Back to all articlesAI action safety

Undo an AI Change Without Overwriting Someone Else’s Work

Alactic Editorial9 min read

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

An AI assistant changes a customer record incorrectly. Before anyone notices, a colleague corrects the contact details and another updates the account owner. Restoring the record’s earlier snapshot would remove the AI mistake—but also erase those legitimate edits.

The safer approach is to reverse only the attributable, still-reversible parts of the AI write. Compare each affected field’s original value, AI-written value, current value, and intervening history. Choose conditional restoration for an untouched AI value, selective correction when valid newer work can be retained, or manual review when evidence cannot settle the decision.

A snapshot is evidence of the past, not an instruction to replace the present. The following process and filled worksheet turn that distinction into a bounded repair.

1. Establish what the completed action actually changed

Start with the action’s affected fields, not the entire record. An action that changed status and ownership should not become an excuse to rewrite contact details, notes, or communication preferences.

For each affected field, collect:

  • Before: the value immediately before the AI write, including whether the field existed.
  • AI wrote: the value actually saved, not merely proposed.
  • Current: the value observed during review.
  • Intervening history: later edits, their authors or processes, and their sequence.
  • Dependencies: related fields or subsequent actions that could change the repair decision.

The assistant’s response is not proof of a saved change. If the saved result is uncertain, mark it uncertain rather than building a reversal around an assumed value. Missing history also limits what can safely be inferred from a current value.

Separate record repair from consequence repair. Restoring a status does not unsend a message or reverse a colleague’s decision based on that status. Track those consequences separately so that a successful field correction is not reported as a complete incident reversal.

2. Choose a repair method for each field

Conditional restoration

Restore the recorded prior value when reliable evidence shows the AI-written value remains untouched. Make the proposal conditional on that evidence still being valid at execution.

Value equality alone is insufficient. Someone may have changed a field away from the AI value and then deliberately changed it back. The text matches, but a later decision now supports it. Use history or a reliable change marker where available.

Selective correction

Prepare a corrected current value when newer work includes both useful information and the AI error. For example, remove an unsupported sentence from a note while retaining a colleague’s later callback request.

This is a new edit requiring approval, not a mechanical reversal. Use it only when the valid contribution and the error can be separated confidently.

Preservation or manual review

Preserve a later edit when it clearly supersedes the AI change. Escalate when intent is disputed, history is incomplete, or related fields cannot be judged independently.

Give each conflict an owner and a specific question. “Check this record” leaves the investigation unfinished. “Does the owner reassignment remain valid after the incorrect handoff status is removed?” identifies the decision needed.

3. Work through a filled rollback worksheet

Fictional worked example: All names, values, times, and events below are hypothetical. An AI action changed six fields: lead status, account owner, preferred contact channel, account summary, handoff priority, and an internal routing label. Colleagues subsequently edited some of those fields.

The reviewer has reliable before-and-after evidence and the relevant change history. These are proposed decisions, not authorised execution instructions. The worksheet can be reused by replacing its record details, evidence, and decisions.

ROLLBACK REVIEW — HYPOTHETICAL
Record: Prospect Cedar Workshop
AI action: A-204, completed at 09:10
Review snapshot: 10:05
Review owner: Priya, operations lead
Scope: Six fields changed by A-204

FIELD: Lead status
Before AI: Discovery scheduled
AI wrote: Ready for handoff
Later edits: None recorded in available history
Current: Ready for handoff
Decision: RESTORE CONDITIONALLY
Proposed value: Discovery scheduled
Evidence: History shows no intervening change.
Condition: Reviewed value and change marker still match.
Dependency: Coordinate with priority and routing decisions.

FIELD: Account owner
Before AI: Maya
AI wrote: Leo
Later edits: 09:35, Jordan reassigned owner to Nina
Current: Nina
Decision: PRESERVE
Proposed value: No write
Reason: Restoring Maya would erase Jordan's later edit.
Question for Jordan: Does Nina's assignment depend on
that incorrect handoff status?

FIELD: Preferred contact channel
Before AI: Email
AI wrote: Phone
Later edits: 09:42, Maya recorded the prospect's preference
for a video meeting
Current: Video meeting
Decision: PRESERVE
Proposed value: No write
Reason: Later confirmed preference supersedes earlier values.

FIELD: Account summary
Before AI: Needs a discovery call before next steps.
AI wrote: Budget confirmed. Ready for handoff.
Later edits: 09:50, Maya appended:
Requested a Thursday afternoon callback.
Current: Budget confirmed. Ready for handoff.
Requested a Thursday afternoon callback.
Decision: SELECTIVE CORRECTION
Proposed value: Needs a discovery call before next steps.
Requested a Thursday afternoon callback.
Reason: Remove unsupported statements; retain callback.
Approval required: Maya confirms the combined wording.

FIELD: Handoff priority
Before AI: Normal
AI wrote: High
Later edits: 09:25, Jordan changed High to Normal;
09:40, Jordan changed Normal back to High
Current: High
Decision: ESCALATE; PRESERVE PENDING REVIEW
Proposed value: No write pending decision
Reason: Current equality hides intervening decisions.
Owner: Jordan; Priya coordinates if Jordan is unavailable.
Question: Is High intentional independently of handoff status?

FIELD: Internal routing label
Before AI: Label absent
AI wrote: Handoff queue
Later edits: None recorded for this label
Current: Handoff queue
Decision: RESTORE CONDITIONALLY
Proposed change: Remove Handoff queue only
Reason: Reverse this addition, not the entire label collection.
Condition: No deliberate reapplication; retain other labels.
Dependency: Review with status and priority.

OUTSIDE SCOPE: Preferred contact time
Current: Thursday afternoon
Decision: PRESERVE; no write
Reason: A-204 did not change this field.

DEPENDENCY HOLD
Status, priority, and routing require a coherent end state.
Jordan must answer the priority and ownership questions.
Unrelated approved repairs may proceed only if they do not
create a misleading or inconsistent record.

EXECUTION RECORD — COMPLETE AFTER EACH ATTEMPT
Approved proposal and approver
Fresh pre-write evidence
Attempted change
Known outcome or unknown outcome
Observed final value and supporting history
Remaining dependency and review owner

The six decisions are: two conditional restorations, two preserved values, one selective correction, and one escalation. That does not mean five writes should proceed. The two preserved values require no write, the summary needs approval, and the handoff-related changes remain dependent on review.

Notice why matching values are not enough: priority is High both after the AI action and at review, but Jordan’s intervening edits make automatic restoration inappropriate. The contact-channel row shows the opposite pattern: a later correction clearly supersedes the earlier values, so preserving it is the useful outcome.

4. Protect the gap between approval and execution

A correct worksheet can become stale before execution. Apply a repair only while its reviewed assumptions still hold.

Where supported, conditional writes or reliable change markers can detect intervening edits. Check the workspace and current plan for available history, permissions, and change controls; this process is a recommended operating design, not a claim that Alactic provides a particular rollback mechanism.

  1. Confirm the operator’s permission to change the specific record and fields.
  2. Read the current affected values and available change evidence.
  3. Compare them with the approved proposal, including relevant dependencies.
  4. If relevant evidence differs, stop that repair and refresh the decision.
  5. Apply only the authorised field-level change.
  6. Read back the result and record what was observed.

A separate read followed by a write still leaves a gap in which another edit can occur. Read-back may reveal a problem; it does not prevent that gap. If the system cannot enforce the required condition at write time, consider a coordinated review period using supported controls. Defer the write if the remaining risk is unacceptable.

Permission is separate from evidence. Knowing a record identifier or possessing its old snapshot does not authorise an edit. OWASP’s authorisation guidance recommends least privilege, denial by default, and permission checks on every request. These are design recommendations, not evidence about a particular workspace.

5. Treat absence, collections, and linked fields precisely

Not every repair is a scalar value replacement. Three distinctions prevent a narrow-looking correction from becoming a broad overwrite.

Absent is not necessarily blank. A missing field, a blank string, and an explicitly cleared value may have different meanings. Record the intended state precisely. In the worksheet, reversing the routing addition means removing that label, not saving a blank label.

Collection membership is not the whole collection. Remove or restore only the attributable item when supported and appropriate. Replacing an old label list could discard unrelated additions. For ordered collections, position can also matter: restoring membership without reviewing order may produce a different result from the one intended.

Linked fields describe a combined state. Discovery status, high handoff priority, and handoff routing may be individually valid values but collectively confusing. Ask whether the proposed combination makes business sense. Do not resolve inconsistency by automatically reverting every related field; a later priority decision might have an independent reason.

If the available write operation replaces more content than the approved repair requires, treat that broader scope as a new risk requiring review, not as an implementation detail.

6. Handle partial repairs and unknown outcomes

A repair can partly succeed, fail without changing anything, or leave an uncertain result. Keep those outcomes distinct.

  • A condition no longer matches: stop that repair, retain the newer value, and obtain a refreshed decision.
  • A write reports failure: if reliable evidence confirms no change was saved, record the repair as not applied. Otherwise classify the outcome as unknown, inspect the current state and available history, and reassess dependencies before retrying.
  • Several operations were attempted: inspect each operation’s outcome. An overall failure does not prove that earlier changes were undone or that later changes were never saved.
  • Some repairs are verified: record those individually while leaving unresolved operations pending or blocked.
  • The correction is itself wrong: treat it as a new change with its own before value and intervening history. Do not blindly reverse it.

A retry is not a neutral response to uncertainty. Repeating a collection change or applying a prepared note again could overwrite work added since the first attempt. Re-establish the current state before deciding whether another write is needed at all.

Keep original-action evidence separate from repair evidence. Otherwise another operator may mistake a repaired value for the AI-written value and attempt the wrong reversal.

7. Close the review with an accurate handoff

Use a short closeout checklist rather than a blanket “rollback complete” label:

  • Every affected field has a supported restore, preserve, correct, or escalate decision.
  • Every attempted write has a verified outcome or an explicit unknown-outcome status.
  • Unresolved conflicts have a named owner, a focused question, and a clear statement of what is blocked.
  • Related fields have been checked for a coherent end state.
  • Consequences outside the record repair remain separately tracked.

In the hypothetical example, an accurate interim handoff might read: “Owner and contact channel preserved; summary correction approved and verified; status and routing held pending Jordan’s priority decision.” That gives the next operator a trustworthy starting point without claiming that unresolved work has disappeared.

AI action safetyRecord managementRollback reviewBusiness 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