The trigger is lower than most readers assume. R220.9 engages where the accountant knows or has reason to believe that information they are associated with is misleading, and “reason to believe” is not certainty about the number — it is a state of knowledge about how the number came about. Someone who knows a forecast was generated by a tool nobody in the function can explain, on assumptions nobody has seen, is not free to treat the resulting doubt as another department’s problem. Association is the other half of the trigger and it is where AI work gets misread: a figure a model produced and an accountant passed on is the accountant’s information. The tool is not a party to anything and cannot be associated with anything.
What the provision requires is action to seek to resolve the matter, and the application material sets that out in ascending order: raise it with a superior; then with management or those charged with governance, asking that the information be corrected or that users be informed; use the organisation’s ethics or whistleblowing policy; consult a professional body, the internal or external auditor, or legal counsel; and determine whether any duty exists to communicate to third parties or to a regulator. Every rung is a step somebody can date, which is the practical reason this is a procedural duty rather than a matter of conscience. Where the underlying failure sits in a system rather than in a person, an escalation of the kind Article 73 contemplates may be running in parallel, to a different recipient and on a different clock.
R220.10 is the provision this page exists for, and it is the only obligation in the corpus whose discharge may require the professional to walk away. Where, after exhausting all feasible options, the accountant determines that appropriate action has not been taken and there is reason to believe the information is still misleading, they shall refuse to be or to remain associated with it — and the application material notes that resignation may be appropriate. Three words carry the weight. *Exhausted* sets a high bar for stopping earlier. *Determines* makes the conclusion an act with an author and a date. And *remain* reaches information already issued, not only information about to be, so an accountant who learns afterwards that a model-produced disclosure was misleading does not discharge the duty by declining to repeat it.
Two features of AI-assisted work make the ladder harder to climb, and neither alters the duty. The first is diffused authorship: when a figure emerged from a tool that three teams configured, the person to raise it with is genuinely unclear — and the provision’s answer is that unclarity about the recipient is not a reason to stop, because the ladder is defined by seniority and role rather than by who built the model. The second is evidential asymmetry. A colleague who produced a number can be asked how. A model cannot, so the accountant’s belief will usually rest on a reconciliation that failed rather than on an account of what went wrong. That is why the contemporaneous memorandum recording *why* the information was believed misleading is the artefact the whole matter later turns on, and why an audit trail over the tool is worth far more to the accountant than to the engineers who usually ask for it.