A person who approves AI-assisted work is responsible for the decision they actually made. The organization also has to answer for the workflow it authorized, the conditions it gave the reviewer and the correction owed to whoever received the flawed result. An approval click does not settle all three questions.
For a workflow owner, the useful response is to reconstruct what happened and give each necessary repair an owner. Establish what the approver could reasonably check, what they did, and which decisions belonged elsewhere. Keep any legal-liability or personnel assessment in the appropriate authorized process; a general AI policy cannot decide an individual case by itself.
Separate approval from organizational accountability
A policy saying employees must check AI output can establish an expectation. It does not establish that the required information was available, the check was feasible or the approved system behaved as intended. Nor does a system defect automatically explain every choice an employee made.
NIST's voluntary AI Risk Management Framework assigns attention to clear responsibilities, personnel training and executive responsibility for AI deployment risks. Its oversight provisions extend beyond the final user. This is governance guidance, not a ruling on who must pay damages after an error. The framework's current page says a revision is in progress.
For practical purposes, distinguish three questions:
| Question | What the investigation needs to establish |
|---|---|
Who accepted this result? | The actual version, intended use, required checks and decision authority. |
Who controlled the conditions? | Who supplied the evidence, designed the handoff and allocated review capacity. |
Who must put things right? | The owner of the affected outcome, each corrective action and verification of the repair. |
The same person can hold several responsibilities in a small operation. In a larger company, they may sit across service management, IT, information owners and leadership. Either way, “everyone is accountable” is too vague to tell a customer who will correct an incorrect decision.
Reconstruct the error before assigning the remedy
Suppose Alex Morgan coordinates warranty claims at an industrial-pump distributor. An assistant drafts a denial using the standard 12-month warranty and an invoice showing purchase 18 months earlier. Alex approves the response. The customer challenges it with a signed service agreement providing 24-month coverage for the fault in question.
The agreement was missing from the packet Alex reviewed. That fact matters, but it does not finish the investigation. Could Alex access the agreement elsewhere? Did the procedure require checking for customer-specific coverage? Was that step performed, skipped or made impractical by the way work arrived?
Preserve the relevant records under the company's access and retention rules:
- The accepted output. What did the customer actually receive, and which version did Alex approve?
- The evidence available then. Which warranty, invoice and agreement could the assistant and Alex each access?
- The expected decision. What instructions and authority applied to a denial with uncertain coverage?
- The working conditions. What review workload, training, support and escalation route existed?
Do not replace the original record with a cleaner account of what should have happened. A later-added agreement cannot show what was visible at approval time. Equally, a missing attachment does not prove Alex had no other way to obtain it.
The existing human-review guide explains the checks before accepting a result. After a failure, investigate whether those checks were specified, possible and actually carried out.
The same denial can require different repairs
The signed agreement was unavailable
The information owner repairs the permitted route to customer-specific coverage.
A required, feasible check was skipped
The manager examines the omission and agrees the appropriate response.
The assigned review could not fit the work
The service owner changes workload, staffing, scope or escalation.
The UK's Health and Safety Executive advises investigating why human failures occur, including the influence of workload, time pressure, competence and interface design, rather than stopping at operator error. Its guidance also distinguishes unintended errors from deliberate deviations and warns that several causes may coexist. This is workplace-safety guidance, not an AI trial. Its useful investigation principle is to connect an observed failure to the conditions and conduct that produced it.
Test whether the review duty was feasible
“Alex should have checked” needs a specific object. Checked the invoice date? The standard warranty? Whether a separate customer agreement existed? Each requires different evidence. A general instruction to be careful leaves the decisive requirement uncertain.
An accounting supervisor's public Reddit reply describes an employee submitting an AI-written work paper that listed procedures the employee had not performed. The supervisor also said management encouraged AI use while training was limited. The underlying document and employer were not independently verified. The concrete distinction remains useful: a record claiming checks occurred is different from evidence that someone performed them.
For Alex's case, ask the reviewer to demonstrate the expected check with the materials and access normally available. Have the manager observe how a customer-specific agreement is found and how an unresolved case is held. Assess ordinary workload too, including interruptions and the time spent obtaining missing information. A successful demonstration with a prepared packet does not prove that a full queue is manageable.
Different findings call for different responses:
- If the agreement cannot be retrieved through the approved process, fix access or route the decision to someone who can establish coverage.
- If checking for an agreement is required and feasible but was omitted, examine that omission and the reason for it through the normal management process.
- If staff are expected to complete checks that the workload makes impractical, the service owner needs to change the incoming volume, staffing, task scope or review arrangement.
More training may help a misunderstood requirement. It cannot, by itself, provide a missing agreement or give a coordinator authority to pause a deadline. Use the manager's work-evaluation guidance to discuss contribution and support without turning one deliverable into a complete performance judgment.
Assign the work beyond the approver
Keep responsibility close to the control someone can change. Alex may be able to hold a reply and correct a draft. Alex may not control the service-agreement repository, the assistant's source access or the team's throughput target.
The following allocation is a starting point for the pump distributor, to confirm against its actual authority rules.
| Owner in the distributor | Work to own after the flawed denial | Evidence the work is complete |
|---|---|---|
Service manager | Resolve the customer's coverage decision and arrange the appropriate corrected response. | The authorized decision and customer follow-up are recorded. |
Agreement information owner | Make the applicable signed agreement available through the permitted review route. | The relevant record can be found for the right customer and equipment. |
Workflow operator | Repair how the assistant or review packet handles missing customer-specific coverage. | Tests show missing information produces the intended hold or escalation. |
Department leader | Resolve gaps in authority, reviewer capacity and incentives. | Staff can perform the required check under the agreed operating conditions. |
This allocation does not remove Alex's checking responsibility. It prevents a correction assigned to Alex from silently standing in for repairs Alex cannot make.
In a reply to Phaedra Boinodiris's October 2026 LinkedIn post, Robert Wood connects this problem to work packaging: a handoff needs someone to own the next action and a way to verify resolution. His visible profile describes work-package data on large construction projects and independent AI exploration. This is a practitioner's perspective, not a measured result from an identified employer.
For the distributor, opening a ticket about the missing agreement is a handoff. Showing that the correct agreement is available and that an incomplete case is held is evidence of a repair. Keep the distinction visible in the incident record.
Close the incident with a checked change
Give the immediate customer correction and the recurring workflow repair separate owners and completion conditions. One can be complete while the other remains open.
- Contain the affected use. The authorized service owner decides whether similar denials need to pause or receive additional review while the scope is established.
- Correct the affected result. Reassess coverage against the applicable records and arrange the appropriate customer response. Investigate whether other accepted cases share the missing-agreement condition.
- Change the failed control. Repair the identified access, instructions, routing, authority or capacity problem. Do not assume a new prompt resolves a source-access defect.
- Retest the condition that mattered. Try permitted cases with standard coverage, extended coverage and missing or conflicting records. Confirm both correct decisions and appropriate holds.
- Accept the remaining limits. Record what was tested, what remains unresolved and who can authorize continued operation within those limits.
A clean retest is evidence about those cases and conditions. It is not proof that all future denials will be correct. Include the receiving staff in the check: a technically successful retrieval can still leave an unusable packet or an unanswered escalation.
If a supplier's defect contributed, give the supplier a specific issue and retain responsibility internally for customer correction and the operating decision. A vendor support ticket is not a completed remedy. Revisit the relevant launch-review questions about evidence and authority when deciding whether the repaired arrangement is ready.
Share the useful lesson through the failure-learning process, with an appropriate audience and sensitive details removed. Keep the fuller investigation available to authorized people. A learning summary should not quietly decide a disputed personnel or legal question.
Questions about approved AI errors
Is the person who approved the AI result solely accountable?
An approval establishes a decision to examine, not sole responsibility for every contributing condition. Determine the approver's duties and conduct, who controlled the evidence and workflow, and who must correct the affected result. For the pump distributor, Alex's approval does not establish that Alex controlled access to the signed agreement. Use the three investigation questions to allocate the work precisely.
What if the reviewer could not check the result properly?
Identify the missing condition: evidence, expertise, time or authority. Arrange a supported hold, different reviewer or manual route while the responsible owner fixes it. A policy acknowledgement does not supply the missing capability. Check the feasibility of the actual review duty, including ordinary workload, before relying on the same approval arrangement again.
Does examining the system mean ignoring a skipped check?
No. Establish whether a check was required, feasible and omitted, and investigate why. Preserve the difference between an unintended mistake, an unclear instruction and a deliberate departure without deciding the explanation in advance. Address individual conduct through the organization's existing fair process, while fixing any contributing workflow defect. One responsibility does not cancel the other.
Can the company leave the problem with its AI supplier?
A supplier may need to repair a contributing product defect. The company still needs an internal owner for the affected customer result, investigation and decision about continued use. Preserve the specific failure evidence and test the supplier's proposed repair in the actual workflow. Contractual responsibility and legal liability require their own assessment; this operating allocation does not settle them.
When can an approved-error investigation be closed?
Close each action when its named owner has supplied the agreed evidence. A corrected customer response can finish before a recurring source-access defect is repaired. Keep that repair open until the relevant cases and escalation path have been tested and the authorized owner has addressed remaining limits. The incident closure steps distinguish correcting one result from accepting the revised workflow.
