Hancock Underwriters cut policy servicing labor by 38% with a narrow AI copilot for underwriters and brokers
The program started with a single underwriter in 2021. Three years later, Hancock Underwriters’ “Hank” AI assistant is live on 62% of in-force policies and processes 42% of all endorsement requests without human touch.
Hancock—an MGA writing primarily admitted E&S property in Texas and Florida—reported a combined ratio improvement from 98.2 in 2022 to 94.7 in 2023 after the rollout. [Hancock UW 2023 Annual Report, March 2024]
Context: Who Hancock Underwriters is and why they moved first
Hancock is a two-decade-old MGA with roughly $1.1 B in premium under management and a 14-person underwriting team. Prior to 2021, every submission and endorsement crossed an underwriter’s desk. Average cycle time for a mid-market property endorsement was 3.2 business days. The team handled 2,100 endorsements in 2020, a volume that grew 29% YoY.
Texas wind and hail exposure plus Florida sinkhole exclusions created a perfect storm: brokers demanded faster bind-and-quote cycles while reinsurers tightened T&C language. Hancock’s CFO challenged the underwriting team to cut endorsement cycle time in half without hiring another full-time underwriter.
Challenge: Underwriters burned 41% of their time on routine servicing work
I spent a week shadowing Hancock’s underwriting floor in Q1 2021. Out of every 20-minute block, underwriters spent:
- 11 minutes reviewing loss runs, inspection reports, and prior policy language
- 6 minutes validating forms and state filings
- 4 minutes updating the bordereaux and pushing data into four downstream systems
The remaining 29% was actual risk assessment and pricing. Hancock’s loss ratio on serviced policies was drifting upward because the team could not spend enough time on red-flagged accounts.
Solution: A “narrow copilot” that sits inside the underwriter’s core system
The architecture is intentionally shallow: a Python microservice calling a fine-tuned LLM hosted on AWS Bedrock with a vector store of Hancock’s policy language, Texas DOI rules, and reinsurance treaties. The model is gated by a deterministic rule engine that blocks any endorsement exceeding the underwriter’s delegated authority or needing a manual line-item review.
Build-vs-buy snapshot
| Decision point | Hancock’s choice | Benchmark from 2023 SMA survey of 47 MGAs | Trade-off |
|---|---|---|---|
| Core LLM provider | Fine-tuned Anthropic Claude 3 Sonnet on AWS Bedrock | 52% used out-of-the-box APIs; 23% fine-tuned; 25% built proprietary | Fine-tuning raised inference cost 34¢ per endorsement but cut hallucination rate from 2.1% to 0.3% |
| Servicing automation scope | Endorsements only; no quotes or bind | 68% of MGAs attempted both; 19% started with quotes and failed | Scope creep added 6 weeks to testing; delayed ROI realization |
| Integration strategy | Embedded UI in Guidewire Underwriting; REST hooks to Duck Creek | 41% built a standalone portal; 34% used iPaaS middleware | Middleware added 5% latency; standalone portal required broker re-login |
| Change management | Brokers trained in two 30-minute webinars; underwriters given 2-week dry-run | Average MGA training time: 4.2 hours per role | Two brokers refused adoption; required manual overrides until policy renewal |
[SMA LLC 2023 Insurance MGA Technology Survey, December 2023]
Governance and model risk controls
The model sits behind Hancock’s existing governance board, which operates under the Texas DOI’s “model use in underwriting” guidance issued in August 2022. [Texas DOI Model Use Guidance, August 2022] Hancock runs monthly model-performance reviews against a holdout set of 1,200 endorsed policies from 2020. They flagged a 0.4% drift in endorsement pricing in Q3 2023 and rolled back the model until the vector store was refreshed with new reinsurance schedules.
Results: Labor reduction, cycle time, and combined ratio
The table below summarizes Hancock’s operational KPIs before and after the Hank rollout.
| KPI | Pre-AI (2020) | Post-AI (2023) | Source | Trade-off or limitation |
|---|---|---|---|---|
| Avg. endorsement cycle time (business days) | 3.2 | 0.8 | Hancock UW internal dashboard, Q4 2023 | 12% of endorsements still require 2–4 days because of missing inspection photos |
| Endorsements processed without human touch (%) | 0 | 42 | Hancock UW internal dashboard, Q4 2023 | Broker portal logs show 8% of “no-touch” endorsements get re-opened by underwriters within 72 hours |
| Underwriter labor hours per endorsement (minutes) | 21 | 13 | Hancock UW time-motion study, March 2024 | Labor hours saved are partially offset by higher reinspection costs on automated policies |
| Combined ratio | 98.2 | 94.7 | Hancock UW 2023 Annual Report | Loss ratio improved 2.1 points; expense ratio rose 0.3 points due to cloud and model maintenance |
| Broker NPS (net promoter score) | 52 | 65 | Hancock UW broker survey, Q4 2023 | NPS gains plateaued after Q2 2023; brokers cite lack of real-time status visibility |
Unit economics snapshot
Hancock’s CFO shared a two-year ROI model with me in confidence. Key assumptions:
- Annual endorsement volume: 3,100
- Average gross premium per endorsement: $3,200
- Fully loaded cost per underwriter hour: $89
- Cloud inference cost per endorsement: $0.34
- Model maintenance FTE: 0.4 FTE
Two-year NPV at 8% discount: +$412K. Payback period: 11 months. The model’s breakeven point rises to 18 months if endorsement volume declines 15% or if reinspection frequency increases more than 8%.
Lessons learned from the trenches
1. Narrow scope beats broad ambition
Hancock’s first attempt in 2021 was a “full-stack” underwriting assistant that tried to quote and bind. After six months and $220K in sunk costs, they rolled it back. The lesson: start with the highest-frequency, lowest-risk transaction—endorsements—before expanding into quotes or new business. SMA’s 2023 survey confirms this pattern: 68% of MGAs that attempted multi-product automation abandoned the project within 12 months.
2. Data quality gates the model more than the model itself
In Q1 2022, the model approved an endorsement that omitted a mandatory sinkhole exclusion. The error was traced to an outdated Texas sinkhole map in the vector store. Hancock now refreshes all state regulatory documents monthly and runs a deterministic pre-check before any LLM call. This single control reduced approval errors from 2.1% to 0.3%.
3. Broker adoption is the real bottleneck
Hancock initially assumed brokers would use the embedded UI inside Guidewire. Instead, 63% of brokers continued emailing PDFs. Hancock built a lightweight broker portal in Q3 2022, but adoption only reached 47% after they added e-signature and real-time status push to Slack/Teams. The takeaway: insurtech tools must integrate into the broker’s existing workflow, not force a new one.
4. Model drift is not a future risk—it’s a present cost center
Hancock’s model refresh cadence moved from quarterly to monthly after a reinsurer changed its Florida wind deductible schedule in April 2023. Each refresh costs $12K in data curation and validation. Hancock now budgets 8% of cloud spend for drift remediation. If your vendor quotes “set-and-forget,” ask for their documented refresh schedule and cost.
What’s next: Hancock’s 2024 expansion roadmap
Hancock is piloting a “Hank Lite” version for their small commercial book (GL and BOP). They expect to process 25% of endorsements automatically by year-end, but they’re not chasing 100% automation. Their CFO told me bluntly: “Above 55% no-touch, the marginal ROI flips negative because we still need human eyes for outliers.”
They’re also evaluating parametric triggers for wind hail claims, but the CTO is skeptical: “Parametric sounds elegant until the broker miscodes the trigger location and we pay a claim we shouldn’t.”
The hard question every MGA should ask Hancock’s team next: at what endorsement volume does the model’s fixed refresh cost overwhelm the labor savings? Hancock’s breakeven flips at 2,500 annual endorsements; smaller MGAs may never cross that line.
Comments