Glossary
›
Autonomous agent

Autonomous agent

Glossary
Also known as:
Autonomous AI agent, self-directed agent

What is an autonomous agent?

An autonomous agent is an AI agent that carries a task from start to finish without a person approving each step: it decides what to do next, acts and moves on, while people set its goals and boundaries in advance and review its results and exceptions afterward.

Autonomy is a matter of degree, not a separate kind of product. The same agent can work alone on rescheduling an appointment and wait for sign-off before offering a payment plan outside the usual range. Calling an agent autonomous says how much of a given task it is trusted to do without supervision, and on which terms.

How an autonomous agent works

Most deployments fit one of four levels. They differ in who makes each decision, not in the underlying technology.

LevelWho decides each stepExample in a customer conversation
SuggestThe agent proposes; a person decides and actsThe agent drafts the reply a representative sends
ApproveThe agent prepares the action; a person signs it offThe agent builds a payment arrangement and a supervisor confirms it
SupervisedThe agent acts; a person watches live and can step inThe agent runs calls while a team lead monitors the live queue
Bounded autonomyThe agent acts alone inside set limits and escalates by ruleThe agent completes a rental application and passes only exceptions to staff

At the higher levels, what keeps the agent safe is the boundary around it: the actions it may take, the thresholds it may not cross, such as a maximum discount or payment term, the situations that trigger a handover, and a log of every step. A person who sets those limits and reviews outcomes, rather than approving each move, is often described as on the loop rather than in it.

Why autonomous agents matter

The range of work agents can do alone is widening quickly. Researchers at METR estimated in March 2025 that the length of tasks frontier AI agents could complete autonomously with 50% reliability had been doubling about every seven months for six years, measured mainly on software tasks.

That figure cuts both ways. Fifty percent reliability is far below what a customer-facing process can accept, which is why autonomy in business runs on narrow, well-defined tasks with checks around them rather than on open-ended goals. Even so, Gartner predicts that by 2028 at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024.

For customer operations, the practical gain is coverage. An autonomous agent works at night, absorbs the spike after a campaign and follows up on a stalled task the moment it stalls, at hours and volumes a staffing plan does not reach.

Where autonomous agents are used

Autonomy fits tasks where the steps are known, the stakes of each step are bounded and the volume is high.

  • Out-of-hours demand. Inquiries, applications and bookings that arrive when staff are off.
  • Routine multi-step requests. Moving an appointment or delivery, updating details and confirming the change.
  • Follow-up. Outreach on stalled applications, missing documents and upcoming renewals.
  • Payment arrangements. Agreeing a plan within options the business has already approved.
  • Peak volume. Taking on surges so queues do not grow.

Complaints, hardship, disputes, high-value negotiations and anything without precedent stay with people, or at the approve level, where a person signs off before anything is committed.

What a good autonomous agent looks like

  • Written boundaries. Permitted actions, amounts and topics are explicit, versioned and reviewed.
  • Escalation by rule. Handover triggers are set in advance rather than left to the agent's judgment in the moment.
  • Confirmation before irreversible steps. The customer confirms what will happen before a payment, cancellation or submission.
  • A stop switch. Staff can pause the agent, or return it to the previous version, at once.
  • Live monitoring. Alerts fire when escalation, error or complaint rates move outside normal ranges.
  • Autonomy earned in steps. A task moves up a level only after performing well at the level below.

How to measure an autonomous agent

  • Autonomous completion rate. Tasks finished with no human involvement, as a share of tasks started.
  • Correction rate. Actions that later had to be reversed or fixed, found through audits.
  • Escalation quality. Whether handovers reached the right team with enough context to act at once.
  • Outcome parity. Results on comparable tasks handled by people and by the agent, side by side.
  • Time to intervention. How quickly a person steps in once an alert fires.

Rules for automated decisions

The more an agent decides alone, the more the rules about decisions apply directly to its work. The main US points follow, with lending as the clearest example; Colorado's and California's rules also reach decisions in housing, employment, insurance and education.

  • Explaining credit decisions. Regulation B requires creditors to state the specific principal reasons for adverse action, and says that pointing to internal standards or a failure to reach a qualifying score is not enough. An agent that contributes to a credit outcome has to leave a record from which those reasons can be given.
  • Automated decision laws. Colorado's revised AI law, which takes effect on January 1, 2027, requires notice before automated decision-making technology materially influences a consequential decision such as one about lending, and an explanation and a route to human review after an adverse outcome. California's privacy regulations on automated decisionmaking technology bring similar obligations into force from 2027.
  • Recurring payments. Under Regulation E, preauthorized electronic transfers from a consumer's account must be authorized in a writing signed or similarly authenticated by the consumer, who must receive a copy. An agent that sets up a payment plan has to capture that authorization.

Frequently asked questions

Is every AI agent autonomous?

No. Every autonomous agent is an AI agent, but many AI agents work at lower levels, suggesting actions or waiting for approval. Autonomy describes how a particular task is set up, not a fixed property of the software.

What is the difference between human in the loop and human on the loop?

In the loop, a person approves decisions before they take effect. On the loop, the agent acts while a person sets its limits, watches results and intervenes when needed.

When should an autonomous agent hand over to a person?

When the customer asks for one, raises a complaint or dispute, shows signs of hardship or vulnerability, or makes a request outside the limits the agent was given.

Can autonomy be increased gradually?

Yes, and it usually should be. A common path runs task by task from suggest to approve to supervised to bounded autonomy, with each step justified by measured results.

Who is accountable for what an autonomous agent decides?

The business that deploys it. Autonomy changes who carries out the step, not who answers for the result.

Autonomy inside the lines you draw.

Encore AI's agents run revenue conversations end to end within a playbook your compliance team approves, with every decision logged.

Scale your best to be the best.

Book a Demo