Personalized AI infrastructure
Approval Ladders for Superintelligent Agents: From Draft to Autonomy
Set staged approval levels so your AI agent drafts first and gains wider write authority only after its supervised work record holds up.

Human approval for AI agents works best as a ladder, not an on-off switch. The agent starts in draft mode, where every consequential action waits for a named person to approve it, and it earns wider write authority one rung at a time. Each step up needs a written scope, a measurable success criterion and evidence from supervised work that the agent handled that specific task reliably.
Key takeaways
- Start every agent on a draft-only rung, with a named person approving each consequential action.
- Define each rung by the actions it allows, the systems it may touch and the highest-impact action it can take without asking.
- Promote an agent only on evidence from its own supervised record, judged against a criterion written before the trial began.
- Make demotion as easy as promotion, and rehearse it so a rollback is routine rather than an emergency.
- Keep identity, credentials, channels and memory under company control at every rung, including the top one.
Why a single approve-or-deny setting fails
Most companies begin with one of two settings: the agent asks before every action, or it runs unattended. Both fail for the same reason. A system that asks about everything trains people to click yes without reading, and a system that never asks gives you no evidence about where its judgment is weak. A ladder fixes this by splitting authority into stages, each with a narrow purpose and a clear exit condition.
The ladder also matches how trust works in a working team. A new analyst drafts reports that a manager reviews, then sends routine reports alone, then handles supplier terms only after a long run of sound work. An agent should follow the same sequence, and its errors deserve more caution than a person's, because it can repeat one mistake across many records before anyone notices. Anthropic's guidance on building effective agents makes a similar point: start with simple patterns and add autonomy only where the task truly needs it. The same logic applies to superintelligent agents, whose reach will grow faster than most teams expect, so the approval design matters more as capability rises.
How do you define the rungs of the ladder?
Four rungs cover most business agents. Write them down before the agent touches a live system.
- Rung 1, draft: the agent prepares outputs such as replies, record updates, quotes or summaries in a staging area. A named person reviews and sends or saves each one.
- Rung 2, supervised write: the agent writes directly to one system, within a narrow template, and a person approves each write before it commits.
- Rung 3, bounded autonomy: the agent acts alone on a defined class of routine cases, with sampling reviews and an automatic stop when an error threshold is crossed.
- Rung 4, delegated autonomy: the agent acts across a wider class of cases, with periodic permission reviews and a formal incident process. Even here, credentials and memory stay under company control.
Each rung needs three written fields: the actions it allows, the systems and data it may touch, and its limit. Express the limit in your own operating units, such as a customer segment, a refund category or a count of records per day, so that a non-technical owner can read it and judge it. For a practical first week on a new agent, see what the first months with a managed AI agent look like.
What evidence should earn the next rung?
Promotion should rest on the agent's supervised record, not on a demo or a vendor's claim. Agree on the criterion before the trial starts, because a criterion written after the results tends to fit them. Use this checklist before any promotion:
- A success criterion stated in business terms, such as drafts accepted with wording changes only, or tickets closed without escalation.
- A sample of supervised cases long enough to include the ordinary variety of the task, plus the unusual inputs that cause trouble.
- A complete log of inputs, actions and reviewer decisions, so the record can be audited later. OpenTelemetry's observability concepts describe the traces that make this possible.
- A failure list in which each failure is marked harmless, costly or unacceptable by the owner of that process.
- A reviewer who did not build the agent and who signs off on the promotion in writing.
- A re-check condition: any change to the underlying model, a connected tool or the data source reopens the evidence for the current rung.
An illustrative scenario: a support agent moving up
Illustrative scenario: a regional distributor, described here as a hypothetical example, uses an agent to answer delivery questions. In the first phase it drafts replies that a coordinator sends. The agreed criterion is that reviewers accept drafts with wording changes only, and that no reply promises a delivery date the logistics system does not show. After enough cases that include late shipments, partial orders and address changes, the coordinator notices that the agent handles address changes well but misreads split shipments. The team promotes the agent to rung 3 for single-shipment status questions only, and split shipments stay in draft. After the split-shipment logic is corrected and tested in draft, that case class is added in a second, separate decision. The promotion stayed narrow, reversible and tied to a named failure.
Step by step: moving an agent up or down a rung
- Write the rung definition, its allowed actions, systems and limits, before changing anything.
- Run the agent on its current rung until the criterion holds across a varied set of cases.
- Review the failure list with the process owner and classify each failure.
- Narrow the promotion to a single case class where possible, rather than widening everything at once.
- Apply the permission change in its narrowest form, using scoped credentials rather than shared accounts.
- Keep sampling reviews running for the first cases after the change.
- Record the decision, the evidence, the reviewer and the rollback trigger in the agent's change log. The process for proposing and approving these changes is described in how to request and approve changes to a managed AI agent.
- Define demotion triggers in advance, such as a costly error, an action outside the written scope or a shift in the data the agent sees.
- Rehearse one demotion, so the team knows how to restore the previous rung quickly.
Which permissions stay human at every rung?
Approval is a permission design problem as much as a quality problem. OWASP's authorization guidance favors checking access on every request and granting only the access a task needs, so an agent's credentials should carry only the scopes its current rung requires. Keep those scopes narrow, and review them on a schedule, using the steps in how to run a permission review for your AI agents.
Text from emails, web pages or documents can also try to instruct an agent, a risk that grows with autonomy. Tool calls triggered by untrusted content should pass the same gate as any other write, and the agent's authority should not expand because a document asked it to. The NIST AI Risk Management Framework treats governance, measurement and management as continuing activities, which suits a ladder that is reviewed rather than set once.
Be Human's managed approach builds the ladder into the first engagement. That engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, all written in the company's own terms. During managed operations, Be Human monitors the agent's record, runs recurring evaluations, reviews permissions, handles incidents and proposes changes that pass through the same approval steps as everything else on the ladder. The company keeps ownership of identity, channels, credentials and memory, so a change in rung never depends on one vendor's settings.
Keeping the ladder honest over time
Ladders drift. A model update, a new tool or a shift in customer mix can make an agent that passed last quarter fail this quarter. Schedule a permission review and an evidence review on a fixed cadence, and treat any change to the underlying model or tools as a reason to re-check the current rung. Keep the change log readable by the person accountable for the process, so they can see who approved each step and why. Over several review cycles, that log shows which rungs hold and which still need more evidence. If you are unsure which task should start on the first rung, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How many approval rungs does a business agent need? Most business agents work well with three or four rungs, starting with draft-only access. Fewer rungs make promotions too large to review, and more rungs add paperwork without adding control. Choose the number that lets each step be narrow and reversible.
Who should decide whether an agent moves up a rung? The process owner who lives with the outcomes should propose the move, and a reviewer who did not build the agent should confirm it in writing. Keep the engineering team responsible for the technical controls, but do not let it alone decide business risk. The decision and its evidence belong in the change log.
What if an agent makes a mistake after it has been promoted? Apply the demotion trigger you wrote before the promotion, restore the previous rung, and record the failure with its cause. Then decide whether the fix belongs in the agent's instructions, its tools, its permissions or its evaluation set. Promote again only after the fix has been tested on the rung where the failure occurred.
Can an agent skip rungs if it performs well from the start? Skipping rungs removes the supervised record that justifies the next step, so it is rarely worth the saved time. A strong early result still needs a varied sample and a reviewer's sign-off before wider authority. If the first results are excellent, the evidence will show that quickly, and the promotion will take less effort.
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.


