AI Readiness Assessment

Regional P&C Insurance Company — Underwriting Automation, Claims Triage & Fraud Detection

Date: 2026-05-30 · Prepared by: Resolvix · Status: Sample Deliverable
Deliverable type: AI & Automation — AI Readiness Assessment (~$850)
Industry: Property & Casualty Insurance

About this sample. This is one example of what a successful Resolvix deliverable looks like at this scope and type — not a template that every engagement follows. Your expert brings their own expertise and judgment to the work: the structure, the emphasis, which angles they dig into, and how they organize their findings will all vary based on your industry, your specific question, and where the research leads. What stays consistent across every engagement is the standard: analysis grounded in evidence, prioritized recommendations, concrete action steps, and a phased implementation plan. All company names, figures, and scenarios in this sample are illustrative.

Executive Summary

Lakeshore Mutual Insurance (illustrative; $380M direct written premium, 450 employees, four regional offices) is evaluating AI deployment across three use cases: underwriting automation, claims triage, and fraud detection. This assessment scores the organization across eight readiness dimensions and maps each score to specific gaps and remediation actions.

Overall readiness: 58/100 — Conditionally Ready. Lakeshore has viable data infrastructure and a clear business case for all three use cases, but faces meaningful gaps in data governance, ML talent, and model risk management that must be closed before production deployment. A structured 18-month program — beginning with fraud detection as the lowest-risk, highest-ROI entry point — is recommended.

Estimated three-year value at stake: $14.2M–$19.8M across all three use cases, net of implementation costs.

Company Context

Readiness Scoring Framework

Each dimension scored 1–10. Scores below 5 represent blocking gaps that require remediation before that use case can go to production. Scores 5–7 represent conditional readiness (mitigations required). Scores 8–10 represent production-ready.

Dimension 1: Data Quality & Availability — Score: 6.2/10

What was assessed

Volume, completeness, label quality, and historical depth of data for each proposed use case.

Findings

Underwriting automation:
- 7 years of structured policy application data in Majesco — sufficient for supervised learning on acceptance/declination and tier placement
- Loss history linkage to policy records is 84% complete; the remaining 16% requires manual reconciliation and represents a material training data gap for loss-predictive models
- Third-party data feeds (LexisNexis, Verisk CLUE) are already ingested but stored in separate schema with no unified entity key — joining to policy records requires custom ETL

Claims triage:
- Guidewire ClaimCenter holds 6 years of structured claim records (coverage type, reserve amounts, settlement outcomes)
- Adjuster notes are stored as unstructured free text; approximately 1.1M historical notes are available but have never been parsed or labeled
- 22% of claims lack a "complexity score" or routing rationale in the structured fields — complexity decisions were made verbally and never recorded

Fraud detection:
- SIU (Special Investigations Unit) has flagged and labeled 3,400 confirmed fraud cases over five years — above the minimum threshold (~2,000) for a supervised classification model but below the volume that supports deep segmentation by fraud type
- Social network linkage data (shared addresses, phone numbers, repair shops) exists in ClaimCenter but is not currently extracted or graphed

Gap Analysis

Dimension 2: Process Maturity & Documentation — Score: 5.1/10

Findings

Lakeshore's underwriting workflows are partially documented at the product line level but lack step-level process maps that would support workflow automation design. Interview with the VP of Underwriting revealed that approximately 40% of acceptance decisions for personal auto involve "judgment factors" not captured in the rating manual — underwriters apply risk intuition that has never been codified. This is a critical gap for underwriting automation: an ML model trained only on structured application data will systematically miss the soft signals that experienced underwriters use.

Claims triage workflows vary by office. The Cleveland office routes claims by reserve size thresholds; the Columbus office uses adjuster availability as the primary routing variable. There is no company-wide triage protocol. This inconsistency means any AI triage model trained on historical routing decisions will inherit conflicting logic and produce unreliable output.

Fraud referral workflows are the most mature: the SIU has a written referral criteria document (last updated 2024), consistent intake forms, and defined disposition codes. This consistency makes fraud detection the most process-mature use case.

Dimension 3: IT Infrastructure & Integration Capability — Score: 6.8/10

Findings

Snowflake's deployment 18 months ago represents a significant infrastructure investment that creates a viable ML data platform. However, Guidewire ClaimCenter remains on-premises at v10.1, which does not have a native REST API for real-time inference calls. Any real-time AI output (e.g., a fraud score at first notice of loss) would require a middleware layer or an upgrade to ClaimCenter Cloud.

Majesco (policy admin) is cloud-hosted with REST API access — underwriting model integration is technically feasible without infrastructure changes.

Current CI/CD tooling (Azure DevOps) is adequate for model deployment pipelines with moderate configuration effort. The IT team has no existing MLOps tooling (no Weights & Biases, MLflow, or equivalent for experiment tracking and model versioning).

Infrastructure gap summary:
- ClaimCenter on-prem creates real-time inference latency and integration complexity for claims triage and fraud scoring at FNOL
- No MLOps platform for model lifecycle management
- No feature store — each model team would rebuild the same feature engineering pipelines independently

Dimension 4: ML Talent & Organizational Capability — Score: 3.5/10

This is the most significant gap. Lakeshore has no in-house data science or ML engineering capability. The data team consists of three business intelligence analysts who work primarily in SQL and Tableau. There is one "data engineer" title, but the role has focused on Snowflake pipeline maintenance rather than ML infrastructure.

To build and maintain proprietary ML models, Lakeshore needs at minimum: one senior ML engineer (model development and MLOps), one data scientist (feature engineering, model validation, business translation), and a part-time model risk function (internal audit or third-party). Hiring all three at current market rates would add approximately $520K–$650K in annual fully-loaded cost before the first model goes to production.

Viable alternatives to a full build-out:
1. Vendor-embedded AI (e.g., Guidewire Predict for claims, Shift Technology for fraud) — pre-built models fine-tuned on carrier data, no in-house ML team required
2. Managed ML service (e.g., AWS SageMaker with a system integrator) — infrastructure managed externally, Lakeshore provides business oversight
3. Hybrid: buy fraud detection (highest complexity, lowest internal data advantage), build underwriting scoring (highest internal data advantage, most defensible moat)

The hybrid approach is recommended. See Recommendations section.

Dimension 5: Data Governance & Model Risk Management — Score: 3.8/10

Lakeshore has no formal model risk management (MRM) policy. This is a material exposure. Insurance regulators in Ohio and Michigan (Lakeshore's primary states) have issued guidance requiring carriers to document, validate, and audit any AI model used in underwriting or claims decisions affecting policyholders. The NAIC Model Bulletin on AI (adopted by Ohio in 2024) requires:

• A model inventory with documented intended use, training data lineage, and validation results
• Pre-deployment independent validation for models affecting adverse action decisions
• Ongoing performance monitoring with defined drift thresholds
• Consumer disclosure obligations for automated adverse actions

Lakeshore has none of these controls in place. Deploying an underwriting automation model without MRM infrastructure would create regulatory exposure — Ohio DOI can require a carrier to cease using a model and reprocess affected policies.

Fraud detection carries lower regulatory risk (models flag for human review, not automated adverse action) and is an appropriate starting point while MRM infrastructure is built.

Dimension 6: Change Management & Stakeholder Readiness — Score: 5.5/10

Interviews with underwriting and claims leadership surfaced meaningful resistance. The VP of Underwriting expressed concern that model scores would be used to evaluate underwriter performance rather than assist decisions — a fear that has caused adoption failures at peer carriers. Three of seven claims supervisors interviewed described AI triage as a threat to adjuster roles rather than a productivity tool.

The CEO and CFO are aligned on the business case. The CTO is supportive but lacks ML credibility with technical staff, making it difficult to address data science skepticism from within.

Change management readiness is moderate. Lakeshore has completed two prior technology change programs (Snowflake migration, Majesco upgrade) with reasonable adoption outcomes, suggesting the organization can absorb change when rollout is structured. The risk is specific to AI/automation anxiety rather than general change fatigue.

Dimension 7: Vendor & Partnership Ecosystem — Score: 6.5/10

Lakeshore has active relationships with Verisk (data and analytics), LexisNexis (risk data), and its Guidewire implementation partner (Infosys). These relationships are assets: Verisk and LexisNexis both offer ML-ready data products that reduce feature engineering burden, and the Guidewire SI relationship provides a credible path to ClaimCenter integration.

No current relationship with an AI/ML system integrator or InsurTech vendor in the fraud or underwriting space. Evaluation of Shift Technology (fraud), Zelros (underwriting AI), and Guidewire Predict (claims) would be appropriate in Phase 1.

Dimension 8: Financial Case & Executive Alignment — Score: 7.2/10

The CFO has approved a preliminary exploration budget of $180K for this assessment and vendor evaluations. There is executive consensus that loss ratio improvement of 1–2 combined ratio points would justify significant AI investment.

Preliminary ROI model by use case:

Three-year net value (all three use cases, net of $3.2M estimated implementation cost): $14.2M–$19.8M

Readiness Scorecard Summary

Two dimensions score below 5 (blocking): ML talent and model risk management. These must be addressed before any use case goes to production. Fraud detection has the highest tolerance for starting while these gaps are closed (because it supports human decision-making rather than automated adverse action), making it the correct entry point.

Use Case Sequencing Recommendation

Sequence: Fraud First → Claims Triage Second → Underwriting Automation Third

Fraud detection (Months 1–9):
Lowest regulatory risk, highest ROI confidence, most mature internal data. Recommended approach: evaluate Shift Technology (vendor-built model fine-tuned on Lakeshore data via federated learning) rather than building in-house. Shift's P&C fraud model achieves 3–5x the referral accuracy of rules-based SIU triggers on comparable carrier datasets. Expected implementation timeline: 6–9 months to production. SIU headcount can be redeployed from intake triage to investigation depth — net of vendor cost, expect $1.8M–$2.4M Year 1 value.

Claims triage (Months 6–15):
Requires process standardization across offices before model deployment (approximately 90 days of pre-work). Recommended approach: a RAG-based (retrieval-augmented generation) routing assistant that pulls from historical claim records and adjuster notes rather than a pure supervised model — this reduces the dependency on labeled complexity scores. Guidewire Predict is a credible option; a custom fine-tuned model on Lakeshore's own notes is possible but requires the ML talent gap to be closed first.

Underwriting automation (Months 12–24):
Highest regulatory exposure, requires completed MRM framework, requires process documentation of "judgment factors" (a 3–4 month effort). Recommended approach: a tiered model where straight-through processing (STP) applies only to the 35–40% of personal auto applications that fall clearly within appetite (score above a validated threshold), with all other applications routed to underwriter review with a model-generated risk summary. This is augmentation, not replacement — which also addresses the change management risk.

Recommendations

1. Appoint a Model Risk Officer (or engage an interim MRM consultant) before any model reaches production. The NAIC AI Model Bulletin compliance gap is the single highest-priority action item. Ohio DOI has issued inquiries to at least three Ohio-domiciled carriers in the past 18 months regarding AI governance. Engage an InsurTech MRM consultant (Milliman, Oliver Wyman, or a specialist boutique) to build the model inventory, validation protocol, and ongoing monitoring framework. Timeline: 60 days to initial framework, 120 days to full policy. Cost: $80K–$140K external.

2. Do not build in-house ML for fraud detection — buy a purpose-built InsurTech solution. The talent gap (Score: 3.5/10) makes a build-first approach high-risk and slow. Issue an RFP to Shift Technology, Friss, and Verisk's Jornaya for fraud detection. Evaluation criteria: federated learning capability (so Lakeshore data stays in-house), Guidewire integration pathway, model explainability for SIU documentation, and pricing model (per-claim fee vs. percentage of recovered fraud). Target vendor selection in 60 days.

3. Standardize claims triage protocols across all offices before beginning any AI triage project. The current inconsistency between Cleveland (reserve-based routing) and Columbus (availability-based routing) means historical routing data is unreliable as training signal. A claims operations working group should produce a unified triage protocol within 90 days. This is a process problem, not a technology problem — attempting to solve it with AI first will entrench the inconsistency.

4. Invest in MLOps infrastructure now, before the first model deploys. Stand up MLflow (open source, deployable on Azure) for experiment tracking and model versioning. Configure Snowflake as the feature store using dbt for feature pipeline management. This work can be done by the existing data engineer with 40 hours of external training — estimated cost $12K including training and tooling. Without this, the second and third models will have no reproducibility, no drift monitoring, and no auditability.

5. Launch a structured AI literacy program for underwriters and claims adjusters before any tool touches their workflow. Peer carrier case studies (Travelers, Erie Indemnity, Westfield) show that AI adoption failure in insurance is almost always a change management failure, not a technical failure. A 4-session program covering: how the models work at a conceptual level, what they can and cannot do, how model output will (and will not) be used in performance evaluation, and how to escalate when they disagree with a model recommendation. Budget: $15K–$25K for facilitation and materials.

6. Resolve the loss history linkage gap (16% incomplete) as a prerequisite to underwriting model development. Assign the data engineering team a 60-day sprint to reconcile unmatched loss records using probabilistic matching on policy number, insured name, and effective date. Document matching logic for the model data lineage record. This is not optional — training an underwriting model on 84% of the relevant signal while ignoring 16% will produce systematic bias in the model's loss predictions.

7. Extract and graph SIU network data from ClaimCenter before the fraud vendor evaluation. Shared address clusters, common phone numbers, and repeat repair shop relationships are among the highest-signal features for P&C fraud models. This data exists in ClaimCenter but has never been structured. A two-week data engineering effort to extract and graph these relationships will materially improve the quality of any vendor evaluation (you can test vendor models against your own labeled data with full feature sets) and will be a required input for the winning vendor's model fine-tuning.

Action Steps

Implementation Plan

Phase 1: Foundation & Fraud Detection (Months 1–9)

Objective: Close the two blocking gaps (MRM, MLOps) and deploy the first production AI capability (fraud detection via vendor solution).

Milestones:
- Month 1: MRM consultant engaged; fraud RFP issued; SIU network data extraction underway
- Month 2: MLflow and feature store deployed; fraud vendor finalist selected
- Month 3: MRM policy draft complete; loss history linkage reconciliation complete; claims triage unified protocol published
- Month 4: Fraud vendor integration begun (Guidewire ClaimCenter API configuration); AI literacy program delivered to Wave 1 (all SIU staff + claims supervisors)
- Month 6: Fraud model deployed in shadow mode (scores generated but SIU not yet acting on them); model output reviewed against SIU human judgments to validate calibration
- Month 7–8: Shadow mode validation complete; model meets go/no-go criteria (see below); SIU begins acting on model referrals
- Month 9: Fraud detection in full production; baseline performance metrics established

Go/No-Go Criteria (end of Month 6 shadow mode):
- Model precision on fraud referrals ≥ 65% (vs. current SIU rules-based system at ~40%)
- False negative rate (missed fraud) no worse than current baseline
- MRM validation sign-off complete
- No open regulatory or compliance flags from MRM review

Success Criteria:
- 25%+ improvement in SIU referral accuracy vs. pre-AI baseline
- $1.5M+ in recovered fraud leakage in Year 1 (partial year)
- SIU staff satisfaction score ≥ 3.5/5 on tool usefulness survey

Budget: $340K–$480K (MRM consultant $120K, fraud vendor Year 1 $180K–$280K, MLOps infrastructure $40K)

Phase 2: Claims Triage Automation (Months 6–15)

Objective: Deploy AI-assisted claims routing across all offices using standardized triage protocol established in Phase 1.

Milestones:
- Month 6: Claims triage vendor evaluation begins (Guidewire Predict vs. custom RAG build)
- Month 8: Architecture decision made (buy vs. build); project plan finalized
- Month 9: AI literacy program delivered to Wave 2 (all claims adjusters)
- Month 10: Claims triage model development/integration begins
- Month 12: Shadow mode deployment across one office (Cleveland pilot)
- Month 13: Cleveland pilot go/no-go review; expansion decision
- Month 15: Full production deployment across all offices

Go/No-Go Criteria (end of Month 12 pilot):
- AI triage routing agreement with experienced adjuster judgment ≥ 72% on holdout cases
- Average time-to-assignment reduced by ≥ 20% vs. pre-AI baseline
- No adjuster grievances or escalations related to perceived role displacement

Success Criteria:
- 25% reduction in average time-to-assignment across all offices
- 10–15% reduction in claims cycle time for AI-triaged claims
- Adjuster productivity improvement (claims handled per adjuster per month) ≥ 12%

Budget: $380K–$550K (software/vendor $220K–$340K, integration $100K, change management $60K)

Phase 3: Underwriting Automation (Months 12–24)

Objective: Deploy straight-through processing for in-appetite personal auto applications, with model-assisted underwriter review for all other submissions.

Milestones:
- Month 12: Underwriting judgment factor documentation complete; MRM framework validated for underwriting use case
- Month 14: Model development begins (in-house or SI-led, decision deferred to Month 10)
- Month 16: Internal model validation complete; pre-deployment regulatory review initiated
- Month 18: Pilot deployment (single personal auto product, one state)
- Month 20: Pilot go/no-go review
- Month 22: Expansion to full personal auto book
- Month 24: Small commercial underwriting model scoping begins

Go/No-Go Criteria (end of Month 18 pilot):
- STP rate for in-appetite submissions ≥ 30% (model only touches submissions it can confidently score)
- Modeled policy loss ratio within 3 combined ratio points of manual underwriting baseline after 6 months of exposure
- Regulatory review complete with no objections
- Underwriter satisfaction score ≥ 3.5/5 (model as assistant, not replacement)

Success Criteria:
- 35–40% of personal auto submissions processed via STP by end of Month 24
- 15% reduction in underwriting expense ratio on in-scope book
- No adverse regulatory findings in first annual model audit

Budget: $520K–$780K (model development $280K–$420K, MRM ongoing $80K, IT integration $100K, change management $60K)

Appendix: Scoring Methodology

Scores were derived from: document review (policy admin and claims system configuration, Snowflake schema documentation, existing analytics reports), structured interviews (VP Underwriting, VP Claims, VP IT, Chief Actuary, CFO, three claims supervisors, two senior underwriters), data profiling of Snowflake tables (row counts, null rates, key linkage rates), and benchmarking against readiness profiles of comparable regional P&C carriers that have successfully deployed AI in underwriting and claims.

Benchmarks drawn from: Celent 2025 P&C AI Adoption Survey, Majesco 2025 Insurance Technology Survey, and the expert's direct experience with comparable carrier AI programs.