Autonomy, yes, but the decisions that matter pass through a person.
Autonomy is not the same as "unsupervised". Our agents do the repetitive load end to end, but you decide which actions run automatically and which need a human sign-off. Nothing ships without your control.
Every department raises this by the second meeting, and it is rarely a question about technology: not whether the agent can perform an action, but who answers for it once it does. Human-in-the-loop is the answer to that question, not a safety feature bolted on at the end.
Guardrails, not brakes
Approvals, audit trails and clear citations mean autonomy never becomes a loss of control. The agent moves fast where it's safe, and asks where it isn't.
The difference matters. A brake slows everything down equally; a guardrail lets you run where the road is straight and only intervenes on the bend. An agent that asks for confirmation at every step isn't being careful, it's being useless: it hands the person back the same load it was meant to lift.
Where the line falls
We don't draw that line in the abstract. It emerges while mapping the department, by looking at which decisions have reversible consequences and which don't. It usually settles like this:
- Runs automatically when the action is reversible and checkable: reading, extracting, classifying, reconciling, preparing a draft, updating an internal record.
- Goes through a person when it is irreversible or exposed: anything that leaves the company, anything that commits it, anything touching a customer or a contractual position.
- Stays human whenever the case is anomalous and needs judgment, which is precisely the work worth leaving to people.
The threshold moves over time, and it has to. A task that needs sign-off on every case in month one can move to sampled review after three months of verifiable history. The direction of travel is towards more autonomy, but earned on evidence rather than granted up front.
The audit trail comes before the autonomy
Every action the agent takes leaves a record of what it did, which sources it relied on and who approved it. This isn't a compliance box to tick afterwards: it is the condition that makes widening autonomy possible at all. Without verifiable history, every expansion would be an act of faith.
Approval must not become the queue
The most common way to get human-in-the-loop wrong is to turn it into a bottleneck: the agent finishes in seconds, then the result waits two days in an inbox nobody owns. Every minute saved is lost right there.
So we design the approval step with the same rigour as the rest: batched rather than one request at a time, inside the tool the department already uses rather than in one more interface, and carrying enough context — the source, the citation, the reason — that the decision takes seconds instead of a fresh investigation.
When it works, the person stops executing the work and starts governing it. That is the point: not an agent under surveillance, but a department deciding more while handling less.