Human-in-the-loop: the only AI automation pattern we trust with money

There is exactly one pattern we will put near a client's money, and it is the one with a human standing in the loop. Everything else is a demo waiting to become an incident.

It comes down to three controls. An approval gate, so anything that moves money or reaches a customer waits for a person to say yes. An action log, so every decision the agent made is recorded with its reasoning and is auditable after the fact. And a hard cap, so the worst possible day is bounded in advance — a number you chose, not a number you discover.

Done this way, the agent drafts and proposes; the human approves and ships. The volume still collapses — a person reviewing 28 prepared decisions is doing a fraction of the work of making 28 decisions from scratch — but the judgement, and the accountability, stay where they belong.

We sleep fine because the system is built to make a runaway impossible rather than unlikely. That is the difference between automation you can run a business on and automation you can put in a pitch.