AI in auto glass claims: how we saved a TPA network $4.1M a year

Case study · Auto glass · Latin America

AI auto claims look simple on the surface, especially for glass. A rock hits a windshield, the driver files a claim, and someone decides whether to repair it or replace it. But that one decision – repeated across hundreds of thousands of claims – is where insurers quietly lose millions of dollars a year.

This case study breaks down how an AI auto glass claims program, acting as third-party administrator (TPA) for an insurer’s entire glass portfolio, turned that single decision point into a $4.1M annual savings opportunity – without sacrificing accuracy or customer experience.

The results at a glance: 300,000+ jobs processed, an 8.1% lower average claim size, and roughly 1,110 tonnes of CO₂ avoided every year simply by repairing more glass instead of replacing it.

The Problem: Left Alone, Networks Always Choose “Replace”

Replacement pays repair partners more than a repair does. So when nobody was independently verifying the call, partners defaulted to replacing glass – even when a repair was clearly justified.

The numbers show why this matters so much:

• Repair cost: ~$26
• Replace cost: ~$218
• That’s roughly an 8× cost difference for every avoidable replacement

Multiplied across a national glass claims portfolio, that bias adds up fast — and the insurer had no independent control over a decision being made on its behalf.

Why Human Review Couldn’t Fix It Alone

Manual claim review doesn’t scale, even in low-labor-cost markets. Every claim still needed a human to look at a photo, during business hours only, which meant:

• 87% human accuracy – and only when the photo was good; quality dropped fast on poor images
• 34% of photos were unusable – blur, wrong angle, or recycled/manipulated images
• 5–10 minutes per review, business hours only, with queues building at peak
• 2× inconsistency – tired reviewers disagreeing on the same photo

The bottleneck wasn’t effort. It was that repair-or-replace decisions needed to be instant, consistent, and available 24/7, something a manual process structurally can’t deliver at scale.

The Solution: Computer Vision Built for Reflective Glass

The fix combined three things: a new repair standard, purpose-built computer vision, and a guided capture process that gets a usable photo from the customer the first time.

A wider repair standard. A new guideline qualified crack repairs up to 10cm — up from the industry’s more conservative default, backed by CV models trained to measure true crack size from a single phone photo.

Three possible calls. The models read chips and cracks and return one of three decisions: no action, repair, or replace. Cosmetic damage and dirt are filtered out automatically, so the model’s attention stays on the cracks that actually drive replacement decisions.

Guided, patent-pending capture. Glass is uniquely hard to photograph — it reflects light and lets you see straight through it, hiding damage in both. So the capture flow guides the customer shot by shot: it prompts the right pose and angle, checks each frame live for quality and authenticity, flags reflections or blur on the spot for an instant retake, and marks every chip and crack directly on the glass before a decision is ever made.

How a Claim Flows Through the System

1. FNOL & triage: A questionnaire flags each claim as repair or replace at first notice of loss.
2. AI inspection:Replace-flagged claims route into the AI pipeline; the motorist self-captures photos in a browser, no app required.
3. The decision: The model validates photo quality live, measures the damage, and returns a decision against the repair guideline.
4. Steer to partner :The job is routed to the right repair partner with the decision already made.

Self-service capture alone lifted photo quality, cut fraud, and surfaced a critical insight: 8% of “replace” cases were actually repairable.

---

AI Auto Glass Claims Results: $4.1M in Annual Savings

Here’s where the AI auto glass claims model paid off in hard numbers, across 300,000 jobs in the program:

• $4.1M net annual savings across 300k jobs
• 8% of “replace” cases were actually repairable
• 8.1% lower average claim size
• 11 percentage points of headroom to lift the network’s repair rate
• 93% AI decision accuracy after tuning
• 82% inspection completion rate in self-service

The math behind the headline number is simple: 300,000 jobs × 8% repairable × ($218 − $26 cost gap) ≈ $4.6M gross savings. Net of the program’s own operating cost, that lands at roughly $4.1M per year — plus an 8.1% reduction in average claim size across the entire book.

The Models Got Sharper, Fast

Within two weeks of go-live, decision accuracy climbed and the mistake rate more than halved:

• Correct AI decisions: 83% → 93% by iteration 2
• Replace-decision quality: 98% precision, 97% recall
• Mistake rate: dropped from 12.4% to 5.5%
• Catching repairable glass: recall improved from 38% at launch to 70% after tuning

That last number matters most. Recall on repairable glass is the lever that lets a network safely repair more claims – it’s not about cutting corners, it’s about not missing legitimate repair opportunities that a rushed human reviewer would.

More Repairs, By Design – Not “Cheaper at Any Cost”

A fair question for any AI claims tool: is it just finding cheaper outcomes, or is it finding correct ones? The data says the latter. The AI consistently found more repairable glass than partners identified unaided — more accurately, faster, and without the fatigue that causes human reviewers to drift.

That freed human analysts to focus on the genuinely complex claims, while the model established 11 percentage points of headroom to lift the network’s overall repair rate from 8.7% toward an 11% target.

---

Built to Fit Any Carrier or Market

An AI auto glass claims platform only works at scale if it can flex to each market and each carrier’s rules. This one is configurable end-to-end, like Automotive Glass Experts considers to implement globally:

• VIN and number-plate detection to identify the right vehicle straight from the capture
• Localized languages and workflows for every market and partner process
• Configurable image validations that catch fraud, including AI-generated fake images
• Eurocodes and NAGS parts matching for additional savings on the correct glass part
• Automatic ADAS recalibration flags so no safety step — or unnecessary charge — gets missed
• White-label, no-app deployment, running in the browser under the carrier’s or partner’s own brand

Multi-tenant account controls let a single TPA platform run many carrier and fleet accounts, each with its own repairability profile — edge and corner safety margins, ADAS field-of-view protection, maximum repairable damages, and more. Change a rule and it applies to every new claim instantly, with no model retraining and no downtime.

---

The Bottom Line

Glass can account for up to a quarter of total motor claims – and it’s often where profitability quietly leaks out of an auto insurance book. An AI auto glass claims system hands the carrier independent control over that entire portfolio: the right decision, the right part, and the right calibration, every time, instead of leaving it to a network with a financial incentive to replace.

With fraud risk rising (fabricated and AI-generated images) and parts and labor costs climbing as windshields get more complex, that independent control is only becoming more valuable.

---

Frequently Asked Questions

How much can AI save an insurer on auto glass claims? In this case study, AI-driven repair-or-replace decisions saved $4.1M net per year across 300,000 jobs, driven mainly by correctly identifying repairable cracks that would otherwise have been replaced.

How accurate is AI at deciding repair vs. replace? The model reached 93% overall decision accuracy after tuning, with 98% precision and 97% recall on replace decisions specifically.

Does AI auto glass inspection require an app? No. Self-service capture ran entirely in the browser, with no app download required for the motorist.

Can the repair guideline be customized per carrier? Yes. Each carrier or fleet account can run its own repairability profile — crack size limits, safety margins, and ADAS zones — and changes apply instantly without retraining the model.

---

See What This Looks Like for Your Portfolio

Repair or replace in about a minute, at 95% accuracy, with no app required. Book a demo to see what AI-driven repair-or-replace decisions could do for your auto glass claims portfolio.