflo.co
- InsuranceIndustry.ai, "Billions left behind: AI and the economics of subrogation" ($15–20B uncollected). insuranceindustry.ai
- Project NANDA, "State of AI in Business 2025" report (exploratory enterprise GenAI research; not an insurance-specific success benchmark), via MLQ.ai. mlq.ai (PDF)
- BCG, "To Win with AI, Insurers Must Go Beyond the Algorithm" (2025). bcg.com
- KPMG, "2025 Insurance CEO Outlook" (AI budget allocation and ROI timeline expectations). kpmg.com (PDF)
- Guidewire, "Guidewire introduces Qusar release" (Aug 2026). guidewire.com
- Duck Creek, "Duck Creek acquires Send" (Jul 2026). duckcreek.com
- NIST, AI Risk Management Framework. nist.gov
- Accenture, "An AI Future for Insurance" (FY26). accenture.com (PDF)
- Datos Insights, "Insurer IT in 2026: bigger budgets, bolder AI" (IT spend ~4.5% of GWP). datos-insights.com
- Capco, "Bionic underwriting" (Jun 2026): SME underwriting is automation-led; large corporate is "bespoke, slow and expert-driven"; 30–40% of underwriter time is often administrative. capco.com
- Deloitte, "Amplifying core modernization in specialty insurance" (2026): many specialty insurers process only 20–30% of incoming submissions; manual reentry and disconnected workflows constrain underwriting. deloitte.com
- Hiscox, AI lead underwriting announcement (Dec 2023), hiscoxgroup.com; Ivans, Arch Insurance submission-intake case (May 2026), ivans.com.
- Deloitte, "Agentic AI: Transforming pricing analysis in insurance" (2026): autonomous filing analysis and competitive pricing comparisons. deloitte.com
- Aviva and hyperexponential, London Market pricing AI pilot (Jul 2025): enhancing pricing tools with new insights and optimizing model code. hyperexponential.com; Actuarial Agent model/rater demonstration, info.hyperexponential.com.
- SOA Research Institute, "Agentic AI for Actuarial Workflows" research scope (2026): rate and assumption development, financial modeling, model governance, documentation, and regulatory reporting. soa.org