A team should not answer "How autonomous is the agent?" with one vague setting. Authority depends on the action. Reading a record, drafting a change, recommending it, making a reversible update and approving a consequential decision are different jobs.
The MAJLS AI action ladder separates those jobs into five levels. It gives an operations leader a way to choose a starting point, record exact permissions and demand operating evidence before granting more authority.
NIST's AI Risk Management Framework playbook advises teams to document the roles and responsibilities for human oversight in a deployed system.[1] UK government guidance adds that humans should validate high-risk decisions influenced by AI and that teams need ways to intervene meaningfully.[2] For an operating team, that means deciding the approval boundary before the system touches live work.
The five authority levels
- Observe. The role reads approved sources, checks their availability and records what it would have done. It cannot create a business record or change a workflow state.
- Draft. It prepares a note, reply, update or request for a person to inspect. Nothing leaves the draft area automatically.
- Recommend. It proposes a specific action and shows the evidence, rule and unresolved exceptions behind that proposal. A named person chooses whether to proceed.
- Execute a bounded, reversible action. It may perform an approved action within explicit limits, such as applying an internal label or moving a complete case to a review queue. The action is logged and can be undone.
- Prepare a consequential action for human approval. It assembles the evidence and prepares the change, but a named person alone authorizes the commitment. Legal, financial, safety, account-access, publication and customer-commitment decisions stop here.
Level five is not the top of a conventional autonomy scale. It is a hard boundary. The system may improve the quality and speed of preparation, but it does not inherit the approver's authority.
Choose the starting level
Start at the lowest level that produces useful evidence about the real workflow. If the inputs are poorly understood, begin with observation. If the team understands the work but needs to learn whether outputs are usable, begin with drafting. Recommendation makes sense only when the role can show why it reached a proposal and route exceptions to the right person.
Level four requires more than accurate-looking output. The permitted action must be narrow, reversible and easy to audit. The owner should be able to describe the maximum impact of one wrong action, the rollback path and the condition that stops further execution.
Do not choose a level from a model confidence score. Confidence is not proof that the inputs were current, the policy applies, the action is authorized or the downstream system behaved as expected.
AI action ladder and promotion-gate worksheet
Complete one worksheet for each action. Do not write "system access" as a blanket permission.
- Action: Name the exact read, draft, recommendation or change.
- Current level: Record a number from one to five and the reason it is useful now.
- Allowed tools and records: List the systems, fields and data classes the role may use.
- Action limits: Set volume, value, time, status and case-type boundaries where they apply.
- Evidence shown: Define the records, checks and policy reference that must accompany the output.
- Audit trail: Record input versions, action, timestamp, outcome, exceptions and reviewer correction.
- Rollback: Name who can reverse the action, how they do it and how quickly reversal must be possible.
- Approver: Name the person or accountable role for any consequential step.
- Stop conditions: Include missing evidence, conflicting records, failed validation, attempted out-of-scope action and unavailable rollback.
Then use one of three decisions at review:
- Promote only when the action has stayed within scope, expected exceptions reached the right reviewer, outcomes reconciled to source records, rollback was tested and the accountable reviewer signs off.
- Hold when the action is useful but the evidence window is too small, an exception pattern remains unclear or the rollback procedure has not been exercised.
- Reduce authority after an unauthorized-action attempt, an unexplained mismatch, a missed consequential case, an audit gap or a change in policy or connected systems.
Promotion should be scoped to one action. Success at drafting access recommendations does not grant permission to modify accounts, and success in one system does not transfer to another.
Illustrative example: an internal software-access request
This example is illustrative. It does not describe a client deployment or measured result.
At level one, the role reads a request, the approved role catalogue and the employee's current access, then records missing evidence. At level two, it drafts a request summary. At level three, it recommends the least-privilege option and links each proposed permission to the submitted evidence and policy.
A safe level-four action could be limited to adding a validated request to an internal review queue. That change is reversible, creates no account and grants no permission. The record includes the request ID, validation results, queue movement and source versions.
The account change remains level five. The role can prepare the change set and show conflicts or missing approvals, but the named system owner authorizes it. If employment status is unclear, the requested permission is absent from the catalogue, records conflict or the change cannot be rolled back, the role stops and hands the case over.
The handover should state the failed condition, evidence checked, proposed action and exact decision needed. It should not hide uncertainty behind a confidence number.
Review authority as part of the operating system
Set a regular review based on the consequence and frequency of the action. Review corrections, exceptions, unauthorized attempts, rollback results and changes to policies or integrations. A promotion decision should name the action being promoted, the new limit, the evidence reviewed and the person who approved it.
The ladder is useful because it makes authority inspectable. A manager can see what the role may do today, what remains human-only and what evidence would justify a change. When that record is missing, the answer is not more autonomy. It is a better operating decision.