Keep an AI Preview Consistent with the Action That Executes
Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.
An AI assistant previews a CRM update for a reviewed set of contacts. The operator approves it. At execution, the system reruns the audience filter, finds additional contacts and updates them too. The filter is unchanged, but the action is not the one the operator reviewed.
To prevent this kind of AI dry run preview execution mismatch, approve a fixed proposal rather than a recipe for generating another proposal. Bind approval to the target identifiers, operations, proposed values, content and exclusions that were actually reviewed. Recheck current permissions and safety conditions before acting, but do not let those checks silently add targets, rewrite content or substitute new values.
The key distinction is between revalidating whether an approved action may still proceed and recalculating what the action should be. The former is necessary. The latter creates a new proposal that needs a new decision.
1. Decide whether the preview is a sample or an approval snapshot
A sample preview helps someone understand a proposed operation. It might show selected rows, representative messages or a total derived from a live filter. It is useful for exploration, but it does not necessarily describe everything that will execute.
A complete approval snapshot defines the entire proposed action. It includes all intended targets and the changes associated with each target. The interface can summarise that snapshot, provided the complete details remain available for review. Showing every row on a single screen is not required; having a determinate action behind the summary is.
Label the distinction explicitly:
- Sample preview: illustrative output; not sufficient to authorise execution.
- Approval snapshot: the complete, versioned proposal to which an approval applies.
A count alone is not a snapshot. Two different target sets can have the same count. Nor is a saved filter sufficient: the records matching it can change between review and execution.
If a team genuinely wants to approve a continuously changing rule, it needs a separate, explicitly bounded standing-authorisation design. A one-time preview approval should not be treated as permission to keep discovering new targets.
2. Fix the parts that determine what will happen
For each approval snapshot, capture enough information to distinguish the approved action from any materially different action. A practical minimum includes:
- Scope: the workspace, action type and destination system or object type.
- Targets: stable record identifiers, not just names, row positions or filter results.
- Proposed changes: the exact field-value pairs for each target, including whether a field is being cleared.
- Content: the final text, attachments and resolved personalisation associated with each target, where applicable.
- Exclusions: omitted records, exclusion reasons and rules that must remain effective.
- Preconditions: relevant current values or record versions against which the proposal was prepared.
- Approval binding: the proposal version, approving person and validity conditions.
For generated content, approving the instruction is not the same as approving the result. An instruction such as “write a short follow-up note” can produce different wording when run again. If the wording matters to the decision, store the reviewed output and execute that output rather than regenerating it.
The same applies to personalisation. A fixed message template can still produce different content if its variables are resolved later. Either resolve the values in the approval snapshot or specify which variables require a fresh preview when their source data changes.
3. Revalidate live conditions without silently rewriting the proposal
Freezing a proposal does not mean ignoring changes that make it unsafe or unauthorised. The executor should still check whether the approver and executing identity have the necessary authority, whether each target remains accessible and whether the action's preconditions still hold.
This distinction is consistent with OWASP's authorization guidance, which recommends checking permissions on every request and denying access by default. An earlier approval is not a reason to bypass current authorization checks.
Classify changing information into three groups:
- Fixed approved outputs: target membership, intended operations, proposed values and reviewed content. A change creates a revised proposal.
- Live eligibility checks: permissions, record existence, relevant status and other execution conditions. A failure blocks or excludes work according to an explicitly approved rule.
- Operational metadata: attempt identifiers, timestamps and progress tracking. These may change without new approval if they cannot alter the business action.
Be careful with apparently harmless refreshes. Refreshing an address label used only in a progress display may be operational. Refreshing a destination address used to send something changes where the action goes. Classify information by its effect, not its field name.
For changed preconditions, the default should be a visible stop or exception—not an AI-generated repair. Choosing a replacement owner, finding a substitute record or rewriting a note may be sensible, but each is a new proposal.
4. Filled fictional checklist: an execution blocked by new matches
Fictional worked example: All identifiers, names, values, counts and times in this example are hypothetical. A sales operations reviewer is preparing a CRM reassignment and internal handoff note. No customer message is being sent. The workflow described is a recommended design, not a claim about an available product control.
The preparation filter selects open inbound leads assigned to Queue North with no scheduled appointment. The reviewer approves a complete snapshot. Before execution, more records begin matching the same filter.
Preview-to-execution comparison checklist
- Proposal and approval: Snapshot
HANDOFF-042-v1, approved by reviewer Mira at 09:10. Execution candidate arrives at 09:14 with the same snapshot reference. Result: reference matches; contents must still be compared. - Action scope: Workspace
SALES-DEMO; update CRM lead owner and append an internal note. No outbound communication and no deletion. Candidate has the same scope. Result: pass. - Approved target identifiers:
L-410,L-411andL-412. These are the complete approved set, not sample rows. Candidate contains those records plusL-415andL-416. Result: fail; two unapproved targets. - Proposed values: For each approved target, change
owner_idfromQUEUE-NORTHtoUSER-27. Leave lead stage and all other fields unchanged. Candidate proposes the same field change, including for the extra targets. Result: approved rows match, but matching values do not authorise additional rows. - Reviewed content: Append the exact internal note: “Reassigned to Priya for an initial review. Check the existing conversation before contacting the lead.” Candidate contains that exact text. Result: pass for content; do not regenerate it.
- Approved exclusions:
L-413excluded because an appointment is already scheduled;L-414excluded because the lead is closed. Both remain absent from the candidate. Result: pass; exclusion checks do not override the target-set failure. - Record preconditions: Approved targets must still be open, assigned to
QUEUE-NORTHand without a scheduled appointment. Current checks confirm those conditions for the approved targets. Result: pass. - Current authorization: The executing identity is checked for permission to update each approved record. Checks pass in this hypothetical example. Result: pass; possessing the record identifiers alone would not be sufficient.
- Execution policy: Whole-batch execution only; any target-set mismatch blocks the batch. Candidate does not match the snapshot. Result: blocked before any writes.
- Reviewer-visible outcome: Approved targets: three. Candidate targets: five. Added:
L-415,L-416. Removed: none. Completed writes: none. Next decision: execute the original snapshot or review a revised proposal.
The block is not evidence that the new leads are unsuitable. It establishes only that they were not approved. If the reviewer wants them included, prepare a new snapshot showing their current values, proposed changes and relevant exclusions.
If the reviewer instead chooses the original snapshot, revalidate its current preconditions and execute only its approved identifiers. Do not rerun the selection filter as the source of execution membership.
5. Make approval refer to an immutable proposal
The execution path should consume the approved proposal, not reconstruct it from the original user request. Preserve the filter and AI instruction as explanation and provenance, but do not use them to regenerate the action after approval.
A useful design sequence is:
- Generate a complete proposal with targets, changes, content and preconditions.
- Assign it an immutable version and retain its details.
- Present a summary with access to the complete action list.
- Record approval against that specific version.
- At execution, load that version and perform live authorization and precondition checks.
- Compare the execution payload with the approved action before making changes.
- Record results against the same proposal and individual target actions.
A content digest can help detect changes to a stored proposal, but it does not prove that the reviewer saw the right information or had authority to approve it. It also needs consistent handling of ordering and formatting. Treat it as an integrity aid, not a substitute for the approval record.
Avoid a gap between the final check and the write. Where supported, use conditional updates tied to the relevant record state or version. If that protection is unavailable, disclose the remaining race condition and consider a narrower action or manual execution.
6. Choose whether changed conditions block one record or the batch
A complete snapshot does not force a single failure policy. The right policy depends on whether the actions are independent.
Whole-batch blocking is easier to reason about when records must move together or the reviewer expects an all-or-nothing business outcome. Its cost is delay: one stale record can stop otherwise valid work. Do not describe a batch as atomic unless the underlying execution mechanism actually provides that guarantee.
Explicit partial execution can suit independent record updates. The reviewer must know in advance that ineligible records will be skipped, not repaired or replaced. The result should identify completed, skipped, failed and uncertain actions individually.
Neither policy permits adding newly matched records. Partial execution is permission to perform a defined subset under approved conditions, not permission to expand the set.
Likewise, define approval expiry around the action's sensitivity and data volatility. Expiry should trigger revalidation or renewed approval, not silent regeneration under the old approval reference.
7. Recover without turning a retry into a new action
A blocked comparison should produce a useful difference report: added and removed identifiers, changed values, altered content and failed preconditions. “Preview expired” is not enough when the operator needs to decide what changed.
For a mismatch detected before execution, retain the original snapshot and offer a clear choice: discard it, revalidate and execute it unchanged, or prepare a new version. Never attach the old approval to the revised version.
For an interrupted execution, first establish what happened. A timeout does not prove that a write failed. Check available execution records or current target state before retrying. Otherwise, an append-note operation could create duplicate notes even when assigning the same owner again appears harmless.
Track recovery at the individual-action level. Retry only unresolved work from the approved snapshot, with current checks applied. Do not rerun the original filter to “finish the batch.” If an action's outcome remains uncertain, route that uncertainty to a person rather than reporting the batch as complete.
8. Run one correspondence check before relying on approvals
Use a non-production workflow to answer one concrete question: does execution consume the reviewed proposal, or does it rebuild it?
Hypothetical exercise: prepare an approval snapshot for three test records, then change the data so a fourth record matches the original filter. Also change a relevant precondition on one approved record. Attempt execution through the normal approval path.
The observable outcomes should be clear: the new record is not acted on; the stale record is blocked or skipped under the stated policy; unchanged records follow that same policy; and any revised proposal requires a new approval. If content is generated, confirm that execution uses the reviewed text rather than producing another version.
Record approved targets, attempted targets, completed actions, skipped actions, unapproved additions and unapproved value or content changes. The acceptance target for the last two measures is zero. Keep the difference report with the approval record so the result can be inspected rather than inferred from a success banner.
This check does not establish every aspect of workflow safety. It establishes something narrower and essential: approval and execution refer to the same business action.
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