AI in finance and accounting
Nothing compels a finance function to follow COSO. What Principle 10 supplies is the vocabulary that reporting regimes borrow, and COSO’s 2026 paper on generative AI draws a line inside it: reliance begins where the evidence for a control is the tool’s output and nothing else. A person who re-performs the work on a sample has moved back across that line.
§ 1Who it binds
Any control activity in a finance function where an AI tool produces, proposes or clears part of the work — matching, reconciliation, classification, extraction, summarisation. Principle 10 binds nobody: COSO’s Framework is voluntary and no consequence attaches to departing from it. It reaches this work because it is the account of internal control that management, boards and auditors already share.
§ 2In practice
Principle 10 sits in the Control Activities component of COSO’s Internal Control — Integrated Framework, and asks an organisation to select and develop control activities that reduce risk to an acceptable level. Nothing in it binds anyone. COSO is a private-sector initiative sponsored by five accounting and auditing bodies; it issues no rules, imposes no penalty, and has no power to require anything of anyone. Its authority is borrowed. It is the settled account of what an internal control system looks like that preparers, boards and auditors reach for when they need one, and in the United States it is the framework management overwhelmingly identifies when a suitable, recognised one has to be named.
COSO’s 2026 paper on generative AI adds the distinction the Framework itself does not carry. The paper’s working definition of reliance is that management depends on an AI system’s output as part of the evidence supporting a control’s design or operating effectiveness. Its illustration is a pair. In the first, a tool matches journal entries to supporting documentation and the control owner approves the posting on the strength of the tool’s output alone: that is reliance. In the second, the tool proposes matches and a person re-performs them on all or a sample of items and documents the review and approval: that is not. The ethics duty at R220.8 asks the accountant to judge what verification a use of technology needs; COSO describes what follows from the answer.
The consequence practitioners miss is that the boundary has nothing to do with how good the model is. A tool with a very high match rate on which nobody re-performs anything is a reliance case; a tool with a poorer one behind a documented sample re-performance is not. Model quality changes how many exceptions arrive; it does not change who is standing behind the control. This is also why "we validated the vendor" does not answer the question Principle 10 asks — validation is evidence about a product, and the Principle is asking about a control activity, which includes the person at the end of it. The same line seen from the audit side is testing a probabilistic control.
Where reliance exists, the 2026 paper’s position is that the pattern should meet the evidence standards already expected of internal control over financial reporting: prompt, configuration and model version documented; a stated sampling rationale; exception resolution recorded; and enough retained to support both design and operation. That is a materially heavier file than most finance functions keep for a tool they think of as a productivity aid. The cheaper route is usually the other one — a documented sample re-performance that keeps the process out of reliance altogether — and it is a choice most control owners have never consciously made, because nobody told them their review note was doing that job.
§ 3What discharges it
The artefacts an assessor asks to see, and what makes each one sufficient rather than merely present.
01
A recorded answer to one question — did a person re-perform the work, and on what — made when the control is designed rather than reconstructed when it is challenged. Most functions have the answer and have never written it down.
02
How many items, chosen how, out of which population. "Spot-checked" is not a basis; a stated rule is what lets a second person repeat the check and land in the same place.
03
Version, thresholds, prompt or rule set, and the date. Captured afterwards it is a reconstruction, and a tool that was updated mid-period cannot be reconstructed at all.
04
What the tool could not clear, who worked it, what they concluded, and when. Exceptions are the visible part of the process and the part finance teams already document; the gap is everywhere else.
05
Where the impact is significant, proof that the tool assists and a named person decides — not a policy saying so, but the record of the decision showing the two apart.
06
A description written before the tool arrived describes something that no longer exists. Naming the tool is what makes the reliance question visible to anyone reviewing the control inventory.
§ 4Worked example
A shared-services team runs an AI tool across intercompany balances at each month end. The tool proposes matches between the two ledgers, clears everything it can, and produces an exceptions list. The team lead opens the exceptions, works those, and signs the reconciliation. The cleared population — around 94% of items — is not opened at all, and the sign-off reads "reconciliation reviewed and approved".
Has the team relied on the tool, and does it matter which answer is right?
Yes to the first, and the second is where the cost sits. Nothing in the file evidences that any of the 94% was looked at, so the only support for the cleared population is that the tool cleared it — reliance on COSO’s 2026 working definition. The exceptions work does not change that, because exceptions are the population the tool could not clear rather than a sample of what it did. Once the process is a reliance case, the evidence expected of it rises to what internal control over financial reporting demands: which version of the tool ran, on what configuration and thresholds, what the matching rule actually was, and how each exception was resolved. Almost none of that is captured today. The alternative costs an hour a month. Pull a defined sample from the cleared population, re-perform the match, and record the sample basis and the result beside the sign-off. The reconciliation is then supported by a person’s work, and the tool is a productivity aid again.
§ 5What a weak answer looks like
The vendor’s accuracy figure offered as the answer. A finance function that has read the benchmark, run a pilot and satisfied itself the product performs well concludes that its control is sound, and it may well be right about the product. It has answered a question about software and left the question about the control open: nobody has said whether a person stands behind the output or merely beside it. The tell is a control description that still describes the process as it ran before the software arrived.
§ 6Elsewhere
Where another instrument addresses the same obligation. These are correspondences, not comparisons — the Council does not rank one framework against another.
The same boundary from the auditor’s side: where management has relied on a model, the engagement has to decide what a sample of outputs establishes about the control as a whole.
Article 14 requires a high-risk system to be designed so that a person can oversee it. Where a finance tool falls in that class, the oversight this Principle expects has a design counterpart on the provider’s side.
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.
§ 7When it applies
1 January 2013
Publication year of COSO’s Internal Control — Integrated Framework. The day is not asserted; the field carries the start of that year so the edition can be dated. A publication date, never a commencement — the Framework is voluntary and no consequence attaches to departing from it.
1 January 2026
Publication year of COSO’s Achieving Effective Internal Control Over Generative AI (GenAI). The day is not asserted. It is thought leadership commissioned by COSO rather than a Framework amendment: it creates no principle and binds nobody.
§ 8Where this is assessed
Examined in 2 credentials, in the domains named on each card.
Verifying AI output before reliance · 20% 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.
AI inside internal control and the close · 20% of the paper
For senior finance and accounting professionals who set the terms on which AI is used across a finance function — and answer for the estimates, controls and disclosures it touches.
§ 9provenance
This page sets out what the instrument requires and what discharges it. The official text is the authority — these go straight to it.
§ 10Also read
Certification
What this page sets out is examined, on work the candidate has not seen, in the credentials 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
AIAC-FNAP
Certified AI in Finance & Accounting, Professional2 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.
SHARE BY 31 DECEMBER, US$199 FEE WAIVED
A free quiz from the AI Assurance Council. Not a certification.