The thesis: the AI already inside your close now has a named control to sit under.
Most finance teams did not wait for a framework before turning on GenAI. Reconciliation agents, forecast assistants, and document-summarizing copilots have been moving into ERP and close workflows since 2025, often adopted by a business team rather than procured through a formal review. COSO’s new guidance does not slow that adoption down. It gives it a control language.
“Achieving Effective Internal Control Over Generative AI” takes the same Internal Control Integrated Framework that already underpins SOX compliance work and translates it into eight capability types with specific, audit-ready control expectations attached to each one. For a finance team that already runs disciplined ICFR, the work is mostly classification and evidence design, not a new program built from nothing.
What COSO actually published.
COSO released the guidance on February 23, 2026, aimed at management, compliance teams, controllers, IT governance, audit committees, external auditors, and internal audit functions. Lucia Wind, COSO’s executive director and chair, framed the risk plainly: generative AI, in her words, “can be confidently wrong, easily manipulated, or deployed outside formal oversight channels.” The guidance exists to close that gap with structure rather than restriction.
The framework applies the five components of the Internal Control Integrated Framework, control environment, risk assessment, control activities, information and communication, and monitoring activities, to a capability-first taxonomy rather than a tool-by-tool or vendor-by-vendor list. That choice matters: the same control expectation applies whether a reconciliation agent runs on an in-house model or a purchased ERP add-on, because the guidance classifies by what the system does, not who built it.
Why the control evidence has to move upstream.
EY’s analysis of the guidance describes the underlying shift clearly: control can no longer focus only on recorded outcomes, it has to extend to how a decision was formed. For a deterministic system, testing the output was usually enough. For a probabilistic model that can produce a different answer from the same inputs, the evidence a control needs is contemporaneous: what prompt ran, what the model returned, who reviewed it, and whether the review actually challenged the output rather than approving it by default.
Deloitte’s summary lands on the same practical point from the audit side. Reconciliations, journal entries, and estimate calculations touched by GenAI still need human review before they reach the books, and that review needs to be documented well enough that an auditor can trace it back: model or version, prompts, key inputs and outputs, and the approval that followed.
The six-step roadmap.
| Step | What it produces |
|---|---|
| 1. Govern | Stand up cross-functional oversight, defined roles, accountability, and an escalation path that covers every GenAI implementation, not just the ones finance built itself. |
| 2. Inventory | Catalog every GenAI use case in use, and map each one to the business process, financial statement assertion, and existing key control it touches. |
| 3. Assess | Evaluate risk using a framework built for probabilistic systems: how a GenAI output could be wrong, how confidently wrong, and how that error could reach a recorded transaction or disclosure. |
| 4. Design | Build COSO-aligned controls addressing access, inputs, prompts, outputs, logging, and performance monitoring for each capability in the inventory. |
| 5. Implement | Roll the controls out with tighter rigor wherever the capability touches a financially sensitive process, transaction posting and reconciliation first. |
| 6. Monitor | Track model performance continuously, watch for drift and anomalies, and keep the evidence of control effectiveness current rather than reconstructed at year end. |
The order matters. Skipping straight to design, buying a control tool before the inventory and risk assessment exist, is the fastest way to end up with controls sized for the wrong capability.
The evidence record a GenAI-touched transaction needs.
Deloitte’s reading of the guidance points to four control building blocks: input and retrieval controls, output validation with human review, logging and traceability, and ongoing monitoring metrics such as transaction volume and override rate. A single record like the one below can carry all four for one transaction.
Example GenAI control evidence record
{
"capability": "automated_transaction_processing_reconciliation",
"coso_component": "control_activities",
"system": "ap_matching_agent_v3",
"model_ref": "finance-tuned-model-2026-06",
"prompt_ref": "prompt-lib/ap-match/v12",
"transaction_ref": "je-2026-10-00483",
"confidence_score": 0.93,
"threshold_policy": "review_required_below_0.95_or_above_10000_usd",
"human_review": {
"required": true,
"reviewer": "ap-controller-214",
"review_date": "2026-10-14",
"outcome": "approved_with_edit"
},
"override": {
"occurred": false
},
"monitoring": {
"period_transaction_volume": 1240,
"override_rate_pct": 2.1,
"drift_alert": false
},
"evidence_ref": "s3://ai-control-evidence/2026-10/je-00483/"
}The structure repeats for every capability type. What changes is the threshold policy and the shape of the review: a forecasting model needs a documented challenge of its reasoning, while a transaction-processing agent needs a dollar or confidence threshold that routes the item to a person before it posts.
Control design by capability type.
| Capability | Control pattern | Evidence |
|---|---|---|
| Data ingestion and extraction | Retrieval-scope limits and source-document validation before content reaches a financial process. | Source log, retrieval config, validation exceptions. |
| Data transformation and integration | Field-level mapping rules reviewed and versioned, with reconciliation between source and transformed values. | Mapping version history, transformation diff, tie-out. |
| Automated transaction processing and reconciliation | Confidence-threshold and dollar-threshold routing to human review, with every override logged. | Override log, reviewer sign-off, match exception report. |
| Workflow orchestration and autonomous task execution | Scoped permissions per task, an approval gate before anything posts, and a kill switch a control owner can pull. | Permission scope config, approval log, execution trace. |
| Judgment, forecasting, and insight generation | Mandatory human challenge before an estimate or forecast is booked, with the AI output kept separate from the final judgment. | Review memo, output-versus-judgment comparison. |
| AI-powered monitoring and continuous review | An independent recalibration schedule with drift thresholds that alert a named control owner, not just a dashboard. | Drift metric history, recalibration log, alert acknowledgment. |
| Knowledge retrieval and summarization | A citation-back-to-source requirement on every summary used in a filing, memo, or disclosure, plus periodic spot checks. | Citation trace, sample review log. |
| Human-AI collaboration | A documented RACI for every AI-assisted task, and a flag in the workpaper wherever AI materially shaped the output. | RACI matrix, workpaper AI-use flag. |
Most finance teams already have three or four of these capabilities live somewhere in the ERP stack. The taxonomy is the fastest way to find out which ones still lack a named control owner.
Implementation checklist for finance and internal audit.
Inventory every GenAI use case touching a financial process today, not just the ones a finance-AI vendor sold, and map each to the ICFR key control it sits inside or next to.
Classify each use case by COSO's eight capability types so the control expectation matches what the system actually does, rather than a generic AI policy applied uniformly.
Set confidence and dollar thresholds that route a transaction to human review, and log every override with a reviewer identity and rationale.
Require logging of model or version, prompt, key inputs and outputs, and approval for any GenAI action that touches a recorded transaction or disclosure.
Build a recalibration and drift-monitoring cadence with a named control owner, so model performance changes surface before they reach the close.
Update the SOX control matrix and walkthrough documentation to name the GenAI capability explicitly, since Section 404 applies to AI-driven controls with no carve-out.
Brief the audit committee and external auditor on the inventory and control design before year end, ahead of any request for evidence.
Retire or redesign any GenAI use case that cannot currently produce the logging and human-review evidence the capability requires.
Start with the inventory. A finance team is often surprised by how many GenAI touchpoints already exist once someone walks every process end to end instead of asking IT for a software list.
Constructive failure modes to design around.
Treating GenAI tools as outside the scope of ICFR because they were adopted by a business team rather than procured through IT, when COSO's guidance and Section 404 both apply regardless of how the tool arrived.
Running a confidence-scored transaction-processing agent without a threshold policy, so review only happens when someone happens to notice an odd result.
Letting a model update or prompt change go live without a version log, so a control that operated effectively in September cannot be shown to be the same control in December.
Building a monitoring dashboard for drift without naming who acts on an alert, which produces a metric nobody is accountable for.
Assuming a human review checkbox is sufficient evidence, when what the framework calls for is a documented, reasoned challenge of the AI output, not a rubber stamp.
Waiting for the external auditor to ask about AI controls before building the inventory, instead of walking in with the mapping already done.
Each of these is fixable with existing control discipline. None require waiting for a new standard, since COSO built this guidance from the framework most SOX programs already run on.
What finance teams should ask ERP or finance-AI vendors now.
Can the system log model or version, prompt, key inputs and outputs, and the identity of any human reviewer for every action that touches a financial transaction?
Does it support configurable confidence and dollar thresholds that route a transaction to human review automatically, with every override captured?
Can it produce a capability-by-capability inventory mapped to COSO's eight use types, ready to hand to internal audit or an external auditor?
Does it track model drift and performance over time, with alerts routed to a named control owner rather than a general dashboard?
Can a reviewer see the AI's reasoning or supporting citations before approving an output, not just the final number?
Does the audit trail survive a model or prompt update, so a control tested in one period can be shown to be the same control in the next?
A vendor that can answer all six with a live demonstration, not a roadmap slide, has already built the evidence layer COSO’s guidance expects.
Practical takeaway.
COSO’s guidance does not ask a finance team to slow down its use of GenAI. It gives that team a way to show, capability by capability, that the AI running inside the close is covered by the same control discipline as everything else on the books. A team that completes the inventory and design steps now walks into next year’s audit with the mapping already done, instead of building it under deadline pressure when an auditor asks the first question about the model that touched a journal entry.
Sources.
- COSO: Achieving Effective Internal Control Over Generative AI (official guidance page)
- COSO via PR Newswire: COSO Releases Practical Roadmap for Managing Generative AI Risks and Controls
- Journal of Accountancy: COSO creates audit-ready guidance for governing generative AI
- Deloitte Heads Up: COSO Releases Publication on Internal Controls Related to Generative AI
- KPMG: COSO releases roadmap on internal control over generative AI
- Doeren Mayhew: COSO Releases Roadmap for Governing Generative AI
- EY: COSO 2026 and the governance shift in AI-enabled decision-making
Targeted searches for public X and Twitter commentary on COSO’s February 2026 GenAI guidance returned professional-firm articles and one LinkedIn post from the Institute of Internal Auditors’ president, rather than a credible set of two to four high-signal posts on X or Twitter from a standard setter, regulator, or independent practitioner. Ordinary citations are used instead, and nothing is fabricated.