AI governance for financial reporting,built as an internal control.
For CFOs and chief accounting officers of public and regulated companies: how AI-assisted FP&A, close, and reporting work fits inside COSO, ICFR scoping, and SOX evidence — with deterministic calculation and lineage on every number.
What changed for public companies
AI in finance is now a controls question, not a tooling question.
In February 2026 COSO published guidance on achieving effective internal control over generative AI, addressed to the people responsible for controls, risk management, assurance, and financial reporting. It does not create a parallel framework — it maps generative AI onto the 2013 framework already underpinning your ICFR assertion.
Deloitte's 2026 CFO Signals work found North American CFOs pressed to deploy AI quickly while worrying about governance, transparency, and cost uncertainty. That is the real tension: the mandate is to move, and the exposure is a number no one can trace.
The way out is not slower adoption. It is AI whose output arrives with its evidence attached.
COSO alignment
The five COSO components, applied to AI-assisted reporting.
Same framework you already report against. New failure modes to control for.
Control environment
Name an accountable owner for every AI-assisted process that touches the financial statements. Board and audit-committee oversight of AI use is documented, and the policy states explicitly which outputs may and may not reach a reported number.
Risk assessment
Assess each AI-assisted workflow for the risks generative systems actually introduce: fabricated figures, silent prompt or model change, drifting source data, and over-reliance on unreviewed output. Rate them by their effect on material account balances and disclosures.
Control activities
Preventive and detective controls at the point the number is produced: deterministic calculation, validated source binding, completeness checks, threshold-based exception routing, and a reviewer sign-off that cannot be bypassed.
Information & communication
Lineage travels with the output. Every figure carries its source system, extraction time, transformation applied, model and prompt version, and reviewer — so the number and its evidence are never separated in transit.
Monitoring
Continuous monitoring of exception rates, review overrides, model and prompt changes, and access. Deviations raise a finding for the control owner rather than surfacing at year-end testing.
ICFR scoping
Five questions to answer for every AI-assisted step.
Is the AI step in scope for ICFR?
If the output feeds a journal entry, an account balance, a disclosure, or a management review control, it is in scope — regardless of whether the tool is labelled a copilot, an assistant, or an agent.
Is it an application control or a manual control?
Generative steps are not deterministic and should not be relied on as automated application controls. In Octopus AI, the numeric layer is deterministic and testable as an application control; the narrative layer is treated as a preparer's draft subject to a manual review control.
What is the evidence of operation?
A retained, immutable record per execution: inputs, versions, calculation trace, exceptions raised, reviewer identity, and timestamped approval — reproducible on demand for a sample selected months later.
How is change managed?
Model, prompt, mapping, and driver changes move through change control with approval and effective dating, so a prior period can be reproduced under the logic that was live at the time.
Who has access, and can they change a number?
Least-privilege, role-based access with segregation between the person who configures a workflow, the person who runs it, and the person who approves its output.
SOX readiness
What testers ask for — and what Octopus AI produces.
| Control area | Auditor expectation | How Octopus AI satisfies it |
|---|---|---|
| Entity-level controls | Documented AI policy, accountable owner, audit-committee visibility | Governance pack per workflow: owner, purpose, in-scope accounts, permitted outputs, escalation path |
| Process-level controls | Preventive and detective controls around the AI-assisted step | Source binding validation, completeness and reconciliation checks, exception thresholds, mandatory reviewer approval before release |
| ITGC | Access, change management, and operations over the AI system | RBAC and SSO, versioned prompts and models under change control, execution logs retained with the period's evidence |
| Management review controls | Evidence the reviewer engaged with precision, not rubber-stamped | Reviewer sees the figure, the lineage, the exceptions, and the assumption set; overrides require a written reason and are logged |
| Evidence & retention | Reperformable population and sample evidence | Immutable per-run record, exportable as an audit package for a selected sample or full population |
How we calculate
Deterministic by construction — the model never does the arithmetic.
Six steps run on every figure. Each one is loggable, testable, and reperformable.
Bind to source, not to text
Every figure resolves to a specific source record — GL account, cost centre, entity, period, or contract line — retrieved through a governed connection. Nothing enters a calculation from free text or from a model's memory.
Validate before computing
Completeness, period cut-off, currency, sign convention, and mapping coverage are checked first. A failed check stops the run and raises an exception; it never degrades into an estimate.
Compute symbolically
Arithmetic, aggregation, allocation, FX translation, and variance decomposition execute in a deterministic calculation engine with declared formulas. The language model does not perform arithmetic and cannot alter a result.
Reconcile and prove
Outputs tie back to the source total. Any residual is reported as a labelled difference rather than absorbed silently, so a reviewer can see the tie-out rather than trust it.
Explain from validated figures only
Narrative and commentary are generated strictly from figures that already passed the steps above, each carrying its lineage reference. A figure without lineage is flagged, never published.
Route for review and record the sign-off
The reviewer approves, requests change, or overrides with a written reason. The decision, identity, and timestamp are retained with the run and are reproducible under audit selection.
The control boundary. Language models retrieve, structure, and explain. A symbolic engine computes. Because the two layers are separate, the numeric layer is testable as an application control and the narrative layer is reviewed as a preparer's draft — a structure your auditors already know how to test.
Auditable AI reporting
What makes an AI-assisted management report auditable
Octopus AI attaches five kinds of evidence to every AI-assisted report: source-level references for each figure, a calculation that can be re-run with the same result, the budget owner's explanation, a list of exceptions still open, and the named person who approved release.
AI does not certify a report's accuracy or its compliance. The evidence lets a finance reviewer, and later an auditor, check the work; the approval and the accountability stay with named people.
Illustrative checklist: an example of what a finance reviewer checks before releasing AI-generated variance commentary. Adapt it to your own control documentation.
Source references
Every figure in the commentary links to its source records: entity, account, period and transaction lines.
Reproducible calculation
Re-running the calculation on the same inputs gives the same figures, and totals tie back to the source.
Figures match the narrative
Every number in the text matches the calculated figure; nothing appears that has no lineage.
Owner explanation
Each material variance carries the budget owner's own explanation, not only a drafted one.
Unresolved exceptions
Open exceptions are listed and either resolved or disclosed; none were closed without a reason.
Versions recorded
The budget version, assumption set, model and prompt versions used are recorded with the run.
Named approval
The reviewer's name, decision and timestamp are logged; any edit or override has a written reason.
See the workflows this applies to: budget variance analysis, rolling forecasts and ERP and planning integration. Evidence from one anonymized reference deployment is in the case study.
Assurance posture
Evidence mapping
Pick an AI-assisted number. See the controls, lineage, and sign-off it needs.
Work down the list with your control owner. Tick each item you can already evidence, then export the result as a working paper for your SOX or internal audit file.
Period variance by entity, department, and account, decomposed into price, volume, mix, and timing drivers.
Controls
What is tested to rely on the number
Model lineage
What must be logged to reperform it
Sign-off artifacts
Who signs, and what evidence they leave
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
FAQ
AI governance and internal control — CFO questions
Does COSO's generative AI guidance mean AI is now an internal control issue for public companies?
Yes. COSO's 2026 publication on internal control over generative AI is aimed at practitioners responsible for controls, risk, assurance, and financial reporting, and it maps generative AI onto the existing 2013 framework rather than treating it as a separate technology topic. Practically, that means any AI-assisted step that affects a reported number belongs in your ICFR scoping, risk assessment, and control documentation — not only in your IT risk register.
Can AI-generated numbers be relied upon for financial reporting?
They can be relied upon when the number is not generated by a language model. In Octopus AI the numeric layer is deterministic — declared formulas executed against source-bound data — while the model handles retrieval framing and narrative. Reliance then rests on testable application controls plus a management review control over the narrative, which is a structure auditors already understand.
How do we scope AI-assisted processes into ICFR?
Scope by output, not by tool. If the output feeds a journal entry, an account balance, a disclosure, or a management review control, it is in scope. Document the workflow, the accounts and assertions it touches, whether each step is deterministic or generative, and the preventive and detective controls around it.
What audit evidence should an AI workflow produce?
A retained, immutable record per execution: source systems and extraction timestamps, transformations and declared formulas, model and prompt versions, validation results, exceptions raised and their disposition, the reviewer's identity, and the timestamped approval. Evidence must be reproducible for a sample selected months after the fact.
Is prompt and model change management a SOX requirement?
Treat it as one. A prompt, model, mapping, or driver change alters how a control operates, so it belongs under IT 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, which undermines reperformance.
How do you prevent hallucinated figures in a reported number?
Structurally, rather than by instruction. Figures come only from validated, source-bound retrieval; arithmetic is symbolic; narrative is generated exclusively from figures that already carry lineage; and anything without lineage is flagged rather than published. The model has no path to invent a number that reaches an output.
Who signs off on AI-assisted financial reporting output?
The human control owner for the account or process — unchanged from today. What changes is the precision of the review: the reviewer sees the figure, its lineage, the exceptions raised, and the assumptions used, and any override is recorded with a written reason.
How does this map to the EU AI Act and other AI regulation for finance functions?
The same primitives satisfy most obligations: documented purpose and ownership, risk assessment, human oversight of output, logging and traceability, change control, and data governance. Building them for ICFR generally puts the finance function ahead of AI-specific regulatory requirements rather than duplicating work.
Which AI financial reporting tools provide source-level data lineage and approval workflows?
Octopus AI links every figure in an AI-assisted report to its source records, records the calculation, model and prompt versions used, and holds the report until a named reviewer approves it. The approval, reviewer identity and timestamp are kept with the run.
How do enterprise finance teams validate AI-generated variance commentary before publishing management reports?
A finance reviewer checks that each number in the commentary matches a calculated, source-linked figure; that the calculation can be re-run with the same result; that material variances carry the budget owner's explanation; that open exceptions are resolved or disclosed; and then records a named approval. AI drafts the commentary but does not certify it.
Practical control requirements: translating COSO's GenAI guidance into controls, lineage, and audit trail.
Preparing for testing: the auditor reliance checklist maps each AI-generated number to evidence, controls, and sign-off.
Building the business case: finance AI ROI calculator and AI in finance.
Bring your control owner and your auditor.
30 minutes on a real workflow: the number, its lineage, the exceptions it raised, and the sign-off record — the same package you would hand to a tester.
Book a DemoPublished AI trust documents
Octopus AI publishes how its AFI capabilities behave, what they may not do, and how each figure is verified.