ML User Clustering for Insurance

The Insurance Industry's Comfortable Lie

Insurance companies have historically sold one product to "the Indian customer" — a monolithic persona built from aggregated surveys and actuarial tables. Young, salaried, urban, safety-conscious. Send them the same email, the same features, the same pricing tiers.

When we ran our first clustering analysis on 180,000 insurance platform users, we found not one customer but seven distinct behavioral archetypes — each with different purchase triggers, different price sensitivities, and different preferred features. The industry's comfort with the "average customer" had been leaving significant money on the table.

The most surprising finding: our highest-converting segment wasn't the young urban salaried buyer the industry optimizes for. It was a 35–50 age segment we hadn't built a single feature for. They were converting despite the product, not because of it.

What Triggered the Analysis

The Conversion Rate Anomaly

We'd noticed an anomaly in our conversion data: a small group of users was converting at 4× the platform average, but their demographics looked ordinary. Same age range, similar cities, similar income proxies. What was different wasn't who they were — it was how they behaved.

Metric	Platform Average	High-Converting Anomaly Group	Difference
Conversion Rate	6.8%	28.4%	+318%
Session Count Before Purchase	3.2	6.8	+113%
Features Explored	2.1	7.4	+252%
Time to Purchase	4.2 days	11.3 days	+169%
Return Visit Rate	31%	74%	+139%

The anomaly group took longer, visited more often, explored more features — and converted at 4× the rate. They weren't impulsive. They were thorough. And we were treating them identically to casual browsers.

The Seven Archetypes

What K-Means Actually Found

We ran K-means clustering with k=3 through k=12, measuring silhouette scores and business interpretability. k=7 gave us the best combination of statistical separation and explainable behavioral profiles. Here's what we found:

Archetype	% of Users	Purchase Trigger	Avg Premium	Conversion Rate
The Researcher	18%	Feature completeness + comparison depth	₹18,400	28.4%
The Price Hunter	22%	Lowest premium, highest deductible OK	₹7,200	19.1%
The Loyalty Seeker	14%	Known brand, even at premium	₹22,100	21.3%
The Life-Event Buyer	11%	Marriage, child, new vehicle trigger	₹15,800	31.2%
The Lapsed Renewer	16%	Reminder-driven, low-friction renewal	₹11,400	44.8%
The Feature Maximizer	9%	Specific add-ons (zero dep, NCB protect)	₹24,600	16.2%
The Passive Browser	10%	No clear trigger; rarely converts	—	2.1%

The Lapsed Renewer converted at 44.8% — highest of any archetype — but had almost no product features built for them. Renewal reminders were generic. The renewal flow had 7 steps. They were converting through friction, not because of it.

What It Meant

One Product, Seven Different Purchase Journeys

The insight wasn't just that these segments existed — it was that the product had been designed around one of them (the Researcher, roughly) and three others (Price Hunter, Life-Event Buyer, Lapsed Renewer) were converting significantly while being actively underserved.

We'd built one checkout flow. We'd written one set of email sequences. We'd designed one pricing page. And seven different types of buyers were trying to navigate all of it — some succeeding despite us, most giving up.

The highest-leverage play wasn't building new features. It was routing each archetype to an experience designed for their specific purchase trigger — using existing infrastructure, different sequencing.

The Framework

Archetype Routing System

We built a real-time classifier that assigns users to archetypes within their first 3 sessions, then routes them to differentiated experiences. The classifier doesn't ask users to self-identify — it infers from behavior.

Behavioral Classifier

Random forest model: session depth, feature usage sequence, time on comparison pages, price filter behavior, brand filter usage. Classifies into one of 7 archetypes at session 3 or after 8+ pageviews — whichever comes first.

Experience Routing

Each archetype gets different: landing page emphasis, email sequence, comparison UI defaults, pricing display order, and support trigger timing. Same underlying product; different journey emphasis based on purchase trigger.

Lapsed Renewer Fast-Track

Highest-value quick win: Renewal flow reduced from 7 to 3 steps. Pre-filled from previous policy. Single-click renewal for users whose policy expired within 30 days. Conversion rate went from 44.8% to 68.2%.

Life-Event Detection

Triggered by UTM (wedding searches, baby searches, car searches in preceding days via platform partner signals). Life-Event Buyers shown curated bundles rather than full comparison. Decision simplified.

Build

Technical Implementation

Clustering Pipeline

→K-means run on 22-feature behavioral vectors (session-level)
→Cluster stability validated with silhouette scores over 3 months of data
→Archetype labels assigned by business interpretability + statistical separation criteria

Real-Time Classifier

→Random forest classifier trained on labeled archetype data (ground truth: eventual cluster assignment)
→Classification triggered at session 3 or 8+ pageviews, stored in user profile
→Confidence threshold: > 70% probability to assign. Below threshold: default to Researcher journey

Experience Layer

→Feature flags control which UI variant each archetype sees
→Email system: archetype tag drives sequence selection
→Pricing page: default sort changes by archetype (Price Hunter: price ascending; Loyalty Seeker: brand prominence)

Rollout

Where We Started and How It Grew

Month 1 · Clustering Analysis

Finding the Seven

→ Seven archetypes confirmed. Lapsed Renewer and Life-Event Buyer identified as highest-leverage quick wins.

✓K-means clustering on 180K users across 22 behavioral features
✓k=3 to k=12 tested; k=7 selected (silhouette + interpretability)
✓Archetypes named, validated with business team

→ Seven archetypes confirmed. Lapsed Renewer and Life-Event Buyer identified as highest-leverage quick wins.

Month 2 · Quick Wins

Lapsed Renewer + Life-Event Flows

→ Lapsed Renewer CVR: 44.8% → 68.2%. Life-Event CVR: 31.2% → 47.6%.

✓Renewal flow simplified: 7 steps → 3
✓Pre-fill from previous policy activated
✓Life-Event bundle page built for wedding/baby/car contexts

→ Lapsed Renewer CVR: 44.8% → 68.2%. Life-Event CVR: 31.2% → 47.6%.

Month 3 · Classifier + Routing

Full Archetype System Live

→ Blended platform CVR: 6.8% → 11.2%

✓Real-time classifier built and validated
✓All 7 archetype journeys differentiated
✓A/B test: archetype routing vs standard journey for Researchers and Price Hunters

→ Blended platform CVR: 6.8% → 11.2%

Results

Across the Platform

+65%

Blended CVR

6.8% → 11.2%

68.2%

Lapsed Renewer CVR

from 44.8% baseline

Distinct Archetypes

each with tailored journey

3 steps

Renewal Flow

reduced from 7

Insights

What Surprised Us

The highest-converting segment was the most underserved

The Lapsed Renewer converted at 44.8% despite a 7-step renewal flow — the worst UX on the platform. They were succeeding through sheer intent. Optimizing for their specific need moved the needle faster than any other single change.

K-means found segments business intuition missed entirely

We'd never explicitly thought about a 'Life-Event Buyer' archetype. Business intuition categorized these users as 'first-time buyers.' The clustering revealed their distinct behavior pattern, and the tailored experience showed their purchase trigger was entirely different.

Confidence thresholds matter as much as model accuracy

A mis-classified user shown the wrong journey had lower conversion than an unclassified user shown the default journey. Setting a confidence threshold (we used 70%) for routing decisions improved outcomes — it's better to not classify than to classify wrongly.

Failures

Where We Stumbled

✕ Initial clusters were statistically clean but business-meaningless

Why it broke

First k=5 clustering produced clean statistical separation but the segments had no interpretable purchase trigger differences. 'Cluster 3' wasn't actionable.

What fixed it

Added business interpretability as a first-class criteria alongside silhouette score. Ran qualitative interviews with users from each candidate cluster to validate behavioral interpretation before acting.

✕ Feature Maximizer journey over-complicated the comparison UI

Why it broke

Assumed Feature Maximizers would convert better with more filter options. Added 12 new filter dimensions. Conversion dropped — too much choice, decision paralysis.

What fixed it

Narrowed to 5 most-purchased add-on combinations for that archetype. Showed 'most popular for similar buyers' as default. CVR recovered and exceeded baseline.

Future

Where This Goes

Dynamic Archetype Migration

Users can move between archetypes over time — a Price Hunter who becomes a Loyalty Seeker. Tracking archetype migration and re-routing in real-time rather than static assignment.

Predictive Archetype at Acquisition

Use pre-acquisition signals (search query, referral source, landing page) to predict archetype before first session. Start the journey correctly from the first pageview.

Cross-Product Archetype Consistency

Extend archetypes across product categories. A Price Hunter in insurance is likely a Price Hunter in health or motor. Cross-category archetype consistency for unified personalization.

There is no such thing as "the Indian insurance customer." There are at least seven — each with a different purchase trigger, a different journey, and a different definition of value. Build for all of them.

Predicting User Drop-Off Before It Happens (Churn Prevention Model)

Machine Learning

From 1K to 100K Monthly Users: Scaling MoneyRateFinder with SEO + Product Loops

Growth

Machine Learning

How We Predicted High-LTV Users Using Behavioral ML (100K+ Sessions)

10 min read

Machine Learning

Building a Real-Time Intent Scoring Engine for Lead Conversion

9 min read

Machine Learning

Using ML to Reduce CAC by 31% via Smart Audience Segmentation

8 min read

All Case Studies

Browse all 30 breakdowns

Work With Me

Let's build something together

Connect on LinkedIn

Follow for more insights

Where to go next

Fintech Product Consultant
The consulting work this sits inside, with the rest of the evidence behind it.
Day 30 retention benchmark, fintech & banking
Published figures, traced to their sources — and where the sources disagree.
Day 1 retention benchmark, fintech & banking
Published figures, traced to their sources — and where the sources disagree.
Product Growth Score
Thirteen questions that locate which part of the funnel is actually constraining growth.

Personalizing Insurance Offers Using ML-Driven User Clustering

Results at a glance

The Insurance Industry's Comfortable Lie

The Conversion Rate Anomaly

What K-Means Actually Found

One Product, Seven Different Purchase Journeys

Archetype Routing System

Technical Implementation

Where We Started and How It Grew

Across the Platform

What Surprised Us

The highest-converting segment was the most underserved

K-means found segments business intuition missed entirely

Confidence thresholds matter as much as model accuracy

Where We Stumbled

Where This Goes

You might also like

Where to go next

The Insurance Industry's Comfortable Lie

The Conversion Rate Anomaly

What K-Means Actually Found

One Product, Seven Different Purchase Journeys

Archetype Routing System

Technical Implementation

Where We Started and How It Grew

Across the Platform

What Surprised Us

The highest-converting segment was the most underserved

K-means found segments business intuition missed entirely

Confidence thresholds matter as much as model accuracy

Where We Stumbled

Where This Goes