Approval Thresholds for AI Agents: A Ladder for Earning Autonomy
Set human approval levels for AI agents by value at stake, reversibility and customer impact, so autonomy grows only as evidence builds.

Human approval for AI agents works best as a ladder of permissions, not a single on or off switch. Each rung defines what an agent may do alone, and the rung for any action depends on three things: how much value is at stake, whether the action can be reversed, and how many customers feel the result. An agent moves up a rung only when its logged decisions show it handled the current rung reliably.
Key takeaways
- A single approval rule either slows routine work or lets risky actions through. Place each action by its stakes, not by the tool it uses.
- Score every action on three axes: value at stake, reversibility and customer impact. The highest placement sets the rung.
- Start every new agent at the observation rung, where it drafts and a person performs the action.
- Raise autonomy only on evidence from logs and sampled reviews, and lower it again after an incident.
- Write the thresholds into your own policy. The exact limits are a business decision, not a vendor default.
Why a single approval rule fails
Most teams begin with one of two rules. Either a person approves everything the agent does, or the agent runs without approval once a short test looks good. Both fail for the same reason: the check does not match the consequence. Full approval turns the agent into a slower draft tool, and reviewers start clicking through prompts without reading them. Zero approval after a brief test gives the agent the same authority over a routine reply and over a refund on a disputed order.
The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) organizes AI risk work around governing, mapping, measuring and managing risk. A ladder of approval levels is one practical way to turn that structure into daily permissions that an operations team can read, audit and change.
Three questions that place an action on the ladder
Ask three questions about every action the agent could take. Answer each one for the worst realistic outcome, not the typical one.
- Value at stake: what is the most this action can commit or cost the company? Include refunds, credits, purchase orders, contract commitments and any spending the agent can trigger. Finance and the process owner should set the limits.
- Reversibility: can the action be undone completely, partly, or not at all? A draft reply is fully reversible. A message already delivered to a customer's phone is not. A payment that has settled needs a separate reversal process with its own approval.
- Customer impact: how many people see the result, and how sensitive is it? One internal note is different from a change that affects every account on a shared list, or a message that exposes personal data.
Use the highest of the three placements. A small, reversible action that reaches many customers is not a low-risk action. If the action touches personal data, also check the rules that apply to it. For Brazilian companies, LGPD and AI Agents: What Brazilian Companies Should Check covers the questions to settle before an agent handles that data.
Four rungs you can adopt
A four-rung ladder fits most small and mid-sized company agents. You can rename the rungs to match your operations, but keep the order and keep the logic.
- Rung one, observe: the agent gathers information and prepares drafts. A person reviews each draft and performs the action.
- Rung two, act with after-the-fact review: the agent performs low-value, reversible actions inside fixed limits. A person samples the results on a set schedule and can reverse any of them.
- Rung three, act with pre-approval: the agent prepares the action with its reasoning and the supporting records. A named person approves each instance before the system executes it.
- Rung four, dual control: actions with large value, irreversible effects or broad customer reach need two named approvers, or a check from a second function. Use it for actions where a single mistake is costly to undo.
Anything outside every rung is not delegated. Treat that as the default, so that new tools start with no permissions until someone places them on the ladder. The question of which actions deserve dual control is covered in more detail in Two-Person Approval for AI Agents: Which Actions Need It.
Illustrative scenario: one support agent, four kinds of request
The following scenario is illustrative and does not describe a real client. An online shop uses an agent to handle customer messages. It can read order history, draft replies, and use a small set of tools.
A customer asks to change a delivery address before the parcel leaves the warehouse. The change is reversible until dispatch, affects one customer, and has a clear record. The agent performs it on rung two, and a person samples the changes each week.
A second customer asks for a refund within the standard return policy. Money leaves the company, so the agent prepares the refund with the order record and the policy clause, and a support lead approves it on rung three.
A third customer asks for an exception to the return window on a large order. This is a policy exception with higher value, so the agent routes it to a manager with a summary and does not offer any outcome to the customer.
Finally, the agent proposes a message to all customers about a delivery delay. The message reaches many people and cannot be recalled. It sits on rung four. The text is fixed in advance, and two named people approve the send.
The same agent, four requests, four placements. The ladder is what keeps routine work fast while the consequential work waits for a person.
A plan for setting thresholds in your company
Work through these steps with the owner of each process, not only with the technical team.
- List every action the agent could take in its first role, one line each, including the tools it would call.
- For each action, write the worst realistic commitment or cost, whether it can be undone, and who would see the result.
- Assign a rung using the highest of the three answers. Mark any action that touches personal data for a privacy check before launch.
- Enforce the rung in the system, not in a sentence inside the prompt. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) recommends denying access by default and checking permissions on every request, which is the behavior the ladder needs.
- Name the approver for each rung, the backup approver, and the time window in which they must respond.
- Define the success criterion for the first role before launch, such as the share of drafts accepted without edits on a sample, and list the events that trigger an automatic pause.
- Decide in advance who may lower a rung, and under which conditions.
What evidence moves an agent up, and what moves it down
Evidence should come from records the system writes at the time of each action: the inputs used, the tools called, the reasoning summary, the approver, and the result. Concepts such as traces and logs, described in the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/), give a practical vocabulary for building those records. Without them, a manager cannot tell whether an agent was reliable or merely lucky.
Promotion should require a sustained period at the current rung, with no unauthorized actions and no sampled error that the process owner judges material. Demotion should be fast. Any unauthorized action, a customer complaint traced to the agent, or a change in its tools or data returns it to the rung below until the cause is reviewed. The reasoning behind gradual promotion is set out in Earning Autonomy: Staged Gates for Superintelligent Company Agents.
Keep a pause path that works during an incident. Stopping new actions should not destroy logs or queued work, and the steps are covered in AI Agent Kill Switch: Halt Actions Without Losing Logs or Queued Work.
How Be Human's managed approach applies the ladder
Be Human designs, implements and operates personalized AI infrastructure for companies, with agents that keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, so each rung is written into the agent's permissions from the start. Managed operations can then include monitoring, recurring evaluations, permission review, incident response and controlled improvements. Evidence from those activities is what decides promotions, and the same review process can lower a rung when a tool, a data source or a policy changes. Service is remote, in English and Portuguese. If you are not yet sure which task deserves the first rung, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Does every action need human approval at the start? Not every action, but most of them should begin on the observation rung, where a person performs the action after reviewing a draft. Low-value reversible actions can move to act-with-review once the process owner has checked the first sampled results. Anything not placed on the ladder stays unavailable to the agent.
Who decides the dollar limits for each rung? The finance lead and the owner of the process should set them, using the company's own risk tolerance and policies. Vendor defaults are a starting point at most, and they rarely match a specific business. Write the limits into the policy, review them on a fixed schedule, and record every change.
What happens if an agent acts outside its rung? Treat it as an incident. Pause the agent, preserve its logs and queued work, find out how the permission allowed the action, and return the agent to a lower rung until the cause is fixed. Record the decision so that the review can be explained later.
Does reaching a higher rung mean the agent is trustworthy? It means the evidence fits that rung, which is narrower than general trust. An agent can be reliable on refund drafts and still not be ready to send messages to every customer. Keep the permissions specific, and judge each rung on its own record.
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.


