Every vendor at SEMA has an AI story. Every collision software deck has a "powered by AI" slide. Every consultant pitching your MSO is offering an AI transformation.
Most of it is marketing. Some of it is real.
Here's what AI is actually doing inside collision repair operations in 2026—from deployments we've built and watched—and, more importantly, what it isn't.
What AI Is Actually Doing
A short list of where AI is earning its seat at the table today:
Writing-Up an Estimate Faster
The intake and write-up process is the single biggest time sink in collision operations. Photos, VIN decode, damage assessment, initial parts list, labor estimate—30 to 60 minutes of an estimator's day per vehicle.
AI copilots are now cutting this to 10–20 minutes for a meaningful share of the work. Photos feed a damage detection model, which proposes an initial estimate. The estimator reviews and corrects. The corrections train the model further.
This is working. The productivity gain is real. It doesn't replace the estimator—it eliminates the mechanical parts of their day so they can handle more vehicles.
Summarizing the Night's Data
A GM walking in at 7 a.m. used to need 20 minutes to figure out what happened overnight: which cars shipped, which stalled, which KPIs slipped, which ROs are at risk.
AI-generated ops briefs now do this in a 5-minute read. Pull the data, run the narrative through an LLM, deliver a structured summary to the GM's inbox or Teams channel before they sit down.
This is working. It scales, it's cheap, and GMs actually read it—because the alternative is opening five dashboards in five tabs.
Flagging Anomalies Before Humans Notice
Static thresholds produce alert fatigue ("WIP over 30 days!" every Monday). Statistical anomaly models—even simple ones—catch real deviations without crying wolf every week.
A cycle-time spike at one shop, a severity drop that doesn't match the claim mix, a DRP volume shift that precedes a carrier review—all of these are catchable with techniques that aren't cutting-edge. They're just not built into CCC.
This is working. It requires a warehouse, data discipline, and a modest investment. Payoff is consistent.
Generating First-Draft Board Narratives
Quarterly board packs take a finance team days to assemble. The data is mostly already in the warehouse. The narrative—"Q2 vehicles-delivered was up 8% driven by the Phoenix shops, offset by a 3-point severity decline in the Dallas cluster"—is what takes the time.
LLMs trained on your KPI data can produce first-draft narratives in seconds. A human edits. The review cycle tightens from days to hours.
This is working for PE-backed MSOs doing frequent board reporting. The draft is never final. It's always 70% done, which is where the leverage is.
Forecasting Vehicles Out
Simple forecasting models (ARIMA, Prophet, even a well-tuned regression) can predict 30/60/90-day vehicles-out with usable accuracy. This drives capacity planning, staffing, DRP negotiations, and cash-flow forecasting.
This is working. It's not LLMs, which is part of why no one markets it. But it's AI, it's valuable, and it's boring enough to actually run in production.
What AI Isn't Doing (Despite the Pitches)
Fully Automating the Estimate
A vendor deck will tell you AI writes the full estimate. It doesn't. Not reliably. Not at the accuracy threshold that avoids supplement rework or carrier pushback.
AI can get you 60–80% of the way to an estimate. A competent estimator still reviews, corrects, adds context a model can't see. The value is in the copilot relationship, not replacement.
Replacing Your Technicians
Nobody serious is claiming this. But the ambient noise around "AI transforming collision" sometimes creates the impression. It doesn't. AI is touching the knowledge-work side of the business—estimating, reporting, planning—not the wrench-turning side.
Detecting Fraud at Carrier-Grade Accuracy
AI can flag suspicious ROs for review. It cannot make the fraud determination. The false-positive rate on every fraud model we've seen is too high to automate a decision that has legal consequences. Use it to prioritize what humans review, not to decide outcomes.
Answering Arbitrary Questions About Your Data
Natural language query ("show me cycle time for the Dallas cluster last quarter") works for simple, well-formed questions on a well-modeled warehouse. It breaks on questions that require joins the model can't infer, time windows with fuzzy definitions, or KPIs where your organization's definition differs from the standard.
It's getting better quickly. It's not a replacement for a good BI tool and a data person yet.
Predicting Customer Behavior
Vendors occasionally pitch AI that predicts which customers will leave bad reviews, which will refer friends, or which will dispute charges. The signal on this is almost always weaker than the model's confidence suggests. Deploy skeptically.
What to Actually Ask Before Buying
When a vendor pitches an AI feature:
- What model? What training data? If they can't answer, it's a wrapper on an off-the-shelf API with no moat.
- What's the error rate in production? Not lab accuracy—production.
- What does a human still have to do? If the answer is "nothing," the vendor is lying or the feature isn't real.
- Where does our data go? Is it used to train models that benefit competitors?
- Can we see it running in a shop our size? With real data, doing real work.
- What happens when it's wrong? Who catches it? What's the blast radius?
The Bigger Picture
The common thread in what's working: AI is most valuable where a knowledge worker does something mechanical that a model can draft, leaving them to review and add judgment. Estimating, ops briefs, narratives, anomaly review, forecasting. Copilot, not autopilot.
The common thread in what isn't: AI is overpromised where it replaces judgment rather than drafting. Fraud determination, autonomous estimating, predictive scoring on fuzzy outcomes. These make great slides and bad deployments.
If you're evaluating an AI pitch, ask which bucket it's in. If the vendor can't tell you, they don't know, and the product probably isn't built.