There is a service manager in this industry who used to walk his shop floor with a syringe in his pocket. Not for anything medical. For radiator hoses.

One puncture. One slow leak. One covered claim. One more repair order that month.

I did not read that in a report. One our team members at PCMI lived it, across 15 years running claims for a large administrator, and he told me the story the way you would tell somebody what you had for lunch.

Nobody caught him. And the reasons nobody caught him are the most useful way I know to explain what AI can and cannot do for claims right now.

There are three.

No rule could catch him. A rule only knows what somebody thought to write down, and nobody ever wrote down “check whether the service manager is puncturing hoses.” That is not a failure of rules. That is what a rule is. It is also why, in 2026, the fraud control most of us run is a filter that says “dealer name does not contain …” followed by a list somebody maintains by hand.

No model could catch him either. Show a model a photograph of that hose and it starts at the point of failure and completes the picture. Hoses fail. This is what a failed hose looks like. Consistent with wear. A puncture is not part of a typical scene, so it does not get seen.

And no system could catch him, because every system was looking at four claims with that shop. Four claims is not a pattern. It needed 400.

Rules. Models. Scale.

This Is a Job Assignment Problem

I should say up front that I sell this. I am the chief revenue officer of a software company that sells AI into this industry, so weigh everything that follows accordingly.

We talk about AI in claims as though it were one thing you either adopt or you do not. It is not. There are two tools on the table, rules and models, and they are good at completely different jobs.

Nearly every AI failure I have seen, including ours, came from handing a job to a model that belonged to a rule, or the reverse.

What Belongs to Rules

Coverage. Whether this component, on this contract, sold on this date, at this mileage, is covered. Contract terms. Deductibles, limits, exclusions, aggregate caps. Money movement. Compliance math.

None of that is simple. Cancellation math alone has ruined more weekends than anything else in our business. But those decisions share three properties.

Being wrong is expensive rather than embarrassing. It shows up in a reserve, a refund or a lawsuit.

Eighteen months from now, somebody who was not in the room will ask you why. A regulator, an underwriter, a reinsurer, an arbitrator. “The system determined” does not survive that conversation. I have watched people try.

And consistency matters more than cleverness. If the same claim arrives twice, from two dealers, on two different days, you need the same answer both times. Not a defensible answer. The same answer. A model that is right 95% of the time and slightly different on Thursday has handed you a fair treatment problem you will hear about later.

Somewhere in the last three years, “rules-based” turned into an insult. But a rule is a decision somebody already made, on purpose, with their name on it, that now happens the same way 10,000 times. In a regulated business, that is the point, and it is the floor that makes everything above it safe to deploy.

What Models Are Genuinely Good At

Reading a contract to find the coverage language. Pulling claim details out of an email, a voicemail or an IVR call. Reading a repair estimate and creating the detail lines. Reviewing photographs and inspection reports. Assembling evidence so that when an adjuster opens a file, the contract, the vehicle history and the prior claims from that shop are already sitting there. Flagging when complaint, cause and cure do not line up. Clearing the low dollar tail nobody wanted to touch anyway.

Read that list again and notice how boring it is. There is no robot adjuster on it. Every item is gathering.

That is the honest state of the art, and it is worth real money. An adjuster who used to spend 20 minutes assembling and two minutes deciding now spends the 20 minutes on the claims that actually need a person.

The Part Almost Nobody Check First

Ask a room of administrators how much they trust the in-service date on a contract and watch them laugh. For 20 years we have all known that field is whatever the dealer typed that afternoon.

Your best adjusters know it too. They have been discounting that field 100 times a day for so long that they no longer know they are doing it. It is not in a procedure. It is not written down anywhere. It is just something they know.

The model does not know it.

Every company reading this has four or five fields like that: fields your people learned to work around and never documented. Automation does not inherit the workaround. It inherits the field. What you are changing is not the data quality. It is the person standing in front of it.

Careful Language, and What Happens to It

The most useful thing I learned this year came out of a photo analysis run on a real vehicle.

The tool found aftermarket intake components in the engine bay, cited the specific photograph, and explained why the hardware was not stock. That is a genuine catch. Nine photos, 90 seconds, an adjuster already behind. The model does not get tired on the seventh photo.

The second finding was a lift kit check. No visible evidence of a raised ride height. And then the tool wrote this: “The vehicle appears at normal ride height in the provided images.”

“In the provided images” means our own tool hedged, correctly, and told you exactly what it was hedging on.

Now watch what happens to that sentence. It goes into the file. Three steps later somebody reads “no lift kit.” Not “we could not see one in nine photographs.” No lift kit.

A colleague sent me a line from a podcast I have not been able to put down since: “Absence of evidence gets consumed downstream as evidence of absence.”

Which is why these tools need three answers available to them, not two. Present. Not present. And cannot determine. And cannot determine should be the most common answer the tool gives.

I am aware how badly that sells. Nobody puts “our AI frequently declines to answer” on a booth banner. But a tool that cannot say it is giving you confidence, not accuracy. Those are different products at similar prices.

Three Questions to Ask, Including of Me

If you take nothing else from this, take these:

1. What does it say when it does not know? If the only answers are yes and no, you are buying confidence. Ask to see “cannot determine” in the product rather than on the roadmap, then ask how often it fires. If that number is small, somebody tuned it to look decisive in a demo.
2. Who writes the decision? Does the tool write into the system of record or hand a recommendation to a person who retypes it? Advice you have to rekey is not automation. It is an expensive second opinion. Ask to see the audit trail, then count how many systems somebody had to open to assemble it.
3. How big is the sample it learned from? Your four claims with a shop look like a fluke, and they are. Somebody else has 400 with the same shop. Most vendors will tell you, correctly, that your data is segregated and never shared. That is the right promise to demand. It also means the sample is capped permanently at the size of your own book.

What This Means for Your Dealers

Your dealers are going to ask you two things this year: Whether this eliminates jobs, and whether they should buy something.

On the first: the gathering goes away, the deciding does not. Straight-through processing is real, it is here today and getting better every day, but a claim that runs start to finish is not a claim where nobody decided. It is one where somebody decided in advance and signed their name to it. Zero touch is the goal. Zero judgment is not.

On the second: before you buy a model, tell your team to go fix four fields. They will know which four. Ask your adjusters which numbers they do not believe, and they will have an answer in 90 seconds and be surprised anybody asked.

The winners over the next three years will not be the ones with the flashiest AI. They will be the ones who know exactly which job they are handing to which tool. Right now, the honest list of what belongs to a model is shorter than the list being sold. Anybody who tells you different is selling.

Branson Smith is CRO of PCMI. This article is an adaption of his opening keynote at the inaugural 2026 AI Dealer Agent Conference in Atlanta.