AI Product Recommendations for Insurance Advisors
A life insurer wanted every advisor, including brand new ones, to offer the right product to the right customer, and wanted to understand what separates its best advisors from the rest.
Customer personas give useful suggestions before any sales history exists
SHAP and LIME show underwriters why two policies were matched
Each model version validated and reproducible for audit
Personal data processed only inside the insurer’s own environment
The short answer: we built a sequential product recommendation pipeline for a life insurer's advisors, with a customer persona framework so advisors with no sales history still get useful suggestions. We added advisor segmentation, an explainable peer matching system for group health renewal pricing, and maker checker validation on every model version.
Project Overview
Industry
Life and group health insurance
Use cases
Advisor product recommendations and renewal pricing
Engagement
Multi phase, 2025 to 2026
Team
Small data science team working with the insurer’s own data scientists
Technology Stack
The Challenge
Advisors with no history and pricing teams that need reasons
Advisor productivity depended on offering the right product to the right customer, but newly onboarded advisors had no sales history to learn from, which is a classic cold start problem. Separately, underwriters pricing group health renewals wanted to compare a policy with genuinely similar peers and see exactly why those peers were chosen.
What We Delivered
Recommendations advisors can use from day one
A sequential recommendation pipeline ranks the next best product for each customer, and a persona framework covers advisors and customers with little or no history. Advisor segmentation and hypothesis testing showed what top performers do differently, and a peer matching system gives underwriters explainable comparisons for renewal pricing. All of it was documented and handed over to the insurer’s own data science team.
Distribution
- Sequential product recommendation pipeline with an engineered feature set
- Customer persona framework to handle new advisors and new customers
- Advisor segmentation and hypothesis testing on what separates top performers
Underwriting and governance
- Peer comparison for group health renewal pricing: feature shortlist, clustering and nearest neighbour matching
- SHAP and LIME explanations of why two policies were matched
- Maker checker validation, versioned pipelines, documentation and knowledge transfer
Challenges & Solutions
New advisors with no sales history
Problem
A recommendation model trained on past sales has nothing to say about an advisor who has just joined.
Solution
A customer persona framework lets the pipeline recommend products from customer and context signals until an advisor builds a history of their own.
Impact
Useful recommendations from an advisor’s first week
Ranking the next best product
Problem
Advisors need a short, ordered list of products, not a single prediction, and the order has to be measured properly.
Solution
A sequential recommendation pipeline built with gradient boosted models, evaluated with ranking metrics such as MRR, nDCG and precision at K.
Impact
Recommendations judged the way advisors actually use them
Peer pricing underwriters can trust
Problem
Renewal pricing against peer policies only works if underwriters believe the peers are truly comparable.
Solution
A validated feature shortlist, clustering and nearest neighbour matching, with SHAP and LIME explanations of each match.
Impact
Explainable peer comparisons for renewal decisions
Regulated models need an audit trail
Problem
Pricing changes in insurance need actuarial review and sign off, not just good model scores.
Solution
Maker checker validation on every model version, versioned and reproducible pipelines, and all processing inside the insurer’s own cloud.
Impact
Models the actuarial and compliance teams can review
Related: AI consulting for insurance · All case studies
Frequently Asked Questions
How do you recommend insurance products to advisors with no sales history?
Use customer personas. Instead of learning only from an advisor’s past sales, the model groups customers into personas and recommends what has worked for similar customers. Once the advisor builds a history, their own data takes over.
How do you measure a product recommendation model?
With ranking metrics, because advisors see an ordered list. MRR, nDCG and precision or recall at K show whether the right products appear near the top, which matters more than raw accuracy.
What is peer matching in insurance renewal pricing?
It finds existing policies that genuinely resemble the one being renewed, using a shortlist of validated features, clustering and nearest neighbour search. Underwriters then price the renewal with those peers in view.
How do you make insurance models explainable?
Tools such as SHAP and LIME show which features drove a prediction or a match. Underwriters and actuaries can then see why the model reached its answer before they rely on it.
How is policyholder data protected in these projects?
Policyholder and advisor data is treated as personal data and processed only inside the insurer’s own cloud environment, with no extracts to local machines or outside services. Every model version goes through maker checker validation.
Want your advisors selling the right product?
Tell us about your distribution data. We will show you what a recommendation engine and advisor insight could look like for your team.