AI-first company

Governance and risk · Run

When AI output goes wrong, correct it fast and in the open

Decide now what happens when an AI error gets out: who is told, who withdraws or corrects the work, and who checks how far the problem goes. In the documented cases, firms that owned up quickly were treated far better than firms that defended the error.

By
Last reviewed
Corrections
Report an error
When Do next. Once your first uses of AI are running.
Effort Low (days, one person)
Evidence Emerging

Eight documented cases bear on it, and in three a court said in its own words that the response changed the outcome. No study has tested it, and most of the cases are from law firms.

Fits
Industries All industries
In the guide What fails

What it is

Agree a short plan now, before anything goes wrong, for the day an AI error reaches a customer, a client, a court or the public. The plan says who must be told, who corrects or withdraws the work, who tells the people affected, and who checks whether the same error is sitting in other work.

What the evidence shows

Eight documented cases are the basis for this advice. In most, what the company did after the error shaped what the error cost.

Firms that defended the error

  • Twelve weeks of standing by fake cases. After a New York firm filed six made-up court decisions, its lawyers stood by them when challenged. The court said the outcome would have been “quite different” if they had admitted the problem when first asked.
  • “Cosmetic errors.” In the English High Court judgment, a law centre’s first reply to being told that five cited cases did not exist was to call them “cosmetic errors”. The court ordered it to pay £2,000 in wasted costs and referred it to the regulator.
  • Three months to concede a small claim. A customer wrote to Air Canada for about three months before the airline admitted that its chatbot had used “misleading words”. The dispute went to a tribunal, and the airline lost in public over a few hundred dollars.

Firms that owned up

  • Withdrawn the next day. Lawyers at a large US firm withdrew their filing the day after the court asked about it, said openly that AI had been used and paid the other side’s fees. The court named the drafter’s “honesty and candor” as a reason for a lighter penalty, and did not penalize the firm. The court also noted that the lawyers could have acted sooner: the other side had questioned the cases eight days before.

  • An outside check of everything else. Another US firm admitted the error the day the court asked and paid a second firm to check more than 2,400 citations in its other filings. The court said the results of that check were why it did not suspend the lawyers.

  • Self-reported. The solicitor at a ten-person English firm apologized and reported himself to his regulator. The court decided not to start contempt proceedings.

  • Reversed within weeks. Commonwealth Bank of Australia announced that 45 customer-service roles would go and be replaced by an AI voice bot. After the union challenged it, the bank said its assessment had been an “error” and that the roles were not redundant. It apologized and offered the staff the choice of staying, moving to another role or leaving with redundancy pay.

A third response: admit the errors, defend the conclusions

  • Deloitte Canada wrote a health workforce plan for a Canadian province, costing about CAD 1.6 million. The press reported at least four citations to research papers that do not exist. Deloitte acknowledged the four citations were wrong and stood by the report’s conclusions. The account comes from press reports.

One more case shows a slow correction. A consulting report for a government client was reissued to fix wrong citations, with the use of AI disclosed. A second correction followed four months later.

How to do it

  1. Name who is told. One person, usually the AI owner or the managing director. Anyone who finds an AI error in work that has gone out tells them the same day.
  2. Make it safe to report. Say in advance that reporting your own error quickly will be treated differently from hiding it. See protect psychological safety during rollout.
  3. Stop the harm first. Withdraw the document, correct the page, switch off the chatbot answer.
  4. Tell the people affected, plainly. What was wrong, what the correct information is, and what you are doing. Do not call it minor until you know it is.
  5. Say that AI was involved, if it was. In the cases above, courts found out anyway.
  6. Check how far it goes. Look at other work done by the same person, tool or method over the same period. Fix everything in one round.
  7. Put right the cost. Refund, redo the work, or pay what the error cost the other side.
  8. Change the process and say what you changed. A new check, a block, a notice in the tool. See keep a person checking.
  9. Write it down. A half-page record of what happened and what changed. Use it in training.

What to measure

  • Time from finding an error to telling the person affected.
  • The number of correction rounds per incident. The target is one.
  • The number of errors staff report themselves, against those found by outsiders.

Where it fails

  • The first reply plays it down. “Minor” and “cosmetic” were the words that made things worse.
  • Fixing only the error someone spotted. The second correction is the one that damages trust.
  • Blaming the tool. Air Canada argued that the airline was not responsible for its chatbot. The tribunal called this “a remarkable submission”.
  • Punishing the person who reports. Next time nobody will.

By company size

  • Micro and small: the owner is told the same day and writes to the customer personally.
  • Medium: a named contact, a half-page checklist of the steps above, and a log of incidents.
  • Mid-market and enterprise: AI errors go into the existing incident process, with legal and communications involved from the first day.

Company cases

Companies that did this

What happened without it

Companies that came round to it

Sources

No published study tests this practice directly, so it is rated Emerging. The rating rests on the company cases above and on the sources below.

  1. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC). Opinion and Order on Sanctions, 22 June 2023. US District Court, Southern District of New York. storage.courtlistener.com
  2. Wadsworth v. Walmart Inc., No. 2:23-cv-118-KHR. Order on Sanctions and Other Disciplinary Action, 24 February 2025. US District Court, District of Wyoming. storage.courtlistener.com
  3. Johnson v. Dunn, No. 2:21-cv-1701-AMM. Sanctions order, 23 July 2025. US District Court, Northern District of Alabama. storage.courtlistener.com
  4. Information Age (21 August 2025). CBA reverses AI-driven job cuts, admits error. ia.acs.org.au
  5. HR Reporter (25 November 2025). AI errors: province grapples with Deloitte report marred by false citations. hrreporter.com