There is a standard arc in AI product marketing. Version one has a human approving everything. Version two has a human approving the exceptions. Version three is "fully autonomous," and the humans are gone. Human review is presented as scaffolding — necessary while the model is immature, removed once it grows up.

We reject that arc, and not primarily for safety reasons. We reject it because the arc misunderstands what customers are buying.

What the approval gate actually protects

When a system drafts a customer message, a scope of work, a claim submission or a payment exception and holds it for approval, three things are protected at once.

The customer relationship. An autonomous system that sends a slightly wrong message to a long-standing client does damage that no efficiency gain repays. The person approving is not checking grammar; they are applying context the system cannot have — that this client is already unhappy, that this quote follows a difficult conversation, that this is the third time we have rescheduled.

Accountability. Someone has to be answerable. In a regulated context this is explicit; everywhere else it is implicit and just as real. An approval gate creates a named human who released the action, which is the difference between a defensible process and an incident report.

The feedback signal. This is the one people miss. Every edit a human makes before approving is a labeled training example, produced for free, in the exact context where the system was wrong. Remove the gate and you remove the highest-quality improvement signal the system has.

The economics are better than they look

The usual objection is throughput: a human gate caps how fast the system can run. In practice this is true and mostly irrelevant, because the bottleneck was never the approval — it was the drafting.

Consider proposal generation. Writing the proposal takes ninety minutes. Reviewing and approving one takes four. Moving from "human writes" to "system drafts, human approves" converts a ninety-minute task into a four-minute one — a 22× improvement. Removing the human afterward converts a four-minute task into a zero-minute one, which is another 1.05× on the original. You capture nearly all the value at the first step and take on nearly all the risk at the second.

The first move captures 95% of the gain. The second move captures 5% of the gain and 100% of the risk.

Where autonomy is appropriate

This is not an argument against automation. Plenty of actions genuinely should run without a human: reading, sorting, classifying, enriching, monitoring, flagging, and any reversible internal operation. Phantom monitors continuously and autonomously; what it will not do is release an outbound action without a gate.

The line we draw is not "risky versus safe." It is outward-facing and hard to reverse versus internal and reversible. Anything that reaches a customer, moves money, or creates a record someone else will rely on gets a gate. Everything else runs free.

What to ask a vendor

If you are evaluating AI systems, three questions separate real governance from a checkbox.

Where exactly is the gate, and can I see the queue? If approval is buried in a settings page and defaults to off, it is decoration.

What does the system record about who approved what? "Provenance" should mean a durable record naming the model, the inputs, the output and the approver — retrievable a year later.

What happens to my edits? If corrections vanish instead of improving the system for your account specifically, you are doing free labeling for someone else's roadmap.

The gate is not the training wheels. It is the steering.