Earning Autonomy: Staged Gates for Superintelligent Company Agents
AI agent autonomy levels explained: a staged model where an agent gains write power only after measured accuracy and a named person signs off.

AI agent autonomy levels are staged permission bands that decide what an agent may do without a person stepping in. A sound model starts the agent in read-only work and grants write power one band at a time, and only when measured accuracy on a defined task meets a threshold written in advance and a named person signs off the promotion. The goal is not to slow the agent down but to make every increase in power something a company can show, explain and reverse.
Key takeaways
- Autonomy is a set of permissions and evidence, not a feeling about how capable the agent seems.
- Each level should add one kind of power, such as reading, drafting, writing with approval, or writing within bounds.
- Promotion needs a measured result on a named task, a threshold set before the test, and a named sign-off.
- Every promotion needs a written way back down, so demotion is routine rather than a crisis.
- Choose the first task where a mistake is cheap and visible, not the task with the largest possible upside.
What an autonomy level actually controls
An autonomy level sets four things: which tools the agent can call, which records it can change, whether it can message a customer or supplier, and whether a person must approve each action before it takes effect. It says nothing about how clever the underlying model is. A capable model with broad credentials is still a risk if nobody decided what those credentials open.
This is why the level should be treated as a governance decision rather than a technical setting. The NIST AI Risk Management Framework organizes trustworthy AI work into functions that include governing, mapping, measuring and managing risk. Autonomy levels are a practical way to connect measurement to a permission change: the agent does not gain a new power because the team feels confident, it gains the power because a defined measurement supports it and an accountable person accepted the risk.
The five levels, in order
Five bands cover most small and mid-sized operations. Each level includes everything beneath it, and a single agent can sit at different levels for different tasks at the same time.
- Level one, observe: the agent reads approved sources and records what it would have done, with no outbound effect.
- Level two, draft: the agent prepares replies, updates or records, and a person sends or saves them.
- Level three, act with approval: the agent executes a specific action only after a named person approves that instance.
- Level four, act within bounds: the agent executes routine actions inside a written boundary, such as a category, a value range or a queue, with sampled human review.
- Level five, act and report: the agent runs a stable workflow end to end, reports on a schedule and escalates exceptions automatically.
Most companies should keep high-impact actions at level three for a long time. Level five suits narrow, reversible, well-measured work. It is not the default destination for every task, and a company that reaches it for everything has usually skipped the evidence step. For how trust tiers can be structured around these bands, see trust tiers for agent permissions.
What must be true before an agent gains write power
Four conditions should hold before any promotion. First, the agent has been measured on a sample of real past cases for that task, including awkward ones, not on a clean demonstration set. Second, the success threshold was written down before the test, by the person who owns the process outcome, so nobody can move the bar after seeing the result. Third, the promotion carries a named sign-off from someone accountable for the outcome, and that person is not the same person who built the agent. The mechanics of who signs at which step are covered in human approval gates for agents. Fourth, the new power has a recorded rollback: who can pause it, how quickly, and what happens to work already in progress.
Severity sets how strict the gate is. An agent that tags internal tickets can be promoted with a modest sample and light review. An agent that can change a customer's billing or send a contractual notice needs a larger sample, a narrower scope and approval on every action. The principle of least privilege, which the OWASP authorization guidance treats as a basic design rule, applies directly: grant only the tool scope the current level needs, and deny everything else by default.
A step-by-step plan for each promotion
Use the same cycle for every task and every level change, so the evidence stays comparable over time.
- Name the single task and the person who owns its outcome.
- Write the success criterion in plain language, including what counts as a failed action.
- Build an evaluation set from real past cases and keep it separate from anything used to tune the agent.
- Run the agent at its current level and record every result, not only the failures.
- Compare results with the threshold agreed in advance, and let the owner decide whether to promote.
- Grant the new tool scope with an expiry date or a scheduled review, and record it in the permission register.
- Watch the first real period with a person reviewing a sample of actions and a clear way to pause.
- Record the demotion triggers, then start the cycle for the next task.
Scheduled reviews of that scope are part of the same discipline, as described in quarterly access reviews for agents. Anthropic's engineering notes on evaluating agents make a point worth carrying into this cycle: grading should reflect the outcome the business cares about, not only whether the agent's answer sounded reasonable. An action that reads well but updates the wrong record is a failure, and the evaluation should count it as one.
Illustrative scenario: a triage agent earns write access
Illustrative scenario, not a real account: a small company receives support requests through a shared inbox. The agent starts at level two. It reads each message, proposes a category and drafts a reply, and a team member sends everything. The owner, the support lead, defines success as correct categorization on a sample of past messages and zero draft replies that promise refunds or policy exceptions.
After the agreed review period, the lead signs off a move to level three for one action: applying a category tag and routing the ticket to a queue. Anything that touches money still requires approval for each instance. Several weeks later a new product line produces messages the agent routes to the wrong queue. The lead returns that category to level two the same day, the team adds the new messages to the evaluation set, and the threshold is reviewed before any new promotion. The company did not lose its trust in the agent. It kept a record of what changed and why.
How Be Human handles staged autonomy
Be Human starts each engagement with one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, so the first autonomy level is written down before any write access exists. Managed operations then keep the staged model running: monitoring of actions, recurring evaluations against the agreed criteria, permission review, integration maintenance, incident response and controlled improvements. Promotions and demotions are both recorded, so the company can show who approved each change and what evidence supported it. The service is remote and available in English and Portuguese.
If you are unsure which task should be the first to earn write access, the free Be Human AI diagnostic is designed to help you find it.
Frequently asked questions
Do we need all five levels? Not necessarily. Many teams run well with levels one through three, using level two for most drafting and level three for the few actions that change records or reach outside the company. Levels four and five make sense only once a workflow is stable, measured and reversible, which is usually a narrower set of tasks than leaders first expect.
Who should sign off on a promotion? The person accountable for the business outcome should sign, not the person who built or configured the agent. For operational tasks that is often a team lead, and for anything touching customers, money, personal data or legal commitments it should be someone with authority over that risk. Keep the sign-off in a written record with the date, the evidence reviewed and the scope granted.
What if an agent passes its evaluation and still makes a costly mistake? That can happen in any real process, which is why the promotion includes monitoring, a review sample and a pause path. Treat the incident as evidence: record what happened, decide whether the scope should narrow, add the case to the evaluation set, and re-check the threshold before any further power is granted. Passing a test is never a permanent license.
Can an agent move back down a level? Yes, and it should be possible without drama. Write the demotion triggers at the moment of promotion, such as a misrouted category, a failed sample or a change in the underlying process, so the owner can act quickly. Demotion narrows the permission, keeps the logs and preserves queued work, which makes it a routine control rather than an admission that the program failed.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


