The approval card is the product.

People ask what the hard part of an AI Operating System is. They expect the answer to be memory, or reliability, or getting the thing to write in the organisation's voice. Those are hard. But the part that decides whether the system gets to exist at all is a message on a phone, four lines long.

To: the grants officer at a trust. What: send the Q2 impact report. Why: their deadline is Friday, and you approved the draft on Tuesday. If you approve: it sends within five minutes. Two buttons that never swap places. And underneath, in small print, the most important sentence we've written: if this expires, nothing sends.

Every design decision in that card is doing safety work. The facts — recipient, attachment, amount — are drawn from the action the system is actually about to take, not from prose that could dress it up. The buttons keep a fixed order so muscle memory can't betray you. Ignoring the card is always safe, because the default is always no. Denying never demands a justification, though the system learns from one if you offer it.

Here's the inversion that took us longest to see: approval isn't the tax you pay for automation. It's the reason you can have automation. The organisations that should buy systems like this are precisely the ones that answer to boards, trustees, regulators. They don't need a system that acts more; they need one that can show its working and wait for a yes. The card is that, made physical.

It asks first. Everything else is engineering.