Codeless Ops“AI with receipts” isn’t a tagline we reverse-engineered onto a product. It’s the constraint we started from.
Built the other way round
The AI tools finance teams were being handed couldn’t be trusted with a number. They’d guess. They’d blend versions. They’d never show their work — and they’d do it all with total confidence.
So we built the whole system the other way round. Provenance and honest abstention are the starting point, not an afterthought. Every answer has to:
- Trace back to the exact figure it came from.
- Treat “which version?” as a first-class question.
- Say “I can’t confirm that” when the documents don’t support an answer.
That’s the only way to build AI you’d put in front of a controller.
Abstention is a feature
An AI that tells you “I can’t confirm that” is worth more than one that’s confidently wrong. In finance, a confident wrong answer isn’t a small error — it’s the one that ends up in the filing. Honest abstention is what makes the rest of the system trustworthy.
The same discipline, everywhere
Grounded is where this shows most clearly, but it’s the same standard we bring to any AI or automation work for a finance team: if it touches a number a decision depends on, it gets built to be verified. Nothing ships that can’t show its work.