Back to all articlesAI action safety

Set Expiry Windows for Approved AI Actions

Alactic Editorial10 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

A manager approves an appointment change, but the queue does not process it until much later. The requested change is identical, and the approval is genuine. Yet the customer may now be preparing for the original appointment. Executing the old decision can be worse than asking again.

Set an approval expiry window according to how quickly the action’s business assumptions can become stale. Then check that window immediately before execution, alongside any required freshness checks. If the approval has expired, stop the action and route it for a fresh decision; do not treat queue delays or retries as reasons to extend it.

An AI action approval expiration policy should distinguish three things: the age of the approval, whether its underlying assumptions still hold, and whether the action actually executed. The completed hypothetical worksheet below shows how to make those distinctions for an internal note update, an appointment change, and an external message.

1. Identify what time can invalidate

Start with the reason the action was approved, not its technical category. “Update a record” says little about how long approval should remain valid. Adding a historical summary is different from adding a note that tells a colleague what to do next.

For each queued action, finish this sentence: This approval makes sense only while…

  • Internal note update: the note still gives the next handler a useful description of the case.
  • Appointment change: the proposed time remains suitable and there is still enough notice for the people affected.
  • External message: the recipient still needs the communication and its timing remains appropriate.

Next, identify the earliest plausible event that could undermine that sentence. A case may close, an appointment may get too close to its start time, or someone may already answer the customer’s question. None requires the approved payload to change.

This article addresses approval age, not changes to permissions or the approved payload. Those need their own controls. A time-valid approval does not establish that every other condition for execution is satisfied.

2. Choose a window from assumption life, not queue convenience

Record the expected wait, but do not let it dictate validity. A busy queue is evidence of an operating constraint, not evidence that an older decision remains sound.

Consider four inputs when choosing a window:

  1. Assumption volatility: How quickly can the reason for acting disappear?
  2. Consequence of late execution: Would delay create minor cleanup, conflicting instructions, or an unwanted external commitment?
  3. Observability: Can the executor check the relevant business facts just before acting?
  4. Recovery difficulty: Can someone correct the result without creating further confusion?

Use these inputs to choose both a maximum age and an execution-time check. A longer window can be reasonable for a lower-consequence internal action whose assumptions are easy to inspect. A shorter window may be appropriate when timing is central to the decision or the consequence leaves the organisation.

Do not make “internal” synonymous with harmless. An internal note can misdirect a colleague. Equally, not every external message needs the same window: a time-sensitive scheduling message and a routine acknowledgement have different assumptions.

All durations below are hypothetical operating choices for a fictional team, not recommended universal defaults. The fictional team has selected them to illustrate a decision method, not to establish safe limits for other businesses.

3. Completed hypothetical expiry worksheet

Fictional worked example: A service team uses an AI assistant to prepare actions for human approval. Approved actions enter a queue. A designated operations owner handles expired items. The team records timestamps in one consistent time standard and displays the local time to reviewers.

Worksheet entry A: Internal note update

  • Approved action: Add a summary of an unresolved service request to the internal case notes.
  • Approval time: Monday, 09:00.
  • Assumed waiting period: Four hours; the scheduled worker is expected to attempt execution at 13:00.
  • Assumption that can become stale: The case is still unresolved, and the summary still helps the next handler.
  • Chosen validity window: Eight hours from approval, ending Monday at 17:00.
  • Execution-time revalidation: Check whether the case has closed or later activity has made the summary misleading or redundant.
  • Expiry decision: At 13:00, execute only if the freshness check passes. At or after 17:00, hold the update even if the case remains open.
  • Reapproval path: The case owner reviews the latest activity. If the note remains useful, they grant a new approval with a new timestamp; otherwise they cancel the queued update.
  • Reason for the choice: The team expects the summary to remain useful through part of the working day, but does not want an unattended note appearing after later case handling.

Worksheet entry B: Appointment change

  • Approved action: Move a fictional customer’s appointment from Tuesday at 15:00 to Wednesday at 11:00.
  • Approval time: Tuesday, 09:00.
  • Assumed waiting period: Twenty minutes, with execution expected at 09:20.
  • Assumption that can become stale: The alternative slot remains available, and changing the appointment still gives adequate notice.
  • Chosen validity window: Ninety minutes from approval, ending at 10:30, subject to a separate hypothetical notice cutoff of 10:00. The earlier boundary governs.
  • Execution-time revalidation: Check current appointment status, alternative-slot availability, and whether the notice cutoff has been reached.
  • Expiry decision: An attempt at 09:20 may proceed if those checks pass. An attempt at 10:05 must stop because the notice cutoff has passed, despite the ninety-minute maximum age not yet being reached.
  • Reapproval path: A scheduler contacts the customer through an appropriate authorised process to confirm whether a change is still wanted. The scheduler then approves a fresh action or retains the original appointment.
  • Reason for the choice: A relative approval lifetime alone would ignore the approaching appointment. A calendar-based cutoff captures that risk.

Worksheet entry C: External message

  • Approved action: Send a prepared follow-up asking a customer to confirm a preferred callback time.
  • Approval time: Monday, 14:00.
  • Assumed waiting period: Ten minutes, with execution expected at 14:10.
  • Assumption that can become stale: The callback preference is still unknown and the conversation has not already moved on.
  • Chosen validity window: Thirty minutes from approval, ending at 14:30.
  • Execution-time revalidation: Review recent conversation activity for an answer, a completed callback, or another follow-up that makes this message unnecessary.
  • Expiry decision: At 14:10, send only if the message remains relevant and separate channel requirements are satisfied. At or after 14:30, hold it for review. If the customer answered at 14:05, stop it even before expiry.
  • Reapproval path: The conversation owner reviews the latest exchange and either approves a fresh follow-up or cancels it. Unclear cases go to a human rather than repeated automated attempts.
  • Reason for the choice: A redundant external message is visible to the customer and may require an explanation; the team chooses a shorter window than for its internal note.

The worksheet’s key distinction is that expiry and failed revalidation are different stop reasons. An expired action may still be useful but needs a new decision. An unexpired action may already be inappropriate and should not execute.

4. Turn the worksheet into an execution rule

Store the approval time, selected validity window, any earlier business cutoff, and the resulting expiry time with the action. Also record the assumptions to check and who owns a stopped item. These are recommended workflow fields, not claims about a particular product’s controls.

A simple rule is:

expiry_time = earlier of:
  approval_time + validity_window
  any applicable business cutoff

Immediately before execution:
  if current_time >= expiry_time: hold as expired
  otherwise, check required business assumptions
  if a check fails or cannot be completed: hold for review
  otherwise, proceed through the authorised execution path

Define the boundary explicitly: in this design, an action is no longer valid at its exact expiry time. A reviewer should not have to guess whether “valid until 14:30” includes an attempt made at 14:30.

Check close to the consequential operation, not only when a worker first picks up the job. A worker can pass an initial check and then spend time waiting on another dependency. If that wait crosses the boundary, it must stop before acting.

A successful freshness check does not renew the approval. Only a new authorised decision creates a new approval timestamp. Otherwise, a queue could keep an old decision alive indefinitely by repeatedly confirming a narrow condition.

5. Keep action expiry separate from messaging windows

For an external message, an internal approval window answers, “Is this business decision still current?” It does not answer every question about whether the channel may be used.

For WhatsApp, recipient permission matters and opt-outs must be honoured. Approved templates are used to initiate Platform conversations and outside the customer-service window, which opens or resets with a user message and lasts 24 hours. Check the current WhatsApp Business Messaging Policy for precise requirements.

Do not copy that channel window into the action’s expiry policy. The fictional callback message expires after thirty minutes because its purpose may become stale, not because of a WhatsApp timing rule. Conversely, a still-current internal approval does not override channel requirements.

Keep campaign preparation, template review, approval of a particular action, and actual delivery as separate states. Template approval is not a guarantee of delivery, and an approved queued message is not evidence that it was sent. Provide a clear human escalation path when automation cannot resolve whether to proceed.

6. Handle retries without reviving old decisions

A retry should inherit the original approval time and expiry. Requeueing, restarting a worker, or recovering from an outage must not reset the clock.

Before retrying, establish what happened in the previous attempt:

  • Known not executed: Retry only if the original approval remains valid and the execution-time checks pass.
  • Known executed: Record the result and do not repeat the action.
  • Outcome unknown: Hold for reconciliation. Expiry does not tell you whether the previous attempt succeeded.

If an outage leaves a backlog of expired actions, do not bulk-refresh their timestamps to clear the queue. Group them by owner and consequence, then review which still serve a business purpose. Some should be reapproved; others should be cancelled because the opportunity has passed.

For actions with several consequential steps, define the boundary of each approval. An appointment update and its customer notification may need separate validity decisions. If one step completes and another expires, route the partial result to an owner rather than silently presenting the whole workflow as complete.

7. Make reapproval a fresh decision, not an unblock button

A useful review screen should explain why the item stopped, when approval was granted, when it expired, and what relevant business activity occurred while it waited. Give the reviewer distinct choices: cancel, approve again, or send for further handling.

Do not preselect reapproval or hide the waiting period. The reviewer needs to decide whether the original purpose still holds, not merely acknowledge a technical interruption.

Record the new approval separately from the old one. Preserve the earlier expiry reason so the team can distinguish genuinely reconsidered actions from repeated queue failures. If an assumption check is unavailable, state that uncertainty instead of treating missing information as confirmation.

Repeated expiry often signals a mismatch between the workflow and the decision’s useful life. The remedy may be faster routing, approval closer to execution, or cancellation of low-value actions—not a longer validity window.

8. Test one queue and measure stale execution

Start with one action type and complete the worksheet using its actual assumptions. Name an owner for expired items before enforcing the rule, so stopping unsafe work does not create an invisible abandoned queue.

Test an attempt just before expiry, one at the boundary, one after expiry, a retry after a delay, an unavailable freshness check, and an unknown previous outcome. Confirm that none silently grants a new approval.

During a defined review period, measure:

  • Approval-to-execution age, including the oldest completed actions.
  • Items stopped for expiry versus items stopped for failed freshness checks.
  • Expired items later cancelled versus reapproved.
  • Time spent waiting for human review.
  • Actions that executed despite an expired approval, with a target of none under this design.

Use the cancellation reasons to adjust the workflow. Many expired items later cancelled suggests the queue is outlasting the decisions. Frequent reapproval with unchanged assumptions may justify reviewing the chosen window, but only alongside the consequences of late execution. The next practical step is to fill one worksheet, test its boundary cases, and inspect every expired item from the first review period before expanding the policy.

AI action safetyApproval workflowsQueue managementBusiness operations

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications