Collision repair has fraud. Some is intentional—staged damage, inflated estimates, phantom repairs. Some is accidental—bookkeeping errors, miscoded supplements, operational sloppiness that looks like fraud from the outside.
AI fraud detection is one of the most pitched features at SEMA right now. Some of the pitches are real. Most are exaggerated. Here's what AI can actually catch, what it can't, and how to deploy it without creating more problems than you solve.
What "Fraud Detection" Actually Means in Practice
Fraud detection does not mean AI identifies fraud and reports it. Every production deployment we've seen uses AI to prioritize which ROs a human should review. The human makes the determination.
This distinction matters. An AI that makes a fraud determination autonomously is a legal liability. An AI that ranks 2,000 ROs so your compliance team reviews the top 20 is a productivity tool.
What the Signals Actually Are
Pattern-Based Signals
Most of fraud detection isn't AI in the modern sense. It's pattern matching on features that experienced estimators and compliance reviewers already know to look for:
- Estimate-to-actual variance. The first estimate said $4,500. The final was $9,200. Not necessarily fraud, but worth a look.
- Supplement frequency by employee. One estimator's work generates supplements at 2x the rate of peers. Could be skill, could be gaming.
- Parts price anomalies. OE parts billed at significantly above or below market prices.
- Labor time vs industry standard. Flag hours exceeding industry reference times for a given repair operation by a significant margin.
- Repeat customer/plate patterns. The same vehicle plate appearing across multiple claims in a short window.
- Total loss threshold adjacency. Estimates landing just below the total-loss threshold repeatedly.
- Tear-down-to-supplement ratio. Tear-down teams with supplement rates that don't match their volume.
None of these require deep learning. They require a warehouse with the right data and the right queries.
Machine Learning Signals
Real ML in fraud detection adds value in a few places:
- Multivariate anomaly detection. Combinations of features that together are unusual, even if each individually is normal.
- Photo analysis. Damage in photos that doesn't match damage described in the estimate.
- Text analysis on estimator notes. Language patterns correlated with historical fraud.
- Network analysis. Relationships between claimants, providers, and shops that suggest organized activity.
These are real AI applications, but they're most valuable as additional signals on top of the pattern-based baseline—not replacements for it.
What AI Can't Do
Make the Fraud Determination
This is the headline. AI cannot reliably tell you "this RO is fraud." It can tell you "this RO has an unusual pattern worth reviewing." The determination is legal and consequential; a human makes it.
Every vendor claim that AI "catches fraud" should be pressure-tested with: "Catches in the sense of flags, or catches in the sense of confirms?" The answer is always flags.
Avoid False Positives
False positive rates on fraud models are high. Often 5–20% of flagged ROs turn out to be legitimate outliers—a complicated repair, a new estimator still learning, an unusual vehicle model.
If your compliance team doesn't have the bandwidth to review the flagged set, automated fraud detection doesn't help. It just produces a list that sits unread.
Scale to High-Confidence Action
Autonomous action on fraud signals—auto-denial, auto-escalation, auto-reporting—is not a safe deployment pattern. The reputational and legal downside of a wrong accusation outweighs the efficiency gain. AI ranks; humans decide; actions follow.
Handle Bias Responsibly
Fraud models can inadvertently proxy for demographic variables through geography, claim patterns, or claimant characteristics. A model that over-flags ROs from certain neighborhoods or certain customer profiles is a problem even if it's not explicitly using those variables.
If you deploy fraud detection, audit what the model is actually flagging. If a pattern correlates with protected characteristics, redesign or exclude the feature.
What a Deployment Actually Looks Like
Data Layer
Your warehouse has the estimate records, supplement history, parts data, labor data, photos (if digitized), estimator notes, and customer/claimant data. Most of what you need is in CCC's exports.
Scoring Layer
A nightly batch job computes fraud scores across ROs. Scores are composite: pattern-match signals + ML anomaly scores. Output is a ranked list.
Review Workflow
Top-ranked ROs go to a compliance queue. Reviewers see the RO, the signals that flagged it, and the comparables (what normal looks like). They triage: legitimate, review further, escalate to SIU (special investigations unit), or close.
Feedback Loop
Reviewer decisions feed back to the model. False positives are labeled. The model learns—or at least the features get tuned.
Realistic Expectations
What a good fraud detection deployment produces:
- Queue of 10–30 ROs per week for a 20-shop MSO to review. Enough signal to justify the review time, not so much that it overwhelms the team.
- Confirmed fraud rate of 5–20% of flagged items. The rest are legitimate outliers or "look funny but fine."
- Early detection of organized patterns—same claimant across shops, suspicious repeat customers, clusters of similar ROs.
- Deterrent effect on internal bad actors who know their work is being scanned.
What it doesn't produce:
- A number you can report to leadership as "fraud detected this quarter."
- Autonomous action against suspicious ROs.
- A replacement for your SIU or legal function.
Whether It's Worth Doing
Fraud detection is worth deploying if:
- You have a compliance or SIU function that can triage a weekly queue.
- Your RO volume is high enough that patterns exist to find (generally 50+ shops).
- Your warehouse is clean enough that the model's inputs are trustworthy.
It's not worth deploying if:
- You don't have anyone to review the output.
- Your data is flaky and the model will flag on bad joins rather than real patterns.
- You're deploying it because it was in the vendor's pitch, not because you have a fraud problem worth the investment.
The Honest Bottom Line
Fraud detection is one of the more hype-adjacent AI features in collision. It's real, it works, it produces value—but only in the context of a human-in-the-loop review process, deployed against clean data, with realistic expectations.
Most of the value is in pattern matching that doesn't require AI in the modern sense. Some of the value is in ML on top. Very little of the value is in the autonomous-AI fantasy that lives in vendor decks.