Does autonomy require a physical robot?

No. Software agents can act through digital tools and services.

Is there one universal autonomy scale?

No. Different domains and frameworks define levels differently, so the applicable scale and scope must be stated.

Can an autonomous system remain under human governance?

Yes. People can define objectives, boundaries and oversight even when they do not direct every individual step.

AI Glossary / Governance, ethics and safety / Beginner

Autonomy

Definition

Autonomy in an AI system is its ability to select and carry out actions toward a goal without a person directing every step, within a defined scope. Its degree depends on the task, environment, permissions and human involvement, rather than a single universal label.

What it means

A writing assistant that suggests an email has different autonomy from an assistant that chooses recipients and sends it. Both may use the same language model. The difference lies partly in the surrounding workflow and the actions the system is permitted to take.

Autonomy does not mean consciousness, personal wishes or unlimited competence. A warehouse robot may navigate independently within one mapped building and still require help with an unfamiliar obstacle. A software agent may choose a sequence of searches while needing approval before changing a customer record.

Human autonomy is another important meaning: people's ability to make their own choices. Increasing machine autonomy can support or restrict that ability depending on the design. This entry focuses on machine autonomy while keeping that governance distinction visible.

How it works

Designers specify an objective, the information a system can observe and the actions it can perform. The system interprets the current situation, selects an action and observes the result. It may revise its plan when conditions change, subject to its capabilities and constraints.

Human involvement can occur at several points: setting goals, approving particular actions, supervising execution or handling exceptions. These arrangements should be described directly. Terms such as “fully autonomous” are incomplete unless the task and operating conditions are stated.

Capability and permission also differ. An agent might be technically capable of drafting a purchase order but lack permission to submit it. Conversely, broad credentials may permit actions that its reasoning cannot perform reliably. Evaluation must check both the quality of its decisions and the effectiveness of the boundaries around execution.

A robust workflow defines what happens when information is missing, a limit is reached or a tool fails. Appropriate escalation can be part of a successful autonomous task.

Three practical examples

A calendar assistant finds suitable meeting times independently but asks the organizer to approve invitations before sending them.
A warehouse vehicle selects a route within its operating area and stops or requests assistance when conditions exceed its design limits.
An inventory agent prepares replenishment orders and submits only those meeting explicitly configured supplier, item and spending rules.

Think of it this way

Think of delegating a bounded assignment to a colleague: “Organise the meeting, but confirm the external guest list with me.” The colleague has discretion over some steps while the approval boundary remains clear.

The analogy does not give an AI human understanding or responsibility. Software permissions, controls and evidence must enforce the boundary; a natural-language request alone may not reliably constrain every connected tool.

Worked example

Describe the action boundary

Consider a fictional stock agent with permission to order approved stationery up to 200 units of currency per order and 500 per day. It has already submitted orders totaling 350 today. It now proposes an approved-item order worth 180.

The proposed order meets the individual limit because 180 is below 200. It fails the daily limit because 350 + 180 = 530. The execution service therefore blocks automatic submission and routes the request for the defined approval process.

The agent can remain autonomous in checking stock, comparing approved suppliers and preparing the request. Requiring approval at this boundary does not mean every earlier step was manual. Nor should it split the order to evade the daily rule. The example shows why autonomy must be described through concrete actions and aggregate constraints, not merely whether the assistant can use a purchasing tool.

Illustrative case study

Fictional Elm Facilities pilots an agent that schedules routine maintenance. Early descriptions call it autonomous because it creates its own task sequence. Testing reveals that the same account can also cancel urgent appointments, although this was never part of the intended scope.

The team separates permissions, allows routine scheduling within defined conditions and routes urgent changes to a responsible coordinator. It records what the agent attempted, what the service accepted and when assistance was required.

During evaluation, a missing technician availability feed causes the agent to stop and escalate. The team treats this as correct boundary handling, not automatically as failure. Its revised description specifies the tasks and conditions the system can manage independently, making the deployment easier to assess.

Common misconceptions and mistakes

More autonomy means more intelligence. Independence of action and capability are related but different attributes.
Autonomous means unrestricted. Useful autonomy is normally bounded by tasks and conditions.
A human can always intervene effectively. Intervention requires timely information, access and sufficient response time.
The model determines autonomy by itself. Tools, credentials, workflow and deployment settings also determine what can happen.

Business use

For businesses, autonomy is a delegation decision with operational consequences. Define what the system may decide, what it may change and which exceptions require another actor. Assess failures at the level of completed actions, including retries and interactions with external services.

Procurement comparisons should use the same task and conditions. A demonstration that completes an ideal workflow says little about interruptions, ambiguous requests or unavailable dependencies. Track successful independent completion alongside appropriate escalation and unintended actions. A low escalation rate is not necessarily good if the system is proceeding beyond its evidence or authority.

Check your understanding

An agent independently researches suppliers but cannot place an order until a person approves it. Is it accurate to say it has no autonomy?

Answer: No. It has autonomy over the research steps and a human approval boundary for purchasing. Describe those capabilities separately.

Related terms

AI agent
Automation
Controllability
human oversight
operational domain
permissions
delegation
human autonomy

Learn next

Read AI agent and controllability. Map one workflow into recommendations, decisions and executed actions. For each, identify who has authority and which conditions require the system to stop or escalate.

Sources

NIST, Autonomous systems: https://www.nist.gov/autonomous-systems. Huang, Terminology for Specifying the Autonomy Levels for Unmanned Systems, version 1.0: https://www.nist.gov/publications/terminology-specifying-autonomy-levels-unmanned-systems-version-10. NIST AI RMF, Human-AI Interaction: https://airc.nist.gov/airmf-resources/airmf/appendices/app-c-ai-risk-management-and-human-ai-interaction/. These establish system and contextual perspectives; software purchasing examples are original illustrations.

Previous: Autonomous agent Next: Autorater evaluation