Approval Prompts Business Agents Should Show a Human Reviewer
A practical list of what an agent approval prompt must show a human reviewer: the exact action, source facts, risk level and undo path.

An approval prompt for a business agent should show the reviewer four things before any consequential action runs: the exact action the agent wants to take, the source facts behind the proposal, the risk if the proposal is wrong, and the undo path if the action must be reversed. A prompt that only asks "Approve this step?" turns approval into a habit rather than a decision. The sections below explain what each element must contain and how to check that it works.
Key takeaways
- An approval prompt is a decision record, not a button. A reviewer should be able to reconstruct what will happen and why from the prompt alone.
- Show the exact action with its parameters, targets and scope, not a summary such as "update the customer."
- List the source facts the agent relied on, and mark which ones it observed and which ones it inferred or could not verify.
- State the risk in terms of who or what is affected and whether the effect can be reversed, not as a single word like "high."
- Attach an undo path, or state plainly that none exists, so the reviewer knows what approval commits the company to.
Why a yes or no question stops working as agents improve
A simple confirmation dialog is enough when an agent drafts a reply that a person then sends. It stops being enough when the agent chains several tools, updates records, contacts customers or changes schedules on its own. Each approval request then carries more consequence, and reviewers are asked to judge more of them in less time.
Capability makes this harder, not easier. A more capable agent can produce a proposal that reads clearly, looks complete and is still wrong in one small but important detail, such as the wrong recipient list or a record that was superseded last week. A reviewer who sees only the agent's conclusion has little to check it against. The prompt therefore has to carry the evidence, not just the recommendation.
The NIST AI Risk Management Framework treats human oversight and accountability as part of managing AI risk across the life of a system, not as a single checkpoint at launch. An approval prompt is one of the places where that oversight becomes concrete. If the prompt is vague, the oversight is nominal, even when a person clicks the button.
Show the exact action, not a summary
The first element is the action itself, written so that a person who was not in the conversation can understand it. Include the following:
- The system and the operation, such as send, update, delete, publish, reschedule or pay.
- The target, named by an identifier a person can look up, such as an order number, a contact record or a channel name.
- The content or parameters. Show short content in full and show long content as a comparison against the current version.
- The scope: how many records are affected, which recipients receive it, and over what period.
- Whether the action runs once or repeats on a schedule or trigger.
Illustrative scenario: an agent proposes to move a supplier delivery. A weak prompt reads "Reschedule delivery?" A useful prompt reads: move order 4471 from Tuesday to Friday in the shared supplier calendar, send the change notice to the contact named in the order record, and leave the price unchanged. The reviewer can now check each part against the order record in under a minute.
Attach the source facts behind the proposal
The second element is the evidence. Show the reviewer which inputs the agent used: the message, the record field, the policy paragraph or the retrieved document that supports the proposed action, along with when it was retrieved. If the agent summarized a document, show the passage it relied on, not only its paraphrase. Retrieval-based designs make this practical because the system already knows which passages were pulled into context.
Separate observed facts from inferences. "The order record lists Friday as the confirmed delivery window" is an observation the reviewer can verify. "The supplier is probably unavailable Tuesday" is an inference, and the prompt should label it as one. Flag anything the agent could not confirm, because an honest gap is more useful to a reviewer than a confident sentence that hides it.
The source list must also show where any instruction came from. If the agent acted because an incoming email asked it to, the reviewer needs to see that email, since untrusted content can contain text designed to steer an agent. The OWASP guidance on prompt injection describes this class of risk. For a deeper treatment, see Prompt Injection Through Email: Defenses for Superintelligent Agents.
State the risk in plain terms
The third element is the risk, stated so that a reviewer can disagree with it. A single label such as "medium" tells the reader almost nothing. Instead, answer four questions in the prompt:
- Who or what is affected: a customer, an employee, a supplier, a financial balance, a public page.
- What the realistic worst outcome is if the proposal is wrong, described in one sentence.
- Whether personal or sensitive data is involved, which changes the review path under the company's data obligations.
- Whether the action touches a rule the company has already decided never to delegate.
Then assign a tier that the company has defined in advance. A common approach uses three tiers: actions a reviewer can approve from the prompt alone, actions that need a second reviewer, and actions the agent may not propose at all. Defining those tiers before the first prompt appears keeps the reviewer from inventing criteria under pressure. For the high-risk tier, see Two-Person Approval for AI Agents: Which Actions Need It.
Give the undo path before the approval
The fourth element is the one most often missing. Before the reviewer approves, the prompt should state what reverses the action, who performs the reversal, how long it takes and what cannot be recovered. A calendar change can be moved back. A notice that has already reached a customer cannot be unsent. A record deleted without a backup may not return at all.
Write the undo path as a concrete step, not a promise. "Can be reverted" is not enough. "Restore the previous calendar entry from the change log within the same working day, then notify the supplier contact with the reversal text" is something a person can execute. Where no undo exists, the prompt should say so in those words, because the reviewer's approval then carries the full weight of a one-way action.
Retries deserve a line too. If the network drops and the agent resubmits the action, the prompt should show whether the operation is safe to repeat. Some providers accept an idempotency key so that a repeated request does not create a second charge or a second message. If the target system offers no such protection, the agent should not retry silently, and the prompt should say that a retry requires a new approval.
A step-by-step plan for building the prompt template
Company teams often start by designing the prompt around the agent's output. It works better to design it around the reviewer's questions. A workable sequence:
- Step one: list every action the agent is permitted to propose, and for each one name the system, the operation and the target type.
- Step two: for each action, write the four required fields: exact action, source facts, risk tier with reasons, and undo path.
- Step three: decide the fields that must be shown in full and those that can be collapsed behind a link to the underlying record.
- Step four: define what the reviewer may do with the prompt: approve, reject, edit the scope or escalate. A prompt with no edit option pushes reviewers toward either blind approval or constant rejection.
- Step five: test the template with the person who will approve. Give them three past proposals, one of them flawed, and check whether they catch the flaw from the prompt alone.
- Step six: log every decision with the prompt text, the reviewer and the outcome, so later reviews can see patterns. The record requirements are covered in Superintelligent Decisions Need Records: What to Explain Later.
The fifth and sixth steps are the ones that keep the template honest over time. Reviewers who never reject anything are a signal that the prompt is not showing them enough.
How Be Human handles approval prompts in managed operations
In a Be Human engagement, the approval boundary is defined in the first phase, together with the single role the agent will perform, the memory and tools it needs and the success criterion that will be observed. The team writes the prompt fields for each permitted action, connects the source passages and decision log to the same record, and tests the wording with the people who will approve. During managed operations, permission reviews and recurring evaluations check whether prompts still show what reviewers need, and changes to an action's scope go back through the same approval tier before they take effect.
Keep approvals meaningful as agents grow more capable
As agents take on more work, the risk is not only a single bad approval. It is a slow drift in which reviewers approve by pattern, because the prompts usually look fine. To counter this, sample a share of approved actions and check them against their sources after the fact. The right sampling rate depends on the risk tier and the volume of work, and it should be written down rather than left to memory. Watch for reviewers who approve a suspiciously high share of proposals quickly, and for prompts that repeatedly lack a source or undo path. Those are signals to revise the template or to reduce the agent's permissions for that action until the evidence improves.
Capability growth should also change what the agent is allowed to propose, not only how it asks. A first tier of actions can move to lighter review only after the team has seen the prompts hold up over real work. For a broader view of how trust is extended, see Superintelligent Agent Authorization: Growing Trust Tiers.
If you want to find the first task worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.
Frequently asked questions
What should an AI agent approval prompt include? It should include the exact action with its target and parameters, the source facts the agent relied on with any gaps marked, the risk tier with the reasons for it, and the undo path. Without these four elements, the reviewer is approving a conclusion rather than a decision. Each element should be written so that someone who was not in the conversation can check it.
How long should an approval prompt be? It should be as long as the reviewer needs to verify the action, and no longer. Short actions can be shown in full on one screen, while long content can appear as a comparison with the current version and a link to the full record. A prompt that hides the actual change behind a summary is too short, even if it looks tidy.
Do all agent actions need a human approval prompt? No. Low-risk, reversible actions inside a defined boundary can run without a prompt, provided they are logged and sampled later. The approval prompt is most valuable for actions that affect customers, money, personal data, public content or anything that cannot be undone. The company should decide these tiers in advance and review them as the agent's permissions change.
What if the reviewer keeps approving without reading? That is a design signal, not only a discipline problem. Check whether the prompts are too long to verify, whether they lack the source passages or whether the reviewer has too many approvals to make real decisions. Reduce the number of prompts by narrowing the agent's permissions, sample approved actions after the fact, and make rejection and editing easy so that the reviewer's judgment is actually used.
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.


