Human-in-the-loop is not a disclaimer, it’s an architecture: the AI gets its own role in the permission system with read-and-propose rights and nothing else, every consequential action waits for a human click, and every decision lands in an audit log. This page shows the actual mechanics, from systems we run in production.
The failed AI project has a recognizable shape: full autonomy on day one, a bad week two, and a quiet shutdown by month three. Production systems climb instead. Each workflow starts at read-and-propose, earns its way up with evidence, and some workflows correctly never leave rung two — money movement being the obvious one.
The loop needs a place for the human, and that place is a queue: everything the AI proposes, waiting with full context and a one-click decision. Done right, reviewing is faster than doing the work ever was, and the queue doubles as training data — every correction teaches the system your operation’s rules.
Telling a model “never send payments” is a wish. Making the AI a database role that cannotsend payments is engineering. In our systems the agent is a constrained principal: its credentials can read and propose, the approval step runs under the human’s credentials, and an append-only audit log records both. The same discipline applies to AI-generated code headed for production — which is why this page and our production-readiness practice rhyme.
Operators trust a system they can overrule, so they actually use it. Regulators and enterprise buyers accept a system with an audit trail, so it survives review. The model improves on your corrections, so the queue gets quieter every month. Rung-two systems compound; rung-four moonshots get unplugged. That’s the whole argument, and it’s why every example in this cluster, from Tandem’s intake to DocuPaint’s inspections, is built this way.