7 min

read

Product Recommendation Engine for P&C Insurance | Higson

Written by

Łukasz Niedośpiał

Published on

09 Dec 2024

Last update

16 Sep 2026

Key takeaways

A product recommendation engine suggests the next policy, coverage or limit a customer is likely to need, based on their data and behavior.
In P&C insurance, are commendation is only useful if the carrier can actually write it: the product must be filed in that state, the risk must fit underwriting appetite, and the offer must be explainable.
The working pattern: a model suggests, business rules decide, and every decision is logged.
The NAIC Model Bulletin on AI explicitly covers marketing, sales and distribution - cross-sell models fall under the same governance expectations as underwriting models.
A business rules engine like Higson runs the eligibility and offer logic in real time and keeps an audit trail of what was recommended, to whom, and why.

What is a product recommendation engine?

A product recommendation engine is software that analyzes customer data and suggests the products or services a customer is most likely to need next. In retail, that means "customers who bought this also bought." In insurance, it means recommending an umbrella policy to a homeowner with two vehicles, or a cyber add-on to a small business buying a BOP.

The engine combines three things: data about the customer, a method for scoring what they might want, and logic that decides which suggestion to show. In e-commerce, the last step is simple - show the top-scoring items. In insurance, it is where most of the work happens. For the fundamentals of that decision logic, see What Is a Rules Engine?.

Why is recommending insurance different from recommending sneakers?

An online store can recommend almost anything in its catalog to almost anyone. A carrier cannot. Before any policy, endorsement or limit is offered, several conditions must be true:

‍State availability. The product and form must be filed and approved in the customer's state.

‍Underwriting appetite. The risk must fit current guidelines - a coastal property may be out of appetite for water backup, even if the model scores it as a likely buyer.

‍Eligibility. Prior claims, vehicle type, construction class or business classification can rule an offer out.

‍Pricing. The quoted premium must come from filed rates and rating tables, not from a marketing estimate.

‍Explain ability. If a regulator or an agent asks why one customer saw an offer and another did not, the carrier needs an answer.

Are commendation that fails any of these checks is worse than no recommendation.It either dies at underwriting, frustrates the customer, or creates a compliance issue.

How does a product recommendation engine work in P&C insurance?

A production setup at a P&C carrier usually has four layers:

Layer	What it does	Example
1. Signals	Collects data about the customer and the moment	New vehicle added mid-term, home purchase, quote abandoned at coverage selection
2. Scoring	Estimates what the customer is likely to want	A propensity model scores umbrella interest at 0.72
3. Decision rules	Filters and ranks offers against state, appetite, eligibility and pricing rules	Umbrella available in this state; household meets minimum liability limits; offer approved
4. Delivery and logging	Shows the offer in the right channel and records the decision	Agent portal shows "Recommend umbrella, $1M"; decision logged with rule version

Layers 1 and 2 answer what might this customer want. Layer 3 answers what can we offer them today. Most carriers already have some version of the first two. The third is where recommendations break down when business rules arehard-coded in a policy admin system or scattered across spreadsheets.

What data does an insurance recommendation engine use?

Insurance recommendations draw on first-party policy data far more than on browsing behavior:

‍Policies in force: lines, limits, deductibles, endorsements, household or account structure.

‍Life and business events: new vehicle, new driver, home purchase, payroll growth, new location.

‍Quote and service behavior: andoned quotes, coverage options viewed, questions asked in service channels.

‍Claims history: frequency, severity and type - relevant both for need (a water loss suggests water backup coverage) and for eligibility.

‍Third-party data: property characteristics, businessclassification, vehicle data.

External consumer data deserves extra care. Colorado's SB21-169 set binding requirements for insurers using external consumer data and algorithms, and other states have followed with their own frameworks. Before a data source feeds a recommendation model, compliance should know about it.

Which recommendation methods work for insurers?

The three classic methods still apply, with insurance-specific limits:

Method	How it works	Insurance example	Limit
Collaborative filtering	Recommends what similar customers bought	Households with teen drivers often add higher liability limits	Needs large volumes; weak for small commercial or niche lines
Content-based filtering	Recommends products similar to what the customer already has	Homeowner with scheduled jewelry → valuable articles endorsement	Tends to recommend "more of the same"
Hybrid models	Combines both, often with gradient boosting or similar ML	Propensity models for bundling at renewal	Harder to explain without feature attribution
Rules-based recommendations	Recommends based on explicit conditions set by product and underwriting	"If home + 2 autos + no umbrella, and state = TX, offer umbrella"	Needs regular review; does not discover new patterns

In practice, carriers rarely choose one. The strongest setups use ML to find patterns and rules to turn them into offers the carrier can stand behind.

Where do P&C carriers use product recommendations?

The use cases with the clearest payback are close to moments when the customer is already making a decision:

‍Bundling at quote. A customer quoting auto online gets a home or renters offer, priced from filed rates, before they bind.

‍Coverage gaps at renewal. The renewal package suggests water backup, higher liability limits or a deductible change, depending on the property and claims history.

‍Agent next-best-action. The agent portal shows the one or two most relevant offers per account, with the reason, so agents stop guessing.

‍Mid-term events. Adding a vehicle or a driver triggers a check for umbrella eligibility or higher limits.

‍Small commercial add-ons. A BOP quote triggers a cyber or EPLI offer when the business class and revenue fit appetite.

‍Embedded and partner channels. A partner's checkout shows only the products the carrier has approved for that partner, state and customer profile.

Why does a recommendation engine need a business rules engine?

Because in insurance, the model is allowed to suggest, but not to decide. A rules engine is the layer that turns a score into an offer the carrier can write, price and defend.

This matters for regulation. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (December 2023) states that AI is used"across all stages of the insurance life cycle, including product development, marketing, sales and distribution." In other words, across-sell model sits in the same governance scope as an underwriting model.The bulletin is not a law; it sets regulatory expectations, and states adopt it through their own bulletins. As of NAIC's April 2026 implementation map, 24states and the District of Columbia had adopted it, while California, Colorado,New York and Texas operate their own insurance-specific frameworks.

For acarrier, the practical requirements are the same everywhere:

‍Guardrails. Underwriting sets the boundaries; the model recommends only within them.‍
Versioning. Every change to eligibility or offer rules has an owner, a date and a reason.‍
Audit trail. For any customer, you can show which rules and which model version produced the offer.‍
Separation of duties. Product teams can adjust offer rules; underwriting approves appetite rules; data science owns the models.

This is also what gets recommendation projects past internal review. Data science teams often build good models that never reach production because the Chief Underwriting Officer cannot see how the output is constrained. Rules that underwriting owns are the answer to that objection.

What business results can carriers expect?

Personalization pays off, but the numbers most often quoted come from outside insurance. McKinsey's cross-industry research found that personalization most often drives a 10–15% revenue lift, with a company-specific range of 5–25%; 71% of consumers expect personalized interactions, and 76% are frustrated when they do not get them (McKinsey, November 2021).

For a P&C carrier, it is more useful to define the metrics before launch than to rely on industry averages:

‍Policies per household or account - the core cross-sell metric.‍
Offer acceptance rate by channel and by offer type.‍
Quote-to-bind rate for sessions where a recommendation was shown.‍
Retention of bundled versus single-policy customers.‍
Loss ratio of cross-sold business - a recommendation that grows premium in the wrong risks is a cost, not a win.

Pick one line of business, one channel and one metric for the first release. That makes the result measurable and the next business case easier.

How does Higson power insurance product recommendations?

Higson is a business rules engine and insurance product configurator built for P&C carriers. In a recommendation setup, it is the decision layer: it holds the eligibility, appetite and offer rules, runs them in real time, and records every decision.

‍Decision tables that product and underwriting teams maintain. Offer rules, state availability and eligibility conditions live in tables that business users edit and test. Rating tables prepared in Excel are imported through CSV.

‍Your models inside the rules. Higson's ONNX runtime (since version 4.2) runs your machine learning models inside rule execution. Your data science team keeps ownership of its models; Higson applies the business constraints around them.

‍Real-time performance. Rule execution takes 0.23 ms per decision, with through put of 9,000 requests/second - fast enough to evaluate offers during an online quote or an agent call.

‍Testing before production. The tester lets teams run a rule change against sample cases before it goes live, and versioning makes rollback straightforward.

‍Audit trail and granular permissions. Higson 4.3 LTS adds role-based permissions at the level of business domains, sessions and operations, with a full audit trail and rollback. Product can change offer rules without touching underwriting appetite rules.

‍AI-ready integration. An official MCP server with 50+ tools lets AI agents work with rules, and the AI Assistant in Higson Studio (beta) drafts functions in Groovy or Python and decision tables from natural-language descriptions.

This setup removes backlog from the IT queue: product and underwriting teams own the rules, and IT gets more time for architecture and integration.

Carriers already use Higson for the rules that recommendations depend on. InterRisk (Vienna Insurance Group) runs the sales rules of its IRON digital sales platform on Higson. Allianz Poland uses Higson as a central product configuration layer and launches product variants in days instead of months. At TUW, actuaries publish a new motor rate in 2 hours instead of 2 weeks.

How do you start with a recommendation engine?

Start small and close to an existing decision point:

‍Pick one use case. Umbrella at renewal or home-and-auto bundling at quote are common first choices.‍
Write the rules first. Before modeling, document which customers the carrier is allowed and willing to offer the product to, by state.‍
Add scoring second. A simple propensity model, or even rules alone, is enough for a first release.‍
Define one metric. Offer acceptance or policies per household, measured against a control group.‍
Plan the timeline realistically. Platform go-live with Higson typically takes 3–6 months. After that, each new product or offer takes weeks, and a single rule change takes hours.

See how your recommendation rules would work in Higson

Bring one use case - bundling, renewal or agent next-best-action - and we will show you how the eligibility and offer rules would look in Higson, tested on your own sample data.

Book a 30-minute call

‍Download the Business Rules Engine Comparison Guide 2026

Sources

NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers -https://content.naic.org/sites/default/files/inline-files/2023-12-4%20Model%20Bulletin_Adopted_0.pdf
NAIC, Implementation of NAIC Model Bulletin - https://content.naic.org/sites/default/files/cmte-h-big-data-artificial-intelligence-wg-map-ai-model-bulletin.pdf
AIPMO, NAIC AI Bulletin Adoption: Q2 2026 - https://aipmo.co/naic-ai-bulletin-q2-2026-status/
McKinsey, „The value of getting personalization right - or wrong - is multiplying" -https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/the-value-of-getting-personalization-right-or-wrong-is-multiplying