Approving AI actions: making the decision meaningful
Design AI action approvals around a specific version, an authorised decision, expiry, rejection and evidence of what happened afterward.
Article language: English

In this article
“Approve” is a short label on a button. Knowing exactly what was approved takes a well-designed process. Human approval of an AI action matters when the person can see a specific proposal, understand its consequences and decide on that particular version. The button will not help if the recipient, attachment or scope of a change can be altered after the click. There needs to be a connection between what the person saw, their decision and the operation the system later performs.
Start with one action: sending a message, changing a record or sharing a document. Decide who may make the decision, what they need to see, when approval stops being valid and how the result will be confirmed. This workshop concerns process design. It does not assume that every small action needs a separate click; it helps identify boundaries that really need an accountable decision.
Separate preparation from an action that has an effect
In a hypothetical company, an assistant might prepare a draft reply to a request. The draft alone is not permission to send it to the customer. Likewise, a proposal to complete a record need not immediately change data in the system. This separation lets someone inspect the result before an action has an effect beyond the working conversation.
The OpenAI Agents SDK describes a human-in-the-loop mechanism (in English): a run pauses for a human decision. Its documentation distinguishes a pending request, approval or rejection, and resumption. This is one example of a technical approach. Your process still needs its own rule for which operations cross that boundary.
List actions according to their actual effects. Reading public instructions, saving an internal draft and sending a document to an external recipient are not equivalent. In a responsible agent workflow, decide which actions fall within permissions already granted and which need a decision for the particular case. For related guidance, see the article on responsible agent workflow (in Polish).
Do not replace a control with a prompt telling the model to “always ask”. The component that performs the action should be able to check for the required decision. Otherwise a model error may bypass a condition that existed only in a description of expected behaviour. Acceptance depends on the actual execution path, not a polite assurance from the assistant.
Show the person what they are deciding
For a message, this means at least the intended recipients, content and attachments needed for review. For a record change, show the object, previous value, proposal and significance of the change. The information needed varies by task, but a generic “Do you approve this action?” usually leaves too much to guess. For an email reply, include a check of the correct thread and actual recipient list, not just the visible message text. The Polish Signal article “check of the correct thread and actual recipient list (in Polish)” discusses this topic.
Do not make someone search a long conversation for an important condition. Provide a concise summary of the operation and a way to inspect the details. At the same time, do not hide a decision-relevant difference in the summary. If an attachment has changed version, show that where the person reviews the message.
Microsoft describes an approval flow in Power Automate (in English), including a request, a decision and subsequent steps. It is a useful reference for designing a flow, not proof that your particular configuration is correct. Check both the request content and the condition that starts the next action.
If the proposal relies on data read from a scan, show material uncertainties. Checking OCR before AI may be needed before a decision. A person should not approve an effect on the strength of a number the system cannot reliably connect to its source. See also the article on Checking OCR before AI (in Polish).
Bind the decision to a version and return changes for review
Record which elements belong to the proposal being approved. If someone changes the content, recipient or another significant parameter after approval, the system should apply the agreed rule for another review. An earlier click must not become open-ended permission for arbitrary later additions to the task.
In a test, show the person version A, then deliberately change an important element before execution. Check whether the operation is stopped or returned for a decision. Do not judge only whether the approver's name was recorded. A name without a connection to the specific version does not explain what was actually approved.
Set an expiry where the state of the case may change. A request left waiting can concern an outdated document, an absent person or a closed ticket. Before resuming, check whether the conditions for execution still hold. The process owner sets the validity period for the task. With an AI calendar proposal, recheck people's and room availability before booking: a previously free time does not establish that the slot is still available. This is also discussed in the Polish Signal article “AI calendar proposal (in Polish)”.
AI activity logs should make it possible to connect the proposal, decision, execution and result. That does not mean copying all content into every entry. You need appropriate identifiers, versions and access to evidence in a location governed by agreed safeguards. See the Polish Signal article “AI activity logs (in Polish)” for related guidance.
Rejection and no response are outcomes too
After rejection, define the next step: revise the proposal, close the case or pass it to someone else. The model should not repeatedly submit the same request until someone clicks a different answer. If a comment is required, it should serve the task rather than merely fill another field.
No response is not approval. Arrange a reminder, a substitute or an end to the wait. The person taking over should see the status and the scope of their authority. Sending a second notification to a wider group is not enough if nobody knows who can actually decide.
Test reopening an old request. Are rejection and expiry still visible? Does a later click avoid triggering an outdated action? Check each channel used, such as a panel, message or app. The same flow can have several entry points that need consistent handling.
Describe the rules in the AI use policy (in Polish) and the instructions for the specific task. A general policy says who decides and when; task instructions show what that person needs to see for the operation. Do not make employees reconstruct the entire ruleset every time they review a message.
Confirm execution after approval
Approval does not yet mean that a message was sent or a change saved. The subsequent operation can fail or end with an unknown outcome. Distinguish the human decision from execution in both the interface and the log. Otherwise a team may consider the case complete merely because a button changed colour.

Test an interruption after the decision but before the effect is confirmed. Retries without duplicate actions (in Polish) need a separate check. The system should recognise the same approved operation and establish what has already happened at the destination before trying it again.
Include pending, partial and manually taken-over tasks in the automation audit (in Polish). An approval should not lose its meaning when a process restarts. If state is stored, check that after resumption it still refers to the right version and person.
Prepare a business continuity plan (in Polish) for an unavailable tool. A person should be able to establish whether a decision was made, whether the effect occurred and what can still be done. The manual procedure must not accidentally bypass the boundary protected in the automation.
Keep the scope of the decision when assigning a substitute
An absent approver does not mean every notification recipient can automatically stand in. Define the substitute's role, access to the material and decision limits. If the organisation requires a different level of approval for a particular operation, handing over the case should respect the same rule.
Test with a primary approver and a substitute. Check who can see the request, who can respond and what happens after the first decision. Two active notifications should not lead to two independent executions. The result must refer to the same version of the operation and take its current status into account.
If several people must approve, state the rule: everyone, one person from a group or decisions in a set order. Agree on the effect of a refusal as well. Do not leave these matters to an incidental tool setting. A technically available option may not reflect how the company actually assigns responsibility.
Record how an operator recognises that a task is waiting for a decision and how they can check its status. A clear status reduces the chance of restarting the whole task simply because someone missed a request. People need a short indication of the next action, not an internal diagram of every dependency.
Group similar decisions without hiding differences
Someone receiving many small requests may start approving them reflexively. Before adding another checkpoint, decide which error it is meant to catch and whether the person has enough information and time. The number of confirmations alone does not measure oversight quality. An interface can make clicking easy while doing little to help someone decide. Sometimes the operation needs a better boundary, not another button.
For a batch approval, show its contents and make exceptions recognisable. Several messages may have different recipients and attachments. A shared summary must not erase those differences. If the whole batch is hard to assess properly, divide it by task or significance of the effect.
Decide what happens when one item is rejected. Can the others proceed, or does the whole batch return for correction? The answer depends on the process. The approver needs to know what their choice does and be able to see the result afterward rather than infer the rules from a button label.
In a small exercise, put one deliberate error and one item needing clarification into a batch. Check whether the reviewer notices them and can handle them differently from the rest. If the interface only offers an all-or-nothing choice, assess whether that boundary is justified and clear to the user.
Limit notification data while keeping access to evidence
A decision request may be sent by email or messenger. It need not contain a complete document copy if secure access to the right version is available elsewhere. Agree on the minimum notification content and how to reach the details. Copying everything to every channel can make information harder to manage later.
Test the notification on a phone and from the intended person's account. Can the recipient recognise the case, open the right version and return to the decision? Must they ask for additional access halfway through? Access should be prepared according to role, not expanded accidentally for one convenient test.
After a request changes or expires, check the older message too. It should lead to the current status of the same operation or clearly state that a decision is no longer available. Do not leave an active shortcut that executes an old version simply because someone kept the notification in their inbox.
Assess the quality of the decision, not just the click
Ask an authorised person to evaluate a proposal using the material they will actually see. Can they identify the recipient, version, effect and a significant limitation? If they need another message from the automation's author, improve the request. The button may work technically while the process still provides no sound basis for a decision.
Rules for integration permissions (in Polish) should separate the ability to prepare, approve and execute when your operating model requires it. In team training (in Polish), practise with both a correct proposal and a deliberate error. A habit of accepting is not, by itself, a useful control.
Also distinguish the source of a request from the person authorised to decide. A document, customer message or note can suggest an action, but should not itself grant approval. Put such an instruction in controlled test material and check that the system still routes the operation through the proper flow. You are assessing the process boundary without using real protected data.
Include one successful flow and one stopped after a version change in the report. The approver should be able to understand why the first was executed and the second required another decision. If the explanation depends solely on an administrator's memory, improve the evidence trail and status display.
Finally, ask about scope after a correction is accepted. Does approval cover only this case, or similar future operations too? Both approaches need a clear rule. A one-time approval must not accidentally become a lasting permission; nor should someone have to reconfirm an action already within a scope they knowingly granted.
To explore the service, see Aeneia Audit on the main site.







