In brief
If you had asked a room of investors five years ago which industry would resist software longer — law firms or auto repair shops — most would have said law. Billable hours, partners in their sixties, a profession that still faxes things.
If you had asked a room of investors five years ago which industry would resist software longer — law firms or auto repair shops — most would have said law. Billable hours, partners in their sixties, a profession that still faxes things.
In 2025, legal tech took $3.3 billion across 97 deals. Two of the fifteen largest vertical rounds of the year were legal: Clio at $850 million and Filevine at $400 million.
Auto repair, meanwhile, still runs on whiteboards.
Both industries are conservative. Both have owners who've been burned by software before. The difference in outcome isn't culture, and understanding what it actually is turns out to be useful for anyone building in an undigitised vertical.
The difference is where the work happens
A lawyer's work product is a document. It's created at a keyboard, stored as a file, and shared as a file. When you build software for lawyers, the work and the record already live in the same place — you're improving something that's natively digital.
A mechanic's work product is a repaired car. It's created in a bay, with hands, on a physical object. The record is a separate artefact that somebody has to make afterwards, as an extra task, in a different place, using different tools.
That gap is the entire problem, and it's structural rather than attitudinal.
Concretely: a mechanic finds the customer also needs a tie rod end. Recording it means stopping, cleaning his hands, walking to a terminal, finding the work order, typing. The repair takes four minutes. The record takes six.
He doesn't make the record. Not out of resistance — because the system asked him to interrupt the work he's paid for in order to document it, and on a busy Friday the work wins.
No amount of interface improvement fixes a ratio like that. You can get the six minutes down to three and you've still lost.
Why partial data is worse than none
What you end up with isn't an empty system. It's a partially filled one, and that's the more dangerous outcome.
A work order with no photo might mean nothing was found, or might mean nobody photographed it. A car with no arrival time might not have arrived, or might have arrived unlogged. You can't tell which absences are real.
A dashboard showing seven cars in progress when eleven are in the shop is worse than no dashboard, because now decisions get made on numbers that feel precise.
In law, this failure mode barely exists. The document either got written or it didn't; there's no version where the work happened and the record silently didn't.
What the legal wave actually proved
The interesting part of the legal tech story isn't that lawyers adopted software. It's what they adopted.
Clio and Filevine didn't win by being generic practice management with better design. They won by encoding how legal work specifically happens — matters, intake, deadlines, trust accounting, the particular way a case moves — deep enough that a horizontal alternative isn't a substitute.
That's the lesson that generalises. Conservative industries don't resist software. They resist software that doesn't fit their work, which is most of it, because most of it was designed by people who've never done the work.
Once something fits, adoption isn't slow at all. Two years and $3.3 billion.
So what fits auto repair
If the constraint is the gap between doing and recording, then the answer isn't a better form. It's removing the moment where a human has to choose to record.
That means starting from the physical event:
The car enters the bay. A fixed camera reads the plate. The system matches or creates the vehicle and customer and opens a work order. Nobody typed, and nobody had the option not to.
The mechanic photographs the defect and dictates one sentence — from the phone already in his pocket, next to the car, with dirty hands. No walk, no keyboard. Photo and transcript attach to the order that already exists.
The customer sees the photo of his own part, the line item, the price, and approves or defers that line. The system stores what he approved and when.
That's the architecture we built at MECH Orbit, and I built it because I own repair shops and had already bought and abandoned the alternatives.
The limit, stated plainly: the camera knows the car stood in a bay for four hours. It doesn't know why — waiting for a part, waiting for the customer, mechanic on another job, all look identical to it. Overselling that is how vendors lose the room.
Why this changes the sequencing
There's a strategic implication worth naming.
In a natively digital industry, you can start with intelligence — put a model on top of the existing system of record and sell insight. The data is already there and reasonably complete.
In a physical industry, that sequence fails. The data is partial for structural reasons, so a model built on it produces confident answers about a dataset with unknown holes. You have to fix the source first, which is slower and demos worse, and only then does the intelligence layer have something true to reason about.
That's the part I'd want a fund looking at physical verticals to be sceptical about. Not whether the AI is good. Whether the data underneath it was generated by an event or by someone remembering to type.
The window
Two costs collapsed recently and made the event-first approach practical for small businesses: a usable IP camera is about fifty euros per bay, and processing moved to a data centre rather than a server in the back room. Neither of those was true five years ago, which is a large part of why nobody built this before.
Law firms got their turn because someone finally built for how legal work actually happens. Repair shops are the same problem with a harder capture layer — and, unlike law, an industry where the top fifty companies control under 10% of revenue and nobody has claimed the position.
Sources
What to check in your workshop
- Verify the primary source and date behind every figure
- Separate reported facts from the author's interpretation
- Compare the conclusion with data from your own workshop