The Human Gate

The pattern we install in every workflow that touches the outside world is the same, and it is simple enough to state in one sentence: the machine proposes, a person disposes.

In practice: research, enrichment, and drafting are carried by systems — they are structure-shaped work, and machines are better at structure than people are. But anything that leaves the building waits in a queue for a human decision. Outreach sits as drafts. Documents sit as drafts. The send button belongs to a person with a name.

Why not autonomy

The argument for full autonomy is throughput. The argument against it is that throughput is not the constraint — trust is. A market that receives machine-sent outreach learns to delete it, and markets talk to each other. The volume you gain is borrowed from a reputation account that refills slowly.

There is also an epistemic reason. The gate is where the system learns. Every draft a person edits before releasing is a labeled example of the difference between almost-right and right — and that difference is precisely the judgment the firm is paid for. Remove the gate and the system stops learning at exactly its current level of taste.

The framework

When we run workflow surgery, the classification step is the whole game. For each step in the workflow, ask: is this search, structure, or drafting — or is it commitment? Commitment means the step changes something in the world that is expensive to unwind: a message sent, money moved, a representation made. Machine-shaped steps get automated aggressively. Commitment steps get a gate, an owner, and an audit trail.

The failure mode to watch for is gate decay: a queue that a person approves by reflex is a gate in name only. We measure pass rates. A gate approving everything untouched either means the drafting system has gotten genuinely good — verify with sampling — or the human has stopped reading. One of those is a milestone; the other is an incident report waiting for its date.