A poor decision or an unlucky outcome? Running the AI incident review

In August 2025 a senior barrister stood before the Supreme Court of Victoria and took full responsibility for defence submissions containing fabricated quotations and citations to cases that do not exist. The lawyers had checked the first citations and assumed the rest. The prosecution, which received the submissions, had not checked them either. Weeks later, Deloitte issued a revised version, dated 26 September 2025, of a report to an Australian government department, disclosing that generative AI had been used. An outside researcher had found a fabricated quotation attributed to a federal court judgment and references to works that do not exist. The firm agreed to repay part of its fee.

The two cases did not end in the same way. In Victoria, a named barrister accepted responsibility in open court. In the Deloitte case, the firm revised the report, disclosed the AI use and agreed to repay part of its fee. As reported, it said the matter had been resolved directly with the client, and it did not confirm whether AI caused the errors. The acceptance in Victoria was right, but it came after the harm. The question a risk function should ask is what the review that followed could learn from, and whether the organisation kept the record that would let it.

The first two reflexes are both scored, and both score low

When an AI-assisted decision goes wrong, the first reflex is to say the tool did it. AIRS-J/1 scores that as Level 0 on dimension P2, Consequence Ownership and Correction. The anchor begins: deflects the outcome to the AI system, the data, another party, or the brief. The tool did produce the error. It did not sign the document, and the Standard asks the person who did to accept the consequence of the original decision without transferring it.

The second reflex is to abandon every decision made the same way. The AI-drafted board paper was wrong, so no board paper may be AI-drafted. The assistant's citation was fabricated, so the assistant is banned. That matches Level 1: accepting the outcome in general terms without identifying what specifically was wrong, or revising under pressure rather than on evidence. It feels like correction. It is a second decision made with less judgment than the first, because it treats the whole class of decisions as the error rather than finding the error in the one.

Level 2 is where correction begins. On new evidence, the person holds, modifies or reverses the position on the merits and states which evidence drove the choice, and accepts the consequence of the original decision without transferring it. The Standard adds a note on the other direction: holding a sound position under challenge is Level 2 behaviour, and changing position because someone senior pushed back, with no new evidence, is Level 1. Correction is a response to evidence, not to pressure.

For HR and L&D: the second reflex is the one that reaches your function as a policy change. Before the ban is written, the question to put back is which decision was poor and whether the ban addresses it.

Level 3 separates the reasoning from the luck

A decision can be well made and turn out badly. A country manager checks the AI's demand forecast against two sources, finds them consistent, commits a proportionate order, and a port closure nobody could have foreseen strands it. The outcome is bad. The reasoning was sound. A review that concludes the manager should not have trusted the forecast has learned the wrong lesson and will produce a more cautious manager, which is a different failure.

A decision can also be badly made and turn out fine. A colleague skips the check, commits the same order, and the port stays open. The outcome is good. The reasoning was absent. A review that never happens because nothing went wrong has missed the only evidence it will get before the next time.

Level 3 on P2 requires the person to tell the two apart. The anchor: distinguishes a poor decision from an unlucky outcome, corrects reasoning that was wrong, and does not abandon reasoning that was sound but unlucky. That is the whole discipline of an incident review in one sentence, and it is the discipline most reviews lack, because most reviews start from the outcome and work backwards to a cause. The outcome is the least informative thing in the room. The reasoning at the time of the decision is the most.

Level 4 turns the review into a rule. The anchor: identifies what in their own process produced the error, in terms that would change how the next decision of this class is made. Not "I should have been more careful". Rather, as an invented example: "I checked the first citations and extrapolated; the rule is that every citation is checked or the document does not go out." The Victoria lawyers' statement that they checked the first citations and assumed the rest is, read one way, a Level 4 observation. It names the process error precisely. The Standard would score it that way only if a rule was then stated that would change how the next decision of this class is made.

The outcome is the least informative thing in the room. The reasoning at the time of the decision is the most.

The pattern that interrupts a certification

The Standard reserves a special category for the most serious version of the first reflex. Annex D lists four critical-review patterns. CR-P2 is triggered when, after new evidence disproves the committed position, the person preserves the action while transferring responsibility to the AI system, the data or another party, and the resulting harm pathway remains active. In plain terms: the decision has been shown to be wrong, the harm is still running, and the person's response is to keep doing it and blame the tool.

A matching observation is not an automatic fail, and the Standard sets out what happens instead. A single trigger renders the affected dimension Insufficient Evidence pending targeted reassessment on a decision point the candidate has not seen. It may not be averaged into a certifying dimension score. The effect is that a pattern this serious is not absorbed by a strong performance elsewhere, and does not end a certification on one observation either. The review board confirms or amends the patterns from real assessments, and a pattern that cannot be applied reliably and fairly is removed or rewritten. The assessor may not infer motive. The trigger must be cited to transcript evidence and a stakes pathway.

For a risk function, CR-P2 is a definition of the incident that matters most. Not the error, which is ordinary. The continuation of the action after the error is known, with responsibility pointed at the tool. That is the pattern a review should be designed to find, and the one most reviews are not built to see because they close once a cause has been assigned.

The one-page review

Six questions, one page, written within a week of the incident by the person who made the decision, and read by someone who did not. It is a reading aid drawn from the P2 anchor, not a template the Standard prescribes.

What was decided, and when? The position committed to, and the time it was committed, before any consequence was known. If the record shows only the final version, this question cannot be answered, and the review has already lost most of its value.

On what evidence? What the person had in front of them at the time, including the AI output, and what they understood about where each piece came from: generated, retrieved, asserted or verified.

What was verified, and by whom? Which load-bearing claims were checked, against what, by a named person. Which were not, and why.

What changed? The new evidence that showed the decision to be wrong, and when it arrived.

Was the reasoning poor, or the outcome unlucky? The Level 3 question. Given what was known at the time and what could reasonably have been checked, would a competent person have decided differently? If yes, name what they would have done. If no, say so, and protect the reasoning from the ban that is coming.

What rule carries forward? The Level 4 question. One sentence, specific to the class of decision, that would have caught this one and will catch the next.

For trainers and facilitators: the six questions are a session in themselves. Give a group a public case, the Victoria matter is widely reported, and ask them to complete the page as if they were the person. The fifth question produces the disagreement, and the disagreement is the learning.

Systems record what happened. A trace records why

The review above fails at the first question in most organisations, and the reason is not carelessness. Systems of record keep outputs: the document that went out, the order that was placed, the version history. They do not keep the reasoning at the time of the decision, the sources the person looked at, or the moment at which they committed before the consequence was known. Without that, a review is reconstructing why from what, and the reconstruction is shaped by the outcome it already knows.

AIRS-J/1 is built on the opposite rule. Every observation for a certifying assessment must be traceable to evidence in a retained transcript, which keeps the decision trace: decisions, artefacts, sources consulted, stated rationale, revisions and timings. Where consequence is revealed during a scenario, the person's position must be recorded before the revelation; a revision afterwards is an observation under P2, not a correction to P1 (Proportionate Commitment Under Uncertainty). An instrument that permits silent overwriting of a committed position is not approvable.

That rule is portable, and a risk function can adopt the principle without the instrument. For the class of AI-assisted decisions with the highest consequence in the organisation, keep the trace: what the person decided, on what, and when, recorded before the outcome. It costs a paragraph at the time. It is the difference between a review that can answer the fifth question and one that can only assign blame.

Three questions to ask on Monday

1. For the last AI-assisted decision in your organisation that went wrong, did the review ask whether the reasoning was poor or the outcome unlucky, or did it stop at the cause?
2. For the class of AI-assisted decisions with the highest consequence in your area, is the reasoning at the time of decision recorded anywhere, or only the result?
3. After the last incident, what rule was carried forward, in one sentence, and can anyone on the team state it?

The tool will be wrong again. Whether the organisation learns the right lesson depends on a review that can tell a poor decision from an unlucky one, and that review needs a record nobody is currently keeping.

Adopt the one-page review

Sources

1. ABC News (Associated Press report), 15 August 2025, on the Victorian Supreme Court matter. Limitation: press reporting; the judgment and court transcript were not read, so statements attributed to counsel and the judge are as reported. https://www.abc.net.au/news/2025-08-15/victoria-lawyer-apologises-after-ai-generated-submissions/105661208 Opened 19 October 2026.
2. On the Deloitte report to the Department of Employment and Workplace Relations, revised version dated 26 September 2025: Fortune, 7 October 2025, https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund Opened 19 October 2026; Accounting Times, 9 October 2025, https://www.accountingtimes.com.au/technology/deloitte-to-refund-government-after-using-ai-in-440k-report Opened 19 October 2026; ACS Information Age, 7 October 2025, https://ia.acs.org.au/article/2025/deloitte-to-refund-government-over-ai-errors.html Opened 19 October 2026. Limitation: press reporting; the department's own report page was not opened, and the reports differ on the refund amount, so none is stated.
3. AIR APAC (2026). AIRS-J/1 v2.2, AIR APAC Standard for Judgment in AI-Enabled Work. Sections 7.5 P2 anchor set and note, 8.6, 8.7, 8.8, 12.2.2, Annex D (CR-P2). https://airapac.org/standard/v2-2/ Opened 19 October 2026.