The question every buyer of agentic automation asks is “how autonomous can it be?” That is the wrong question, and the projects that ask it tend to end badly in one of two directions: an agent trusted everywhere before it earned trust anywhere, or a blanket no-autonomy policy that leaves the whole investment doing autocomplete. The right question is smaller and repeats: how autonomous should this agent be, in this workflow, given the evidence we have today?
The approval gate is the design, not a placeholder
Our standing rule: wherever an action moves money, data or risk, a person approves it. Sending a payment, changing a customer record, emailing an external party, granting access — each of these gets an approval gate, and the gate is part of the architecture from day one, not a training-wheels phase to be engineered away once everyone relaxes.
This changes what the agent is for. Its job is to do the mechanical work — gather the documents, reconcile the fields, draft the response, propose the action — and to package that proposal so a human can judge it in seconds rather than minutes. A well-designed gate is not a bottleneck. It is the point where machine throughput and human accountability meet, and it is where the audit trail gets written.
Autonomy expands with evidence, per workflow
Autonomy is earned per workflow, not granted per project. An agent that has proven itself on invoice matching has proven exactly nothing about supplier onboarding. Each workflow gets its own ladder, and the agent climbs it on eval evidence, not on vibes or a steering-committee decree.
The ladder looks roughly like this in practice:
- Draft only — the agent prepares, a person does. Every output is reviewed, and every review becomes eval data.
- Approve each — the agent proposes complete actions; a person approves one at a time. Disagreements are logged and fed back into the golden dataset.
- Approve by exception — routine cases inside well-evidenced bounds proceed; anything novel, ambiguous or high-value is routed to a person. The bounds themselves are defined by where the eval suite shows consistent agreement with human judgement — and they contract again if drift monitoring says quality moved.
Moving up a rung is a decision made on eval results against a golden dataset the domain owners adjudicated — the same discipline we describe in taking RAG from demo to production. Moving down a rung has to be just as easy, which is an argument for building the ladder into the system rather than into a policy document.
An audit trail per action, not per quarter
Auditors ask one question about automation: who approved this action? The system has to answer per action. Every agent step records what was proposed, what evidence it saw, which model and prompt version produced it, who approved or overrode it, and when. That record is boring to build and priceless twice: once when an auditor or regulator asks, and continuously as the highest-quality training signal you will ever get — every human override is a labelled example of where the agent’s judgement ends.
What “human in the loop” concretely means in the back office
The phrase has been abstracted to death, so here is the concrete version. In an accounts-payable workflow it means the agent matches the invoice to the purchase order, flags the discrepancies, drafts the posting — and the controller approves the payment from a queue, with the agent’s evidence attached, instead of assembling that evidence herself. In customer operations it means the agent drafts the reply with citations to the actual policy documents, and the handler sends it. In access management it means the agent prepares the change with its justification, and the owner of the system clicks approve.
Notice what the human is doing in each case: judging, not typing. The mechanical work moved to the machine; the accountability did not move at all. When the agent is unsure, it says so and abstains — an escalation is a success mode, not a failure mode. That is the whole design, and teams that get it right find the approval queue becomes the place where the organisation actually learns what its own judgement looks like.
This is how our Applied AI & Automation practice builds every agentic system: humans gate consequence, autonomy is earned rung by rung, and every action can answer for itself. If you are designing approval gates for an automation of your own, bring the workflow to a tech lead — a technical call, not a sales call.