COSO's GenAI Guidance in Practice: Internal-Control Requirements, Model Lineage, and Audit Trail for Financial Reporting
In February 2026, COSO published guidance on achieving effective internal control over generative AI. It was not addressed to data scientists. It was addressed to the people who own controls, risk management, assurance, and financial reporting — controllers, chief accounting officers, SOX leads, internal audit, and the CFO who signs the certification.
The substance is deliberately unglamorous: generative AI is not a new framework. It maps onto the COSO 2013 components already underpinning your ICFR assertion. Which means the question in the room has quietly changed. It is no longer "should we use AI in finance." It is "which of our reported numbers now depend on an AI step, and can we evidence the control over it?"
Deloitte's 2026 CFO Signals work describes the same tension from the other side: CFOs under pressure to deploy AI quickly, while naming governance, transparency, and cost uncertainty as their live concerns. Speed without evidence is the exposure.
This piece translates the guidance into requirements you can put in a control matrix.
1. Scope by output, not by tool
The most common scoping error is inventorying tools. Tools get renamed, bundled into existing platforms, and adopted informally. Outputs do not.
The test: does the output feed a journal entry, an account balance, a disclosure, or a management review control? If yes, it is in ICFR scope — whether it came from a "copilot," an "assistant," an "agent," or a spreadsheet plugin somebody enabled last quarter.
Practical requirement: maintain a workflow register with, for each AI-assisted output, the accounts and assertions affected, the systems involved, the accountable owner, and the permitted use of the output. Refresh it annually and on material change.
2. Separate the calculation layer from the language layer
This is the single most consequential design decision, and it determines whether the rest of your control work is even possible.
A language model is non-deterministic. Run it twice on the same inputs and you may get two different sentences — and, if it is doing arithmetic, potentially two different numbers. Non-deterministic steps cannot be relied on as automated application controls, because reperformance is the basis of that reliance.
The workable architecture is a hard boundary:
- Numeric layer — deterministic. Aggregation, allocation, FX translation, variance decomposition run in a calculation engine with declared formulas against source-bound data. Testable and reperformable as an application control.
- Narrative layer — generative. Commentary, framing, and explanation are generated only from figures that already passed the numeric layer. This is a preparer's draft, subject to a management review control.
Both structures already exist in your control matrix. What is new is drawing the line explicitly and documenting which side each step sits on.
3. Bind every figure to a source record
A number that cannot be traced is not evidence, however plausible it reads.
Requirement: every figure resolves to an identifiable source record — GL account, cost centre, entity, period, contract line — retrieved through a governed connection. Nothing enters a calculation from free text, from a pasted email, or from the model's memory of a prior session.
The preventive control is a validation gate that runs before calculation:
- Completeness and control-total checks against the source
- Period cut-off and effective dating
- Currency, sign convention, and elimination policy
- Mapping coverage, where unmapped items raise an exception rather than defaulting to zero or "other"
A failed check must stop the run and raise an exception. The failure mode to design out is a validation gap that quietly degrades into an estimate.
4. Model lineage: what "lineage" actually has to contain
"Full lineage" is the most over-claimed phrase in AI-for-finance marketing. Here is the concrete list a tester will reconstruct a figure from:
| Lineage element | Why it exists |
|---|---|
| Source system, object key, extraction timestamp | Traces the figure to a record and a point in time |
| Validation results | Shows the preventive control operated, not just existed |
| Declared formula and calculation trace | Enables exact reperformance of the arithmetic |
| Model identifier and version | Establishes which system produced the narrative |
| Prompt or instruction version | Prompts change how the control operates — they are control logic |
| Mapping and driver versions, effective-dated | Lets a prior period be reproduced under the logic live at the time |
| Assumption snapshot | Assumptions drift; the run's assumptions must be frozen at execution |
| Exceptions raised and disposition | Shows the detective control operated and how items were resolved |
| Reviewer identity, decision, timestamp | Evidence of the management review control |
| Override reason, where applicable | Distinguishes judgement from bypass |
The test of whether lineage is real: pick one figure from three months ago at random and reproduce it, including the narrative's inputs, without asking anyone who was involved. If that requires a conversation, you have logs, not lineage.
5. Prompts and models are change management, not configuration
This is the requirement most teams miss, and it is the one that breaks reperformance.
A prompt change alters how a control operates. So does a model version upgrade pushed by a vendor, a mapping change, or a revised driver definition. All of it belongs under IT change management with approval, effective dating, and retained history.
Two consequences worth planning for:
- Vendor-initiated model updates need contractual notice. If your provider can silently change the model behind your reporting process, you cannot assert control over it. Put change notification in the contract.
- Effective dating is not optional. Without it, reproducing Q1 under Q3 logic gives a different answer, and reperformance testing fails through no fault of the original work.
6. Design the audit trail as an exportable package
The deliverable at testing time is not a dashboard. It is a package, per selected sample, that a tester can work through unaided: lineage, validation results, calculation trace, exceptions and disposition, assumption snapshot, and the sign-off record.
Two disciplines make this survivable:
- Test the export once per reporting cycle, not for the first time when a sample lands. Export paths rot quietly.
- Retain to your records policy in an append-only form. Evidence that can be edited after the fact is weak evidence.
7. Monitor the things that predict a control failure
COSO's monitoring component is where most AI control frameworks are thinnest. Four metrics carry most of the signal:
- Exception rate by workflow — a falling rate can mean improvement or a loosened threshold; look at both.
- Override frequency and concentration — overrides concentrated in one person or one account are a review-quality issue.
- Change volume — prompt, model, and mapping changes per period, and how many were retrospective.
- Access drift — who gained the ability to configure, run, or approve since the last review.
Report these to the audit committee on a fixed cadence. Entity-level monitoring is what turns a set of process controls into a governance posture.
What this looks like when it is done
The finished state is unremarkable, which is the point. A monthly variance pack arrives with each figure traceable to its GL line, the arithmetic reperformable, the commentary drawn only from validated figures, exceptions listed rather than smoothed, and a named reviewer's approval attached — reproducible six months later by someone who was not there.
That is not a constraint on using AI in financial reporting. It is the condition that makes it usable at all.
Where to go next
- Framework view: AI governance for financial reporting maps the five COSO components, ICFR scoping questions, and SOX expectations against what a governed AI workflow produces.
- Testing prep: the auditor reliance checklist turns the above into requirements with named evidence and sign-off, and exports as a PDF.
- Using LLMs safely in the meantime: the CFO AI prompt pack adds citation and uncertainty-handling rules to FP&A and risk prompts.
This article is provided for planning purposes and is not accounting, audit, or legal advice. Confirm scoping and evidence expectations with your external auditor and internal controls team.
Frequently asked questions
What does COSO's generative AI guidance require of finance teams?+
It applies the existing COSO 2013 internal control components to generative AI rather than creating a separate framework. For finance teams that means scoping AI-assisted outputs into ICFR, assessing AI-specific risks such as fabricated figures and undisclosed model change, designing preventive and detective controls at the point the number is produced, retaining traceable evidence, and monitoring exceptions, overrides, changes, and access.
Are AI prompts and model versions subject to SOX change management?+
Treat them as such. A prompt, model, mapping, or driver change alters how a control operates, so it belongs under change management with approval, effective dating, and retained history. Without effective dating you cannot reproduce a prior period under the logic that was live at the time, and reperformance testing fails.
What must model lineage contain to satisfy an auditor?+
Source system and object key with extraction timestamp, validation results, the declared formula and calculation trace, model and prompt versions, effective-dated mapping and driver versions, the assumption snapshot, exceptions raised and their disposition, and the reviewer's identity, decision, and timestamp — enough to reproduce a figure selected months later without consulting the people involved.
Can generative AI steps be relied on as automated controls?+
No, because they are non-deterministic and reperformance is the basis of that reliance. Keep arithmetic in a deterministic calculation layer that can be tested as an application control, and treat generated narrative as a preparer's draft subject to a management review control.
How do we scope AI-assisted processes into ICFR?+
Scope by output rather than by tool. If an output feeds a journal entry, an account balance, a disclosure, or a management review control, it is in scope regardless of what the tool is called. Maintain a workflow register recording the accounts and assertions affected, the systems involved, the accountable owner, and the permitted use of each output.
Analyze this article with AI
Get an instant executive brief right here, or open it in your preferred AI assistant.
Executive brief — generated in-app
Tailored for CFOs and VPs of FP&A. Takes ~5–15 seconds.
Prefer another assistant? If the prompt doesn't autofill, use the fallback below.
Deep link didn't work? .
CFO briefing
Get the CFO governance briefing
Practical internal-control guidance for AI-assisted financial reporting — written for controllers, SOX leads, and CFOs. Roughly monthly, no marketing.
- COSO GenAI mapping updates as guidance evolves
- ICFR scoping and SOX testing patterns for AI-assisted numbers
- New templates, checklists, and auditor-ready evidence formats
Next step
See the full picture of AI in Finance
The comparison above is one slice. Explore the complete AI in Finance hub for the pillar view — workflows, vendor archetypes, governance, and the 12-month sequence — then book a working session on your own numbers.
Stay ahead in AI-powered finance
Get insights on AFI capabilities, FP&A automation, and enterprise AI delivered to your inbox.

