Nobody dreams of building software for workplace safety compliance. It doesn’t demo well. There’s no screenshot of it that gets shared around. When I tell people I co-founded a company in this space, the reaction is polite confusion: why would someone who builds AI systems spend his time on paperwork?
Here’s why: paperwork is never really the problem. It’s the symptom everyone agrees to complain about instead of asking what’s actually broken underneath it.
BTF, Bureau of Technical Safety, exists inside a market most people would call boring on purpose: workplace safety, food safety, regulatory compliance for small and mid-size businesses. The incumbents are consultancies that have run the same way for decades: a consultant visits a client, takes notes, goes back to the office, and manually assembles the documents the law requires. Every deadline is tracked by memory, a spreadsheet, or a wall calendar. Every document is drafted almost from scratch, because the last one lives in someone’s inbox, not in a system.
The popular explanation for why this stays slow is that compliance is inherently bureaucratic: heavy, cautious, resistant to change because the stakes (worker safety, food safety, legal liability) are too high to move fast. I don’t think that’s true. What I found when I actually mapped the workflow is that the slowness has nothing to do with the subject matter and everything to do with information never being structured anywhere. The client’s history lives in one place, the contract terms in another, the deadlines in a third, and the document templates in whoever’s head remembers writing the last one. None of it talks to any of it. That’s not caution. That’s just an unstructured system, and unstructured systems are slow no matter how careful the people running them are.
So we built the thing that was actually missing: one platform that holds the client and contract relationship, tracks every compliance deadline against it, ingests the documents that come in, and generates the documents that go out from that same structured base. What used to take a consultant days or weeks per client (chasing down the last version, drafting from a blank page, cross-checking the deadline calendar by hand) now takes minutes. The model does the assembly. A human still reviews and signs off on every document before it reaches a client, because in a market where the output is a legal safety certification, removing the human from that decision isn’t automation, it’s abdication.
That last part matters more than the speed. The interesting version of AI-native isn’t “replace the person who used to do this.” It’s “give the person a system that removes the repetitive, error-prone assembly work so their judgment is the only thing left for them to spend time on.” The consultant still decides what’s true about a client’s compliance status. The system just stops making them retype it five times to prove it.
What surprised me is how little of this is actually about AI. The hard part was mapping how the work really moved between people before writing a line of code: who touches a document, in what order, what has to be true before the next step can happen, where the current process quietly loses information. The AI is what makes the resulting system fast and cheap to run. It’s not what makes the system correct. That comes from actually understanding the mechanism first.
That’s the pattern I keep finding in industries everyone assumes are already efficient because they’re old, regulated, or unglamorous: the problem was never that the work is hard. It’s that nobody structured it. Boring isn’t the opposite of disruptable. Boring is just where the incumbents stopped looking, and where the next system that actually gets built first wins by a wide margin.