Every platform has an autonomy dial
An AI agent should have as much autonomy as it has earned trust in that domain, and no more. That autonomy should be enforced by limits on what the agent can touch, not by a person approving each action.
Earlier this year I wrote that the governance model built for AI that talks doesn't work for AI that acts, and that the better analogy was onboarding: you give a new hire a scope of authority and widen it as they earn it. That argument has held up, but what I didn't anticipate is how quickly the industry would turn it into a product feature.
Every agent platform now ships a way to set how much the agent does without asking. In Copilot Studio it is a switch on each tool: require approval, or don't. Relevance AI gives each action three options: auto run, approval required, or let the agent decide. Salesforce and ServiceNow have guardrails, verification gates and escalation thresholds. Above the products sit the ladders: the Cloud Security Alliance published a framework in March 2026 with five levels modelled on the self-driving scale, and the consultancies have their own. None of them say what moves an agent from one rung to the next.
In practice, the answer is the vendor's confidence (or the project sponsor's impatience, or a demo that went well). An agent gets switched to level three because level three is what the business case assumed. That is not governance: it is a setting with a default.
Autonomy should track trust
Autonomy should track trust, and trust is measurable.
At CDI we describe the maturity of an AI conversation on five levels: functional, trustworthy, behavioural, contextual, memorable. An assistant that works (functional) is not yet one a customer relies on (trustworthy); one a customer relies on is not yet one that reads the situation and adapts (behavioural), and so on up. These levels were built to describe the customer's experience, but they turn out to be a better basis for an autonomy policy than any capability checklist. An agent that hasn't reached trustworthy in a domain should not be acting unsupervised in that domain because the customer hasn't yet extended it the trust for that unsupervised action.
But it's not what actually happens. Most teams grant autonomy on the basis of what the agent can do and then wait to discover what it shouldn't have done. The discovery is rarely cheap. A policy cancelled instead of amended, a refund paid twice, a letter from the regulator; and the cost lands in a budget line nobody had connected to the agent.
Why approval steps stop working
There is a second problem, which Jim Fisher at Granola pointed out. His argument is about coding agents, but it applies more widely. The default in most agent tooling is a human approval step: the agent proposes an action, a person clicks approve. This feels safe. The first few approvals are considered but the next few are quick. After ten minutes, the person is clicking approve on reflex because the last twelve were fine, and the approval step has become theatre with all the friction of supervision with none of the protection. It is also the most expensive way to run an agent: a person at the keyboard approving what the agent could have done alone is the automation you paid for, minus the saving. Fisher's conclusion is that real control is not approval but authorisation. Give the agent credentials scoped to what it is permitted to touch, enforce that at the systems it talks to, and let it get to work.
This is how organisations already treat people. Nobody reads and approves everything a new hire does. A new hire is trusted to do certain things without asking, and that list grows with their track record. Whether they are an engineer or an accountant, it works the same way. The approval step is the exception, reserved for actions that are rare and consequential. For agents, we have inverted this: approval for everything, which really means approval for nothing.
A two-part autonomy policy
So an autonomy policy should have two parts. The first is a trust threshold: what level has this agent demonstrably reached, in this domain, with these customers, and what actions does that level license. The second is a structural boundary: what the agent is technically able to touch, regardless of what it decides to do. The first is a design question. The second is a systems question. Both are cheaper to answer up front than in production, which is where most organisations currently answer them. Most organisations have answered neither and have instead bought a slider.
In a bank, that might mean an agent trusted to explain a fee and refund it up to a set amount, with no access at all to the customer's payees. In a health insurer, an agent trusted to confirm what a policy covers, but not to decline a claim. In a government service, one that can book and move appointments but cannot change the record that decides who is eligible for them. The trust threshold sets what each agent may do; the boundary makes sure it cannot do the rest, however confident it becomes.
Measuring trust, and losing it
Measuring trust is harder than measuring containment rate. It means instrumenting the conversation for signals the dashboard doesn't currently show: whether customers accept the agent's answers or re-ask them, whether they escalate after an action or before it, whether they come back. Those signals are the test for the threshold: an agent has reached trustworthy in a domain when customers accept its answers without re-asking, and escalate before its actions rather than after, at rates agreed before it went live. Measuring them is also the only way to put a number on the assumption every agent business case rests on, which is that customers will keep using it. Trust is earned one domain at a time. An agent trusted to reschedule appointments is not automatically trusted to cancel policies, and the level it has reached on one doesn't transfer to the other.
And it means the autonomy setting changes. An agent that has earned unsupervised action in a domain can lose it after a bad week. That is not a failure of the model; it is the policy working. The organisations that find this uncomfortable are usually the ones that wanted the slider to be a one-time decision.
The word autonomous is doing a lot of work right now, and most of it is marketing. The useful question is never how autonomous an agent is. It is who granted it that autonomy, on what evidence, and what happens when the evidence changes. Increasingly that question is being asked by the person who signed off the budget, and they are entitled to an answer.
This is the first in a series on what autonomous agents ask of the organisations deploying them. The series, and CDI's wider work on designing agents for high-stakes settings, is collected on the Agentic Experience Design page.
