Building an AI maturity model for insurance: a step-by-step implementation guide Building an AI maturity model for insurance: a step-by-step implementation guide

I’ve reviewed six failed AI initiatives in insurance where the root cause was the same: no maturity model. Teams jumped straight to model deployment without first defining what “good” looks like. The result was predictable—misaligned stakeholders, inconsistent KPIs, and budgets that evaporated when executives couldn’t see ROI. After watching this pattern repeat, I built a lightweight maturity model specifically for insurance. It starts with a 4-level framework, uses a 24-question assessment, and is designed to be run in under 90 days with one FTE. Below is the exact playbook I’ve reused across four property & casualty insurers, one life & health carrier, and two MGAs. If you’re responsible for shaping AI strategy, this is the tactical guide you need.

Why a maturity model beats random pilots Insurance AI initiatives fail in three predictable stages:

Stage 1 (0-6 months): One-off pilots (e.g., “Let’s try generative AI on claims narratives”). Stage 2 (6-18 months): Tactical tooling (e.g., OCR for FNOL, RPA for bordereaux).

Stage 3 (18+ months): Re-platforming when executives realize they have 17 different AI tools and no way to measure them.

• A maturity model prevents this by forcing two things: a shared definition of “AI” and a shared vocabulary for progress. In practice, that means answering “What does Level 3 look like for underwriting?” before anyone writes a single line of Python. The model I use has four levels, which map to the insurance value chain:
• Level Scope
• Primary KPI Typical insurance use cases

Budget range 1. Ad-hoc

Playbook exists, training log, KPI dashboard 3 weeks

3. Managed AI projects are measured and controlled

Model risk framework, model inventory, audit trail 4 weeks

4. Optimizing AI projects drive measurable business outcomes

• Loss ratio impact, combined ratio delta, reinsurance savings 4 weeks
• The assessor scores each practice as “Not achieved,” “Partially achieved,” “Largely achieved,” or “Fully achieved.” The scores map to CMMI’s 0-100 scale. I’ve found that setting a target of 70+ for “Largely achieved” is realistic for most insurers. Step 3: Run the 24-question assessment
• The assessment is split into four tracks: Data & Infrastructure, Model Development, Operations & Governance, and Business Impact. Each track has six questions. The assessor must collect documentary evidence for each answer. Below is the exact question set I reuse:
• Data & Infrastructure Do you have a centralized data catalog that includes AI training datasets?
• Can you trace any AI output to its source dataset? Do you automate data quality checks for AI inputs?

Is your data infrastructure scalable to petabyte-scale AI models? Do you track data lineage for regulatory audits?

Is your data infrastructure PCI-DSS and SOC 2 compliant? Model Development

Do you use a standardized model development playbook? Do you validate model assumptions against historical claims?

• Do you bake model explainability into the development process? Do you test models against edge cases (e.g., catastrophe scenarios)?
• Do you document model limitations and failure modes? Do you version-control your model code?
• Operations & Governance Do you have a model risk framework approved by compliance?

Do you audit AI models annually? Do you run red-team exercises on AI systems?

I once had a CFO argue that a model risk framework was “Fully achieved” because “we have a policy.” The assessor pushed back—until the compliance counsel produced the actual framework document, which was three years old and hadn’t been updated. The score dropped to “Partially achieved.”

The output is a gap analysis: a list of practices that need improvement to reach the target level. The steering committee then prioritizes these gaps based on business impact and effort. Below is a template I reuse: Gap

Current Score Target Score

Business Impact Effort (FTE-weeks)

1. Dependencies Centralized data catalog
2. 0 4
3. High (loss ratio impact) 8
4. Data engineering team Model risk framework
5. 2 4
6. High (regulatory risk) 12

Legal, compliance AI KPI dashboard

1. 1 4
2. Medium (stakeholder alignment) 4
3. Data team
4. The steering committee then builds a 12-week roadmap. I’ve seen this roadmap fail when it’s treated as a wish list. To avoid this, each gap must have a single owner, a clear success criterion, and a budget. If any gap lacks these, it’s dropped from the roadmap.
5. Step 6: Build the 90-day execution plan The plan must be concrete, measurable, and resourced. Below is a template I reuse for a $5B P&C carrier targeting Level 3:
6. Week Action

Owner Success Criterion

1. Budget 1-2
2. Hire data catalog vendor (Collibra, Alation, or open-source) CTO
3. Data catalog live with 80% of AI training datasets cataloged $50k
4. 3-4 Build model inventory spreadsheet
5. Model risk team Inventory includes 100% of models in production
6. $10k 5-6

Run model explainability workshop Underwriting lead

1. All underwriting models include SHAP/LIME explanations $5k
2. 7-8 Build AI KPI dashboard (Power BI)
3. FP&A Dashboard includes loss ratio, cycle time, combined ratio
4. $15k 9-10
5. Run red-team exercise on pricing models Compliance
6. Red-team identifies 3 edge cases for pricing models $8k

11-12 Publish AI maturity report to board

The central team owns: Data catalog

Model registry AI KPI dashboard

Model risk framework The federated teams (claims, underwriting, actuarial) own:

1. Use case ideation Data labeling
2. Model explainability Business validation
3. The central team must be staffed as follows for a $5B insurer: Role

Headcount Responsibilities

> 2024-AI-Maturity-Assessment
> ├── 01-Data-Infrastructure
> │   ├── Data-Catalog.pdf
> │   ├── Data-Lineage-Screenshot.png
> │   └── PCI-Compliance-Letter.pdf
> ├── 02-Model-Development
> │   ├── Model-Playbook.pdf
> │   ├── Explainability-Report.pdf
> │   └── Version-Control-Log.pdf
> ├── 03-Operations-Governance
> │   ├── Model-Risk-Framework.pdf
> │   ├── Audit-Report.pdf
> │   └── Incident-Log.csv
> └── 04-Business-Impact
>     ├── ROI-Calculation.xlsx
>     ├── A-B-Test-Results.pdf
>     └── Peer-Benchmark-Report.pdf

AI Product Manager 1

Roadmap, prioritization, stakeholder alignment Data Engineer

2 Data catalog, data quality, infrastructure

ML Engineer 1

Model deployment, versioning, MLOps Model Risk Analyst

$88k (vendor tools, infrastructure) Pilot fails to show ROI

Scale 24 weeks

6 (central team) + 3 (federated data scientists) $600k (salaries, tools, training)

import pandas as pd
from openpyxl import load_workbook

# Load assessment data
scores = pd.read_excel("assessment_scores.xlsx", sheet_name="scores")
evidence = pd.read_excel("assessment_scores.xlsx", sheet_name="evidence")

# Map scores to CMMI levels
score_map = {
    "Not achieved": 0,
    "Partially achieved": 25,
    "Largely achieved": 75,
    "Fully achieved": 100
}

# Calculate total score
scores["numeric_score"] = scores["score"].map(score_map)
total_score = scores["numeric_score"].sum()

# Map to CMMI level
if total_score <= 24:
    level = "1. Ad-hoc"
elif total_score <= 49:
    level = "2. Repeatable"
elif total_score <= 74:
    level = "3. Managed"
else:
    level = "4. Optimizing"

# Generate gap analysis
gaps = scores[scores["score"] != "Fully achieved"]
gaps[["question", "score", "evidence_missing"]] = gaps.apply(
    lambda row: [row["question"], row["score"], evidence[evidence["question"] == row["question"]]["document"].isna().all()],
    axis=1, result_type="expand"
)

# Save output
with pd.ExcelWriter("ai_maturity_report.xlsx") as writer:
    scores.to_excel(writer, sheet_name="scores", index=False)
    gaps.to_excel(writer, sheet_name="gaps", index=False)
    pd.DataFrame({"total_score": [total_score], "level": [level]}).to_excel(writer, sheet_name="summary", index=False)