AI you keep, not AI you clap for.

There’s a version of AI built to be clapped for. It demos beautifully, makes the right noises in a pitch, and quietly falls apart the first time it meets a real workflow and a user who can’t afford to be wrong. We’ve built enough software to be allergic to it.

The two products in our lab both use AI, and the interesting decisions were about where we refused to. In VANTA — a system people hand their books and their tax filings to — the auditor that checks every invoice is not an LLM. It’s a deterministic engine: given the same numbers, it returns the same verdict every time, with a reason you can point to. A model that’s right ninety-something percent of the time is a wonderful demo and a catastrophe in a system whose whole promise is that the number is correct. So the load-bearing part stays boring on purpose, and the model waits for the job it’s actually good at — explaining, drafting, narrating — where being occasionally wrong costs nothing.

AI belongs where a mistake is cheap and a human is still in the loop. Keep it out of the place where “usually right” isn’t good enough.

Tawrio is the other side of the same coin. There, AI isn’t a chat box stapled to a form — it does the tedious labour a procurement team dreads: it reads a plain-English request and turns it into proper quote requests, chases suppliers, lines up the responses so they can actually be compared, matches invoices back to orders. Real work, removed. And around all of it, an audit trail — because a decision you can’t explain later isn’t one you can stand behind.

The unglamorous bar

None of this is anti-AI — we build with it every day, and it’s changing what a small team can take on. It’s about a bar. A feature earns its place when it still works on the boring Tuesday, after the demo energy is gone and someone just needs the thing to be right. That’s the AI we want to be known for: not the kind you clap for — the kind you keep.