How Much Autonomy Should AI Agents Have?

Every AI agent has an autonomy setting, whether or not anyone chose it. The agent either acts alone on a given action or it does not, and if leadership did not decide, the project team did. This guide gives executives a framework for deciding deliberately: per action, tied to consequence and evidence, and reversible.

Why is autonomy the central decision?

Because it is where the value and the risk of agentic AI meet. An agent that requires approval for everything delivers little more than a drafting tool. An agent that acts alone on everything is a liability waiting for its first bad input. The value is in setting autonomy correctly for each kind of action, and the framework for that is a leadership responsibility, not a technical default. FISTA's AI agent human oversight guide covers the mechanics of review; this piece covers the decision.

What is the autonomy dial?

Autonomy is set per action class, not per agent. A single customer-service agent might run autonomously on order lookups, with sample review on address changes, with exception-only review on refunds within policy, and with full review on anything involving a complaint.

Level	What happens	Fits when
Full review	A person approves every action before it takes effect	New action class; high consequence; little evidence
Sample review	Actions take effect; a defined sample is reviewed after; a trigger returns to full review	Evidence accumulating; consequence moderate and reversible
Exception-only	Agent acts within policy; escalates what falls outside	Strong evidence; policy is written and testable
Autonomous with monitoring	Agent acts; monitoring and scheduled evaluation watch for drift	Low consequence; reversible; strong evidence over volume

The glossary entry what is an autonomy level in AI gives the formal scale.

What sets the dial?

Two inputs.

Consequence of a wrong action. Is it reversible? What does it cost? Who is affected: an internal team, a customer, a regulator? Does it create a legal commitment? The higher the consequence, the higher the starting level and the more evidence required to move down.

Evidence. Agreement between the agent's proposed actions and reviewers' decisions over a meaningful volume; the evaluation pass rate for that action class; the incident history; and whether monitoring exists to catch regression. Evidence moves the dial down; its absence keeps it up.

A useful rule: the burden of proof is on the evidence. An action class stays at its current level until the numbers justify moving it, and the numbers required are set before the review period starts, so the decision is not relitigated each month.

Which lines does policy hold regardless of evidence?

Some actions stay human because the company would not delegate them to a junior employee without sign-off, however competent: legal and contractual commitments, regulated decisions with legal effect on individuals (credit, employment, insurance, and similar, where automated-decision rules may also apply), large financial transactions, personnel decisions, and communications during incidents. Risk, security, and legal set these lines; evidence does not move them. This is general guidance, not legal advice; the specific lines depend on jurisdiction and industry.

How is autonomy withdrawn?

Autonomy is reversible by design. An incident in an action class returns it to full review until root cause is found and fixed. A drift alert or a falling evaluation pass rate does the same. A change to the process, the policy, the tools, or the model resets affected action classes to review until re-evaluated. The AI agent kill switch design guide covers the extreme case of stopping an agent entirely.

Who decides?

The business owner of the process proposes level changes with evidence. The executive accountable for the function approves changes in authority, because they are accountable for the outcome. Risk, security, and legal hold the policy lines. Engineering implements the decision as permissions and approval gates, and reports the evidence monthly. The AI decision rights framework guide extends this to the whole program.

What does the progression look like in practice?

A new agent launches with full review on every consequential action. After a defined volume, low-consequence action classes with high agreement move to sample review. After a further period, those with clean histories move to exception-only or autonomous with monitoring. High-consequence classes move slowly or not at all. The whole progression is recorded, so at any point the company can show why each action class is at its level. That record is what boards, auditors, and regulators increasingly ask for; the AI oversight for boards whitepaper explains why.

What should executives ask?

• For each agent, which action classes run without review, and what evidence supported each release?
• Which lines does policy hold regardless of evidence, and are they written down?
• When was autonomy last withdrawn, and why?
• Who approves changes in authority, and is the evidence standard set in advance?

How can FISTA Solutions help?

FISTA Solutions builds AI agents with autonomy implemented as per-action permissions and approval gates, with the agreement tracking and evaluation that produce the evidence for each decision, and its AI enablement practice helps executive teams set the framework and the policy lines. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries.

To set an autonomy framework for the agents you are deploying, talk to FISTA on WhatsApp, or continue with human-in-the-loop AI explained.