There's a moment in every diagnostic where someone asks, half-worried: "So the AI just... sends things? On its own?"
Almost never. And the reason isn't caution theatre — it's that the best-performing systems we've shipped all share the same shape: the machine does the keystrokes, the human makes the call.
Take a quote generator. Reading the enquiry, pulling the rates, applying the surcharges, formatting the document — that's typing. Forty minutes of it. Deciding whether this particular customer gets a discount because they've had a rough quarter and you want to keep them — that's judgement. It takes four seconds and it's worth more than everything else in the quote.
A system that automates the typing returns the forty minutes and leaves the four seconds exactly where it belongs.
This design rule has a practical payoff beyond the philosophy. Draft-for-review systems can ship weeks earlier than fully autonomous ones, because the cost of an imperfect output is a human edit, not an angry customer. The review step is the safety net that lets the system go live before it's perfect — and going live is how it gets perfect, because every correction teaches us where the edges are.
Then something interesting happens. After a month or two of reviewing drafts, clients stop reading most of them. Not because they've gotten lazy — because the drafts have earned it. The review narrows to the flagged cases, the weird ones, the judgement calls. Which is exactly the work the human should have been doing all along.
We could pitch "fully autonomous from day one." It would demo better. But the businesses we build for don't need a demo — they need Tuesday afternoon back, without lying awake wondering what the machine sent while they weren't looking.