Back to all articlesAI action safety

Decide Which Changes Invalidate an AI Action Approval

Alactic Editorial9 min read

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

A manager approves an AI-prepared CRM update. Before execution, one target record is replaced, a proposed value changes, and the preview is sorted differently. Those differences should not all receive the same treatment.

Renew approval when a change alters the targets, intended effects, or business conditions on which the reviewer relied. Preserve approval for a presentation-only change that leaves the action unchanged. Block execution when a required condition fails; another approval cannot make an ineligible action eligible.

The practical task is to define those boundaries before a mismatch occurs. This guide covers changes between review and execution, not approval expiry, changes in user permissions, or general data freshness practices. Its examples and proposed controls are fictional workflow designs, not claims about available Alactic capabilities.

1. Define the action the reviewer is approving

An approval attached only to a conversation or preview screen is ambiguous. The reviewer may approve specific updates while the executor interprets that approval as permission to rerun a selection rule against whatever records match later.

Instead, create an approval snapshot: a stable description of the reviewed action. Include:

  • Operation: the action to perform, such as updating a field.
  • Targets: exact record or recipient identifiers, not just names or counts.
  • Proposed effects: values to write, or message content and destinations.
  • Dependencies: current values and business conditions that justify execution.
  • Batch meaning: whether order matters, whether items must succeed together, and whether partial execution is permitted.
  • Decision context: totals or relationships that materially shaped the review.

Capture the facts that make this action the one the reviewer intended, rather than copying every field. A decorative profile image normally does not matter. A changed communication destination does.

State the batch rule explicitly. Approval for three independent updates is different from approval for one indivisible update affecting three records. That distinction determines both the execution mechanism and the recovery options.

2. Separate approval decisions from execution guarantees

Use three outcomes for drift:

  • Preserve approval: the difference is presentation-only or covered by a narrow, documented equivalence rule.
  • Require fresh approval: the candidate may be valid, but it differs materially from the reviewed action.
  • Block the candidate: an eligibility condition fails, the target is uncertain, or a required safeguard cannot be enforced.

These are approval decisions, not transaction guarantees. Invalidating approval for a whole batch means the old approval cannot authorise the revised batch. It does not establish that no writes have already occurred.

Likewise, conditional writes and atomic batch execution solve different problems. A conditional write can prevent one record from being updated after its expected starting state changes. But earlier records may already have been updated when a later record fails its condition.

An all-or-nothing batch requires a mechanism that enforces the relevant conditions and commits the entire batch atomically. If that mechanism is unavailable, hold the batch or obtain approval for a separately defined partial-execution workflow. Do not quietly replace the approved batch rule with best-effort updates.

3. Use this filled hypothetical drift decision table

Fictional worked example: A sales operations manager reviews updates to R-041, R-052, and R-063. Each has an outreach status of “Needs review.” The proposal changes each status to “Ready for outreach,” provided campaign membership and outreach eligibility remain unchanged.

This is CRM preparation only; it does not authorise messaging. The manager requires all-or-nothing execution. The design therefore requires an atomic mechanism covering the whole batch and its relevant conditions. Without that capability, this batch must remain on hold.

The first four rows below concern that CRM batch. The final row concerns a separate hypothetical messaging action.

ChangeReviewed stateExecution candidateDecisionReasonRecovery
Changed target recordsR-041, R-052, R-063R-041, R-052, R-074Require fresh approval for the revised CRM batch.An unchanged count conceals a different target.Show R-063 removed and R-074 added, with R-074’s dependencies, then review the revised snapshot.
Changed proposed valuesSet R-052 to “Ready for outreach.”Set R-052 to “Contacted.”Require fresh approval.The candidate asserts a different business state.Correct the proposal or present the changed value in a revised batch.
Changed current valuesR-063 is “Needs review.”R-063 is “Do not contact.”Block the entire CRM batch.A required eligibility condition fails; the approved batch cannot proceed.Investigate the change. Prepare a new eligible batch without overwriting the newer state.
Reordered display rowsR-041, R-052, R-063 displayed by identifier.The same records appear in reverse display order; effects and dependencies match.Preserve approval.Display order has no operational meaning in this example.No new review; retain exact identifier-to-value mapping.
Newly added recipients: separate messaging exampleA send addressed only to P-011 and P-012.P-013 is also included.Require fresh approval for the expanded send, subject to eligibility checks.A new person would receive a communication.Remove P-013 or review a new send proposal. Block an ineligible candidate rather than treating approval as permission to bypass eligibility.

To reuse this artifact, replace the fictional identifiers, values, dependencies, and recovery instructions with one real action type. Keep preparation and communication separate: approval to mark a record ready is not approval to contact it.

4. Compare meaning without hiding material differences

Comparing raw screen text creates unnecessary interruptions. Comparing only record counts misses changed targets. Use stable identifiers and field-specific comparison rules.

Compare target sets when execution order is irrelevant. Compare ordered sequences when order affects dependencies or results. A reordered preview is harmless only when it does not alter execution order or assign values to different records.

Next, compare proposed effects. In a hypothetical numeric field, “10.0” and “10.00” might mean the same value. That does not justify stripping punctuation or whitespace from message text, where those differences may affect meaning.

Do not use approximate name matching to preserve approval. Two records with the same name are not necessarily the same target. Conversely, an unchanged contact identifier with a changed messaging destination can still represent a materially different action.

Keep equivalence rules inspectable. Record which rule accepted a difference and what it normalised. If an operator cannot explain why two candidates are equivalent, hold the action rather than treating uncertainty as approval.

5. Identify the current values that approval depends on

The proposed output can remain identical while the reason for applying it disappears. In a hypothetical assignment workflow, a reviewer approves changing an owner from “Unassigned” to a team. Another operator then assigns the record elsewhere. Writing the original proposal would erase that newer decision.

For each action type, identify the fields that justify the change, the values it would overwrite, and related conditions that make it eligible. Also identify changes that are irrelevant. In the CRM example, status and campaign membership are dependencies; an unrelated display preference is not.

Avoid making every field a dependency. Excessive review noise encourages rubber-stamping. But do not exclude a field merely because it changes frequently if the action genuinely relies on it.

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

Internal approval, campaign preparation, template review, and actual delivery are separate states. Template approval does not guarantee delivery. Any automated messaging workflow also needs a clear human escalation path.

6. Enforce the approved conditions at commitment

A preview-time comparison cannot protect against a change made afterward. Revalidate the candidate where its effects are committed, using a mechanism appropriate to the approved batch semantics.

  1. Retrieve the approved snapshot and identify the exact execution candidate.
  2. Compare target membership, proposed effects, and declared dependencies.
  3. Classify differences as preserve, renew, or block.
  4. Enforce the relevant conditions during commitment, not only in a preceding read.
  5. Record the actual outcome against that specific candidate.

For independently approved updates, per-record conditional writes may be appropriate. For the all-or-nothing CRM example, they are insufficient by themselves: R-041 and R-052 could succeed before R-063 fails.

The required mechanism must check the relevant conditions and commit all approved updates as one atomic unit, or commit none. A preliminary check followed by unconditional writes does not satisfy that requirement. Nor does attempting to undo earlier writes after a failure establish atomic execution; intervening changes or failed reversals can leave a different result.

If the available system cannot provide the required guarantee, stop before starting. A manual recheck can inform an operating decision, but it is not equivalent protection. Verify availability in the relevant workspace and current plan rather than assuming a recommended control exists.

Authorisation remains independent. OWASP’s object-level authorisation guidance calls for checking access to each requested object; knowing its identifier is not enough. A matching approval snapshot does not replace that check.

7. Recover using observed outcomes, not assumptions

A drift notice should identify the changed fact, explain its significance, and report only the execution outcome actually established.

For the fictional CRM batch, the following notice is appropriate only after a confirmed atomic rejection: R-063 changed from “Needs review” to “Do not contact.” The atomic batch was rejected; no records were updated. The no-update statement depends on the enforced mechanism and confirmed outcome, not merely on the all-or-nothing policy.

If a non-atomic executor has already updated R-041 and R-052 before detecting R-063’s change, report those completed writes. Treat that as a failure to meet the approved batch rule, not as a successful all-or-nothing stop. If the outcome is uncertain, reconcile execution records before retrying or claiming that nothing changed.

A separately approved partial-execution workflow should define which items are independent, which conditions block individual items, and what batch-level changes require renewed review. After execution, distinguish completed, blocked, failed, and unattempted items.

Do not replay completed actions merely because fresh approval is needed for the remainder. Build a new candidate from the remaining work and check its conditions again. Any revised approval must bind to that revised snapshot, not to an open-ended instruction to accept future changes.

8. Test the rule sheet before expanding the workflow

Start with one action type and the filled table above. For a hypothetical pilot, prepare ten cases: the five table scenarios, an unchanged action, an irrelevant field change, a changed destination, an unresolved target, and a partial-execution recovery. Specify expected decisions before running them.

Add an execution-boundary test in which a dependency changes after comparison but before commitment. For the all-or-nothing design, include a later-record failure and verify that earlier records were not committed. A preview rejection alone does not test atomicity.

Record three observable outcomes:

  • Decision agreement: did the workflow preserve, renew, or block as specified?
  • Execution containment: did commitment respect the approved per-record or atomic batch semantics?
  • Recovery clarity: could the operator establish what completed, what stopped, and what needed review?

Investigate mismatches before expanding the workflow. The useful deliverable is an agreed rule sheet supported by recorded outcomes—not an approval button whose relationship to execution remains undefined.

AI action safetyApproval workflowsCRM operationsChange control

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