Back to all articlesAI action safety

Define What Stop Means for an AI Action Already in Progress

Alactic Editorial10 min read

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

A staff member asks an assistant to stop a bulk update. The assistant acknowledges the request, but several records change afterward. Did the stop fail, or were those writes already beyond the cancellation boundary?

To stop AI actions already in progress safely, define three separate things: when new work stops being dispatched, what happens to work already submitted, and how final outcomes are verified. Acknowledging a request does not prove execution has stopped. Cancelling pending work does not reverse completed writes.

The operating rule should be explicit: stop prevents whatever remains preventable, reconciliation establishes what actually happened, and reversal requires its own decision and authorisation. The worksheet below makes those boundaries visible for a bounded bulk update.

1. Make the stop contract specific before execution

A useful stop contract describes behaviour, not just a button label. Before authorising a batch, agree on the scope of the stop and the evidence required to declare it complete.

  • Scope: Does stop apply to this batch, one step, or every run of the same workflow?
  • Dispatch boundary: At what checkpoint does the runner stop submitting new writes?
  • Submitted work: Can a queued request be cancelled, or can the system only wait for its result?
  • Retries: Does stopping suppress retries and resumed work as well as initial attempts?
  • Verification: What evidence establishes the result of each submitted write?
  • Reversal: Who may approve a corrective update, and what checks must precede it?

Check the controls actually available in the workspace and current plan. Do not assume that a conversational stop request reaches an executor, that a queue supports cancellation, or that an undo feature exists. These are recommended operating practices, not claims about a particular product control.

Where cancellation is unavailable, describe the narrower capability honestly: no further requests will be submitted after the runner observes the stop, but submitted requests may finish. If even that dispatch boundary cannot be verified, do not describe the batch as safely stopped. Escalate to the authorised operator who can establish its status.

2. Separate the request from the effective boundary

A stop event needs more than one timestamp. Record when the person requested it, when the execution system acknowledged it, and when the relevant execution boundary took effect. These moments may differ.

Requested means the instruction has been received through the chosen channel. Effective means the defined execution control has been applied. Reconciled means submitted actions have known outcomes and pending actions have been accounted for.

A write can therefore finish after the request time without proving that new work was dispatched after the effective stop boundary. Conversely, a reassuring acknowledgment can conceal a failure if a worker continues dispatching afterward.

Use evidence that distinguishes dispatch from completion. An action identifier, submission event, result receipt, and final record check are more useful than a single batch status. If different systems record time differently, prefer a reliable event sequence or correlated identifiers over a claim based solely on clock order.

The immediate response to staff should stay provisional: the stop has been requested; submitted work is being checked; no final change count is available yet. Save the definitive message for reconciliation.

3. Use action states that answer practical questions

For each action, record its state and what the operator can still do. A single label such as stopped is too broad to guide recovery.

  • Pending: The write has not been submitted. It can still be withheld if the execution path honours the stop.
  • Queued: The write has entered a queue but has not been confirmed as executing. It is only cancellable if the queue provides an effective cancellation mechanism.
  • In flight: The write has been submitted and has no definitive result. Treat its outcome as unresolved unless cancellation is confirmed.
  • Completed: The write has a confirmed successful outcome. Stopping cannot prevent it; changing it back is a separate write.
  • Failed without a write: Evidence establishes that no update occurred. Suppress any retry associated with the stopped batch.
  • Unknown: Evidence is insufficient to establish whether the write occurred. Investigate rather than treating it as failed.
  • Cancelled: Evidence establishes that the action will not execute through this batch, including its retry or resume paths.

Keep execution state separate from recovery state. A completed action may also be marked reversal proposed, reversal approved, or reversal completed. Calling it cancelled would erase the fact that it changed a record.

Do not infer success merely because the record now has the desired value: another operator could have made that change. Where possible, connect the result to the specific action. If attribution remains uncertain, retain that uncertainty in the worksheet.

4. Completed hypothetical stop-status worksheet

Fictional worked example: Every identifier, time, count, control, and outcome below is hypothetical. The example assumes an execution mechanism that checks a stop flag before dispatch, suppresses retries, and supports cancellation of the stated queued work. Verify whether equivalent controls exist before adopting it.

Batch definition and agreed stop behaviour

  • Batch: BULK-HYP-17, containing eight individually tracked CRM record updates.
  • Authorised change: Set the internal review-status field from Awaiting review to Ready for review on records H-01 through H-08.
  • Exclusions: No messages, deletions, owner changes, or writes to other fields. No downstream write is part of this authorisation.
  • Before-state evidence: Each target's review-status value and available revision evidence were captured before execution.
  • Stop contract: Stop dispatching new writes at the next dispatch check; request cancellation for queued actions; suppress retries and resume paths for this batch. Submitted writes may finish. Do not reverse completed writes automatically.
  • Responsible roles: The batch operator reconciles results. The record owner approves any corrective batch.

Stop timeline

  • 10:04:12: Operator requests stop for BULK-HYP-17.
  • 10:04:13: Runner records the effective dispatch stop. One write submitted between the request and this boundary is included in the in-flight count.
  • 10:04:14: Runner confirms no further dispatch or retry from this batch; queued cancellation remains under verification.
  • 10:04:20: Queued cancellation and all submitted outcomes are reconciled.

Filled action ledger

  1. H-01 — Completed at the effective boundary. The write had already succeeded. Nothing remains preventable. Final evidence confirms Ready for review. Returning it to Awaiting review requires a separately authorised corrective write.
  2. H-02 — Completed at the effective boundary. The write had already succeeded. It cannot be cancelled. Final evidence confirms Ready for review. Separate reversal would be required if the owner decides the change should not remain.
  3. H-03 — In flight at the effective boundary. Submission preceded the stop boundary; no effective cancellation was available for that request. It succeeded at 10:04:15. Final state: completed. Separate reversal is required to restore the previous value.
  4. H-04 — In flight at the effective boundary. Submitted at 10:04:12, after the request but before the effective boundary. It succeeded at 10:04:16. Final state: completed. Its completion is not evidence of post-boundary dispatch. Separate reversal is required if approved.
  5. H-05 — Queued at the effective boundary. Cancellation was requested, then confirmed before execution. Final state: cancelled. The field remains Awaiting review; no reversal is needed.
  6. H-06 — Pending at the effective boundary. Withheld before submission, with its resume path suppressed. Final state: cancelled before dispatch. The field remains Awaiting review; no reversal is needed.
  7. H-07 — Pending at the effective boundary. Withheld before submission, with its resume path suppressed. Final state: cancelled before dispatch. The field remains Awaiting review; no reversal is needed.
  8. H-08 — In flight at the effective boundary. Final result established failure without a write. Its retry was suppressed. Final state: failed without a write. The field remains Awaiting review; no reversal is needed.

Reconciled totals and final message

Final accounting: Eight targets equal four completed writes, three cancelled actions, and one failure without a write. No actions remain pending, queued, in flight, or unknown. No post-boundary dispatch was found. No corrective writes have been authorised.

Stop reconciliation is complete for BULK-HYP-17. New dispatch stopped at 10:04:13. Four records changed: H-01 through H-04, including two submitted writes that completed after the stop took effect. H-05 through H-07 were cancelled before execution. H-08 failed without a write and will not retry through this batch. No outcomes remain unresolved. The four completed changes remain in place; returning them to their previous values requires a separately approved corrective update.

5. Reconcile unknown outcomes before retrying or reversing

The worked example has definitive outcomes. Real execution records may be less clear. A missing response does not establish that a write failed, and a cancellation request does not establish that cancellation succeeded.

For an unknown action, first preserve its identifiers and available execution evidence. Check the receiving system's action result or record history where available. Compare the current value with the captured before-state, while checking for intervening edits. Keep the action unresolved if that evidence cannot establish the outcome.

Do not blindly retry an unknown write to discover whether it worked. A repeat might overwrite newer work or trigger another effect. Likewise, do not issue a reversal merely because the batch was stopped: there may be nothing to reverse, or the record may now contain a legitimate later change.

If reconciliation cannot finish promptly, send an interim report that names the unknown actions, the assigned investigator, and the next check. A partially reconciled stop is an acceptable reported state; a falsely definitive all-stopped message is not.

6. Treat reversal as a new, bounded operation

Stopping answers whether remaining execution can be prevented. Reversal answers whether a completed change should be corrected now. Those are different decisions with different risks.

In the hypothetical worksheet, H-01 through H-04 are candidates for review, not automatic rollback. Before approving a corrective batch, compare each current field value and available revision evidence with the state left by the original action. If someone has since changed a record, route that record for review rather than overwriting it with an old snapshot.

A field correction also does not necessarily undo consequences that followed the original write. Check whether other authorised processes reacted to the change. Record any separate effects as their own recovery tasks rather than expanding the correction silently.

Keep corrective authority narrow. The OWASP authorization guidance recommends least privilege and permission checks on every request. Applied here, an original batch authorisation should not be treated as unrestricted permission to perform arbitrary recovery writes. This is design guidance, not evidence about a specific product.

7. Test the boundaries that commonly fail

Before relying on the procedure, run a small authorised test on non-critical records in a suitable test workspace, if available. Choose cases that expose the boundary rather than simply confirming that an idle batch stops.

  • Stop while work is queued: Verify that cancellation removes executable work, not merely its visible entry.
  • Stop during submission: Identify requests dispatched between the user's request and the effective boundary.
  • Stop before a retry: Confirm that delayed retries cannot restart the stopped batch.
  • Stop with multiple workers: Check every worker associated with the batch, not just the one that acknowledged the request.
  • Resume after stop: Confirm that reopening a session or restarting a runner does not silently resume cancelled work.
  • Lose a result: Verify that an uncertain outcome stays unknown until evidence resolves it.

Smaller batches and fewer concurrent submissions generally make the boundary easier to inspect, at the cost of throughput. Larger batches may be operationally convenient but leave more work submitted when a stop arrives. Choose the trade-off explicitly; do not use a fast acknowledgment as a substitute for measuring outstanding work.

8. Make the next authorised batch auditable

Before the next bulk update, fill in the worksheet header: target set, permitted fields, stop scope, effective boundary, retry behaviour, evidence sources, and recovery approver. If the available controls cannot support that contract, narrow the batch or use a more directly supervised execution method.

At the end of a stop test, require three observable results: every target has a final state or a named unresolved investigation; completed writes are separated from cancelled actions; and the final message states what changed and what remains unchanged.

The decisive check is simple: can another staff member account for every authorised action without interpreting the word stopped? If not, the stop status is not yet specific enough to guide the next decision.

AI action safetyExecution controlBulk updatesReconciliation

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