Human Approval Gates for Superintelligent Agents: Where People Sign
Learn which agent actions need a named human approver, how to sort them into approval classes, and how to write gates you can test before launch.

Human approval for AI agents means a named person must review and accept specific actions before the agent carries them out. The gate belongs on actions that spend money, change records, speak for the company externally, or widen what the agent can touch, not on every routine step. Decide these gates in writing before the agent receives its first tool, because a gate defined after an incident is usually defined too late.
Key takeaways
- Place approval gates on actions that are costly, external, hard to reverse, or that extend the agent's own permissions.
- Name one approver for each gate and a backup who can decide when the first person is unavailable.
- Treat the agent's own confidence as an input to review, never as the control itself.
- Raise gates when an agent's scope or capability grows, and lower them only after recorded evidence.
- Test the path a person uses to reject or pause an action before the agent goes live.
Why model behavior cannot be your approval control
Many teams begin by trusting the model to recognize when it should stop. That trust is understandable, since capable models often explain their reasoning and decline obvious mistakes. But a model's behavior shifts with the input it receives, the tools it can call, and the instructions it reads, including instructions hidden inside documents or emails it processes. The OWASP Top 10 for LLM Applications lists prompt injection as a standing risk for this reason, so an agent that follows only its own judgment can be steered by content it was never meant to obey.
A control worth the name sits outside the model. It is a rule in the tool layer or the workflow that holds a specific action until a specific person accepts it. Authorization checks belong in that layer as well, as the OWASP authorization guidance describes, so the agent cannot take an action its credentials do not allow even if it asks. The NIST AI Risk Management Framework treats this as governance: roles, documented decisions and oversight around the system, not only properties of the system itself. Your approval gate is where that governance becomes something people can operate.
Which actions need a named approver
Sort every action your agent could take into four classes before deciding anything else. Where the agent should stop and ask is a related question, and the guide on where an AI agent should stop and ask goes deeper on triggers. The classes below are simple enough for a manager to apply in one meeting.
- Read and draft: the agent gathers information or writes a draft that no one outside the company sees. Usually no gate, but log the source of each fact it uses.
- Internal write: the agent changes a record, creates a task, or updates a status inside your systems. Gate it when the change affects pay, inventory, scheduling, or a customer file.
- External send: the agent contacts a customer, supplier, regulator or the public. Gate every first message that uses a new template, then sample later messages on a fixed schedule.
- Money and access: the agent commits funds, issues a refund or credit, grants a permission, or changes who can use a tool. Gate every instance, and require a second approver for the largest amounts or the broadest grants.
As the agent's capability grows, some of these thresholds should move. That change belongs in a written review, not in a quiet edit to the tool configuration.
A step-by-step plan for placing your gates
1. List every tool the agent will hold, with the exact actions each tool exposes. A tool that can send, delete or pay is split into separate actions, because each one carries a different risk. 2. Assign each action to one of the four classes. Where two people disagree, choose the stricter class and write down why. 3. For each gate, write the condition that triggers it, such as an amount, a recipient outside the company, or a record type. Vague triggers like anything risky do not hold up under real work. 4. Name the approver and a backup by role and by person. Confirm that each one has enough context to decide without rereading the whole case. 5. Define what happens during the wait. The agent should hold the action, keep the request in a visible queue, and continue unrelated work where that is safe. 6. Define the response options: approve, reject with a reason, or edit and approve. Record every decision with the time, the approver and the reason given. 7. Run a rehearsal in which an approver rejects one action. Confirm that the agent stops, the queue shows the item, and the record is complete.
Who should hold the approval
The approver should be the person accountable for the outcome, not the person who built the agent. For a refund, that is usually the finance or customer lead. For a message to a regulator, it is the compliance owner. For a change to who can use a tool, it is someone who can see the full permission picture. Engineers can confirm that a tool call is well formed, but they should not be the only approvers of business decisions they do not own.
Name a backup, and make sure the backup is not the person who requested the action. Gates that only one person can clear will be bypassed during holidays and busy weeks, and the bypass is usually the part nobody records. For the largest or broadest actions, require two approvers. The guide on two-person approval for AI agents explains which actions justify that extra step.
What an approval request must show
A person can approve responsibly only when the request is complete. A request that says the agent wants to issue a refund forces the approver to go hunting for context. Build every request so it can be decided in about a minute, and let the approver open the underlying records when the stakes justify it.
- The action in plain words, with the exact system and record it touches.
- The triggering rule that sent it to a person, so the approver knows why they are seeing it.
- The evidence the agent used, with links to source records rather than only its summary.
- The amount, recipient or scope, stated once and never inferred.
- The expected outcome and how to reverse it if the decision turns out to be wrong.
- The options available: approve, reject with a reason, or edit and approve.
Illustrative scenario: a clinic's scheduling agent receives a request to waive a late cancellation fee for a patient who has cancelled several recent appointments. The rule sends every fee waiver to the front desk lead. The request shows the fee amount, the cancellation history pulled from the booking record, and the clinic's written waiver policy, which allows one waiver per patient each year. The lead rejects the request because this patient already received one, and the agent logs the rejection and tells the patient that a staff member will call back. No one had to read the full booking history, and the rejection becomes evidence for the next review of the rule.
How Be Human runs approval gates in managed operations
Be Human designs the gate map during the first engagement, alongside the one role the agent will perform, the memory and tools it needs, and the observable success criterion that shows the role is working. Managed operations keep that map current: reviewing triggers when tools change, confirming that named approvers are still in post, rehearsing the rejection path on a schedule, and bringing any new action class to the business owner before an agent gains it. Approvals, rejections and pauses are written to a log the company can read, so the people who own the business keep control. Each gate can be reviewed, changed or removed by a person without rebuilding the agent.
Take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How many agent actions should require human approval? Only the actions that are costly, external, hard to reverse, or that widen the agent's access. Gating every step tends to produce approvals that people click through without reading, which is worse than fewer gates that get real attention.
Can an agent approve its own actions if it is confident it is right? No. Confidence is useful input for the reviewer, but the approval has to come from a named person outside the agent's own process, recorded in a log the company controls.
What if a small company has only one person who can approve? A single owner can hold the gates, but the company still needs a named backup, even if that backup is a trusted manager who can decide during the owner's absence. Without a backup, the gate will eventually be bypassed when the owner is unavailable.
How do we know whether our approval gates are working? Review the rejection log, check how long requests wait, and look for actions taken outside the gate path. Rehearse the rejection path on a schedule, and read the blocked action history for patterns that suggest a gate is too loose or too strict.
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.


