Six levels of AI agent autonomy

Autonomy means how much an AI system may do without a person stepping in. Sociotek uses six levels, A0 to A5. There is no best level. The right one depends on the work and on what could go wrong.

Warm marks the human system: judgement, relationships, accountabilityCool marks the agent system: identity, permissions, execution
A0

Human execution

AI retrieves or generates information. People perform every consequential action.

For example: You ask an AI tool to summarise a policy, then you write the reply yourself.

What it calls for: Ordinary care with information.

A1

Adviser

AI recommends. A person interprets, decides and acts.

For example: The AI points out which claims look unusual. An assessor decides what to do about each one.

What it calls for: Know where the advice comes from, and review its quality.

A2

Approved executor

AI prepares or executes bounded actions after explicit human approval.

For example: The AI prepares the refund and waits. Nothing happens until a person presses approve.

What it calls for: Approval, a record of what was done, and a way to reverse it.

A3

Bounded operator

AI completes a defined workflow and escalates exceptions.

For example: The AI handles standard refunds from start to finish and passes the unusual ones to a person.

What it calls for: Its own identity with limited access, monitoring, and a tested way to stop it.

A4

Persistent delegate

AI pursues continuing goals, initiates work and adapts between interactions.

For example: The AI watches the queue around the clock, starts work when something arrives and remembers earlier cases.

What it calls for: A mandate that expires and must be renewed, rules for what it remembers, and continuing checks.

A5

Orchestrated agent system

Multiple agents dynamically divide work, use tools and hand off tasks.

For example: Several AI agents pass work between them: one checks documents, one assesses, one pays.

What it calls for: Controls on what agents may pass to each other, identity between agents, containment, and monitoring of the whole system.

The line between A2 and A3

Up to A2 a person approves each action before it happens. From A3 the agent acts first and people review afterwards, or only when something unusual comes up. That is the point at which a short list of controls stops being good practice and becomes a condition of going ahead. Read: who can switch it off?

The goal is the right level for each workflow. Moving everything to A5 is nobody's aim.