The Best AI Agent Knows When to Do Less
Autonomy is not a universal setting. The same action can feel efficient to one person and alarming to another.
Published 2026-08-28 ยท 5 min read
The Best AI Agent Knows When to Do Less
One Autonomy Setting Creates Two Bad Experiences
An AI agent that always asks before acting feels slow to a user who wants it to move. An agent that acts first and explains later feels unsafe to a user who wants to review every consequence.
Most products solve this by picking one default. That is understandable: a single behaviour is easier to design, explain, and support. It is also why many useful AI features are loved by one group of users and quietly avoided by another.
Autonomy tolerance is not a plan-tier preference. It is a combination of personal decision style, task stakes, and current state. Someone who happily lets an agent organize a calendar may want approval before it changes a payment, publishes a message, or touches customer data. The same person can move between those modes within a session.
The Signals Are Already In The Product
You do not need to ask every user to complete an autonomy questionnaire. The product often sees their preference in ordinary interaction:
These are not proof of intent individually. Together, and compared with a personal baseline, they form useful context. A user who usually accepts low-stakes recommendations but starts opening every detail before a contract approval is not rejecting automation. They are signalling that this moment deserves more control.
Reversibility Comes Before Confidence
Model confidence is useful, but it should not be the only input to autonomy. A confident action can still be a bad action to automate if it is difficult to undo.
The better ordering is:
1. Is this action reversible and visible?
2. What does the person's behaviour say about their desired control right now?
3. How confident is the system in the recommendation?
For a reversible, low-stakes action, an agent can propose a default and make the correction easy. For an irreversible or high-stakes action, the right design is usually explanation and confirmation, even when the system is highly confident.
This is not caution for its own sake. It protects the trust that makes automation valuable. A person surprised by an agent once starts supervising it everywhere. The time saved by automation disappears because the product taught the user that it might act beyond their comfort level.
Design The Handoff, Not Just The Action
Adaptive autonomy should feel legible. The user needs to understand what the system is about to do, why it is offering that level of control, and how to change it.
Three product patterns work well:
Suggest, then let the person decide. Use this when the action carries meaningful consequence or the behavioural read is uncertain. The agent can do the preparation work without taking the final step.
Act with a visible undo. Use this for reversible actions where the cost of a wrong read is low. The undo cannot be buried in settings; it is part of the action itself.
Explain a change in behaviour. If the agent normally acts automatically but asks this time, say why in plain language. A stable mental model matters more than a clever adaptation users cannot see.
What Teams Should Measure
The wrong success metric is the percentage of tasks the agent completed without a person. That number can rise while trust falls.
Measure acceptance without correction, undo rate, time to completion, repeat use, and the difference between a user accepting assistance and merely tolerating it. Look for harm as carefully as lift: unexpected actions, escalations, or people turning the feature off after a surprise.
An agent that does less on the right moments can create more long-term value than an agent that does more everywhere. Behavioural context gives a product the missing input: not whether automation is generally good, but how much automation this person is ready to delegate right now.