AI in finance and accounting
Two of R220.4’s six sub-paragraphs decide what an AI-assisted number has to survive: the accountant must avoid undue reliance on technology, and must be aware of the risk of bias. Both are mandatory. Neither came from the Code’s technology revisions — they arrived with Role and Mindset, before generative AI was a public phenomenon.
§ 1 — In practice
R220.4 requires the accountant preparing or presenting information to do six things, and two of them are what an AI-assisted process most often fails. Sub-paragraph (e) requires them to “Avoid undue influence of, or undue reliance on, individuals, organisations or technology”. Sub-paragraph (f) requires them to “Be aware of the risk of bias”. Both are mandatory, and neither arrived with the Code’s technology project — both came in with the Role and Mindset revisions. What (e) does structurally is place technology in the same clause as people and organisations, which means the test is the ordinary one: not whether the tool was good, but whether the reliance was undue. Undueness is a property of the decision to rely, assessed against what the accountant knew and what they could have found out.
The other four sub-paragraphs supply the standard an output has to meet before it becomes information the accountant has presented. Preparation in accordance with a relevant reporting framework, where one applies. Presentation “in a manner that is intended neither to mislead others nor to influence contractual or regulatory outcomes inappropriately”. The exercise of professional judgement to represent the facts accurately and completely in all material respects, to describe clearly the true nature and impacts of transactions, and to collect, classify, record and measure information in a timely and proper manner. And no omission intended to render the information misleading. Together with (e) they answer the question practitioners actually ask about a model-drafted note: the standard has not moved, and the only thing that changed is how much of the drafting a person did — which Article 50 treats as a disclosure question and the Code treats as a preparation one.
R220.5 is where this becomes a planning and forecasting page rather than a general one. It forbids exercising the discretion that professional judgement involves with the intention of misleading or of influencing contractual or regulatory outcomes inappropriately, and 220.5 A1 gives examples: determining estimates such as fair value; selecting or changing an accounting policy among permitted alternatives; determining the timing or structuring of transactions; selecting disclosures; and — the one that reaches model output head-on — “Preparing forward-looking information by relying on assumptions that are unrealistic or inconsistent with management’s decisions or objectives in order to mislead.” A forecasting model runs on assumptions, and the accountant either chooses them, accepts the supplied defaults, or lets the tool infer them from history. All three are exercises of the discretion this provision governs, and the judgement R220.8 requires about an output does not reach back to cover the assumptions that went into it.
R220.6 completes the sequence and is the provision nobody applies to model output. It requires the accountant to identify and consider the purpose, the context and the audience of the information, and 220.6 A1 notes that including relevant estimates, approximations and assumptions allows those who might rely on the information to form their own judgements. Applied to an AI-assisted forecast that is a disclosure duty stated in the Code: the reader of a board paper is entitled to enough about the assumptions to judge the number, and a forecast whose assumptions were set inside a tool and never surfaced does not give them that. The duty runs to the audience, which is why it is not discharged by a working paper the audience will never see.
§ 2 — What discharges it
The artefacts an assessor asks to see, and what makes each one sufficient rather than merely present.
01
Separating what the tool generated from what management set, with the source of each. It is what makes the difference between a forecast and an output nobody can interrogate.
02
The check against a separate source or a prior-period expectation, and what it showed. The point is not that the figures agreed; it is that something other than the model was consulted before the number was used.
03
Who the information is for, what decision it supports, and what they are being asked to take on trust. R220.6 asks for this in terms, and very few management packs carry it.
04
Which known distortions could reach this figure through the data or the method, and what was done about them. Sub-paragraph (f) is a duty of awareness, and awareness either leaves a note or leaves nothing.
05
Tool, settings and date, held with the deliverable. Where a covenant or an incentive turns on the number, the question of what produced it arrives months later, by which time the configuration has moved on.
§ 3 — Who it binds
Any professional accountant in business preparing or presenting information, financial or non-financial, for use inside or outside the employing organisation: management accounts, forecasts, board papers, regulatory returns, tax computations. The provisions reach information a tool drafted, estimated or generated from the moment the accountant prepares or presents it. The Code takes effect through each IFAC member body that has adopted it.
§ 4 — Worked example
An FP&A team replaces its bottom-up revenue forecast with a model that projects each customer’s spend from three years of history, setting churn and expansion assumptions itself. The output is four per cent above the sales team’s own number, which the CFO prefers, and it goes into the board pack labelled “model-generated forecast”. The assumptions sit in the tool’s configuration and nobody has written them down. The covenant headroom calculation uses the same figure.
Where does R220.4 bite, and does the label help?
The label is the weakest part of it. Calling the number model-generated tells the board where it came from and nothing about what it rests on, which is the opposite of what R220.6 asks: purpose, context and audience, with enough of the estimates and assumptions for a reader to form their own judgement. That reader is being asked to accept a churn assumption nobody has stated. Sub-paragraph (e) then bites on the choice itself — preferring the higher number because a model produced it, in circumstances where the people closest to the customers produced a lower one, is the pattern it describes. And the covenant calculation moves this out of internal reporting altogether. Once the figure influences a contractual outcome, R220.5 is live, and the question a member body would put is not whether the model was accurate but how the discretion over its assumptions was exercised. None of that requires the forecast to be wrong. It requires the assumptions to be visible, which is a schedule rather than a project.
§ 5 — What a weak answer looks like
The tool named instead of the assumptions. A board pack marks the figure as AI-assisted, sometimes with a sentence about the model’s accuracy on historical data, and the finance team treats that as the disclosure. It discloses provenance, which nobody was going to dispute, and withholds the four or five assumptions on which the number entirely depends. What R220.6 asks for is the material that lets a reader form their own judgement about the information. A provenance label lets them form a view about the software.
§ 6 — When it applies
29 August 2024
Publication date of the 2024 edition of the IESBA Handbook, used as an edition marker. Sub-paragraphs (e) and (f) came in with the Role and Mindset revisions, whose IESBA commencement date has not been verified; the remainder carries no separate effective date. Not a deadline.
§ 7 — Exposure
No fine under the Code; the exposure is personal and runs through the adopting body
R220.4 and R220.5 are mandatory provisions, and a breach is investigated by the IFAC member body that has adopted the Code rather than by a regulator. Sanctions run from reprimand and fine to conditions on practice and withdrawal of membership and the practising certificate. Where the same information reaches a market or a lender, separate statutory and contractual consequences may attach under other instruments; the Code neither creates nor measures those.
§ 8 — Elsewhere
Where another instrument addresses the same obligation. These are correspondences, not comparisons — the Council does not rank one framework against another.
Article 50 requires certain AI-generated content to be disclosed as such to the people who receive it. R220.6 asks a different question of the same deliverable: what the audience needs about the assumptions in order to judge the number.
Finance and accounting standards
R220.8 governs the judgement about using an output. R220.4 supplies the standard that output must meet once the accountant has prepared and presented it as information.
A correspondence indicates that two instruments address the same underlying obligation. It is not a mapping endorsed by either body, not a statement that one satisfies the other, and not a judgement about which is more demanding.
§ 9 — Where this is assessed
Examined in one credential, in the domains named on each card.
AI in analysis, forecasting and FP&A · 25% of the paper
For finance and accounting professionals who use AI in analysis, close and reporting work — and verify the output before the numbers are relied on.
§ 10 — provenance
This page sets out what the instrument requires and what discharges it. The official text is the authority — these go straight to it.
§ 11 — Also read
Certification
What this page sets out is examined, on work the candidate has not seen, in the credential beside it. Every paper is machine marked against its key and every result is reviewed and released by a named person, and every credential is on the public register.
AIAC-FNAF
Certified AI in Finance & Accounting — Fundamentals2 hours · remote, proctored · valid 2 years
Ten questions on using AI at work. About two minutes, free. You finish as one of six types, with a card to share.
A free quiz from the AI Assurance Council. Not a certification.