AI Agent Autonomy Checker

Thinking of letting AI handle a job in your business? Describe the task and answer eight questions. The checker tells you whether AI should only advise, prepare work for a person to approve, or act on its own, and which safeguards to put in place. The rule it follows is simple: the risk of the task sets the most freedom AI can have, and no amount of good performance can raise that limit.

Decision support, not a legal ruling. This checker gives general guidance based on your answers. It is not legal, regulatory or compliance advice and does not decide whether a use of AI is lawful in your situation. If a task involves regulated activity, decisions about people or significant sums, take proper advice. Your answers stay in your browser and are not sent to us.

What this tool does

It places a task on a four step ladder of AI autonomy and explains why:

Advise only. AI suggests; a person does the work and makes the decision.
Prepare for approval. AI does the work, such as a draft letter or a payment run, but nothing takes effect until a person approves it.
Act within limits. AI acts on its own inside firm boundaries, and a person reviews a sample of what it did afterwards.
Act on its own. AI runs the job routinely, with monitoring and alerts when something looks wrong.

When to use it

Before switching on an AI agent or automation for a new job.
When a supplier offers to "let the AI handle it" and you want a second opinion on how far to go.
When writing an AI use policy for your staff, to decide which jobs need a person to sign off.

How it works

Each risk answer sets a ceiling. A task that can pay out money or cannot be undone can never go above "prepare for approval". A task that makes decisions about people with legal or similarly significant effects, such as hiring or credit, stays at "advise only", which reflects the protections in UK GDPR on solely automated decisions. The lowest ceiling wins.

Only then does the checker look at how well the AI has performed. A poor or untested track record pulls the recommendation down. A good one never pushes it above the risk limit, because a mistake that is rare is still unaffordable if it is large, permanent or unlawful. The answers that set the limit are listed with the result, so you can see what would need to change.

We use the same approach in our own business: see what we hand over to AI and what we don't, and if you are not sure what counts as an agent, start with what is an AI agent.