"Human in the loop" is too blunt
Most vendors say a person stays in the loop. The phrase hides the decisions that matter: which loop, how often, and what has to be true before the person can step back. Left vague, it ends one of two ways. Either a person checks everything for ever, and the agent becomes an expensive way to draft. Or someone decides one day that it seems fine, and the business takes on risk nobody signed for.
Our view is simple. An agent should earn autonomy the way a senior hire does: against evidence, signed off by the business, and taken back after a serious mistake.
Three levels of oversight, set for each workflow
Autonomy belongs to a workflow, not to an agent. The same agent can be trusted to chase an overdue quote and still need a person to approve every price.
- Level 1: a person approves. The agent prepares the work, and a person approves every output before it takes effect. Every new workflow starts here.
- Level 2: a person handles the exceptions. The agent acts on its own. Anything that breaks a rule or falls below a confidence threshold goes to a person, and a fixed sample of everything else is checked.
- Level 3: monitored. No single output is reviewed. Oversight comes from a monthly review, alerts on errors and drift, and tracking against the value agreed for the project.
What has to be true before oversight steps down
We publish our thresholds because a buyer should be able to test them. These are our starting values. The method is new, and we'll tune them as real run data builds up.
| Condition |
|---|
| At least 4 weeks and 200 reviewed runs at level 1. |
| Fewer than 5% of outputs changed or rejected by a person. |
| No serious errors. |
| Written exception rules, tested against past runs. |
| At least 70% of planned use actually happening. |
| Condition |
|---|
| At least 8 weeks at level 2. |
| Exceptions on fewer than 2% of runs. |
| Sampled checks at least 98% correct. |
| Value on track: at least 100% of what was expected at that point in the project. |
| Drift alerts and a way to roll back, both live and tested. |
You co-sign every step down
Each move down a level changes the risk your business carries, so it isn't ours to decide alone. Our delivery lead and your process owner both sign it. For our first workflows at level 3, our CEO signs as well.
Oversight steps back up automatically
A workflow moves up one level straight away, and the move is reviewed afterwards, in any of three cases:
- there is a serious error;
- corrections or exceptions run above their threshold for two weeks;
- the value delivered falls far enough behind the agreed plan to trigger a recovery review.
You're told within one business day. It steps down again only by meeting the thresholds and being co-signed, as before. Trust that can't be taken back isn't being managed.
Some decisions never leave a person
No level of evidence moves these to an agent:
- commitments that change scope, price or risk;
- financial changes above agreed limits;
- legal or regulatory exposure;
- anything that can't be undone.
An agent may prepare them; a person decides them.
The same principle governs our own business. Our constitution keeps people accountable for judgement-heavy and high-consequence decisions, and sets out which actions need a person's approval.
What this looks like in practice
In a platform we're building for a design and procurement business, an AI agent drafts follow-ups and questions to suppliers. It doesn't send them: a person reviews and sends every one, which is level 1 in our method. Anything that commits money, price or dates goes to the project manager.
Questions to ask any vendor before an agent acts on its own
- At what level does each workflow start, and who decides when it moves?
- What evidence, in numbers, justifies the move?
- Who in our business signs it?
- What happens automatically after a serious mistake?
- Which decisions will never be handed to an agent, whatever the evidence?
If the answers are "it depends", "we'll know", "the vendor" and "we'll review it", the autonomy hasn't been earned. It has been assumed.
Futuvara builds, implements and runs AI platforms where AI agents take on real work in your business, under rules you set, and earn the right to do more.
