AI in the Close and Consolidation: A Workflow Playbook

Four kinds of AI in the close

“AI” covers tools with different risks and different controls. The Center of Eminence uses four categories. They describe roles, not products, and they overlap: one platform can contain all four.

• AI built into finance platforms. Features delivered inside the ERP, planning, close, consolidation and reconciliation software the organization already licenses. The vendor builds them and changes them through its releases. The organization decides which ones to enable and how to configure them, within the options the vendor offers.
• General-purpose AI assistants. Chat assistants and productivity copilots that individuals use alongside finance systems to draft, summarize, analyze or build files.
• AI agents and automated actions. AI configured to take actions in other systems, such as posting entries, approving, changing master data or initiating payments. These may be built in house, bought or embedded in a platform.
• Established machine learning and rules-based automation. Matching engines, anomaly scoring, forecasting models and document capture. Much of this predates generative AI, and organizations often already test it as automated controls.

Two questions matter more than whether something is labeled AI:

• Does it suggest or act? A tool that drafts, recommends or flags leaves a person between the output and the books. A tool that posts, approves, changes or pays does not, unless the system enforces one.
• Who controls it? The vendor, finance, IT or an individual user may each be able to change its configuration, prompts, data connections or behavior. Whoever can change it can change the result finance relies on.

A third distinction matters for evidence. A deterministic rule gives the same output for the same input, so it can be tested and then monitored for change. Generative output can differ for the same request, so each result that is relied upon needs its own support.

Examples from vendor documentation

The examples below show selected vendors’ documented approaches to ERP, reconciliation and planning. They are vendor descriptions, not a complete market survey or comparative product assessment.

• ERP. As described in Microsoft’s Dynamics 365 Finance documentation, updated September 2026, the Account Reconciliation Agent, labeled a production ready preview, evaluates exceptions raised in the Account reconciliation workspace and recommends an action for each; for a voucher amount mismatch it recommends creating a journal entry, which the user can accept or replace with another action. The same page says the agent does not run until it is switched on in feature management and its modules and exception limits are set and confirmed; that Microsoft might automatically update the AI model, so the model in use might differ from the one the interface shows; and that each exception’s action history is logged and the action that reconciled a transaction can be undone, returning the exception to an unmitigated state. Microsoft’s 2026 release wave 1 plan describes a goal of moving the agent from “guided automation” toward a “fully autonomous reconciliation orchestrator,” and notes that planned functionality may not be released.
• Close and reconciliation. As described in Oracle’s Account Reconciliation documentation, Transaction Matching Assistance uses machine learning trained on the organization’s historical manual matches to predict matches for unmatched transactions; users can review a predicted match and confirm or discard it. The same documentation says the Enable Predictive AI setting must be enabled and that predictions are limited to one-to-one matches. Oracle’s April 2026 What’s New adds that each predicted match carries a confidence score.
• Planning. As described in Anaplan’s Anapedia documentation, Forecaster is a machine learning-based time series forecasting tool, and in finance a forecast item might be a budget line item such as a spend category or cost center. The same page says Forecaster comes in two tiers, obtained through the customer’s Customer Success Business Partner.

How to assess AI at each decision point

For each AI-assisted activity, whether newly proposed or already in place, ask five questions:

1. Evidence: Can the output be traced to reliable source records, and can the completeness and accuracy of its inputs be tested?
2. Judgment: Is the task primarily mechanical, or does it call for accounting interpretation, estimates or significant professional judgment?
3. Authority: Does it suggest or act? Can it recommend, draft, post, approve or change records, and are those permissions technically constrained?
4. Control: Who can change its configuration, prompts, data connections or model, and will finance know when the vendor, IT or a user changes it?
5. Review: Is there a specific test that could detect a material error or omission before the output is relied upon?

These are CoE | AI Cost & Governance evaluation criteria, not a classification set by any regulator. The appropriate control level depends on materiality, complexity, system architecture and the organization’s financial reporting obligations.

Work through the close, one decision at a time

These starting postures are CoE | AI Cost & Governance analysis. Adjust them for materiality, complexity, the system architecture and the organization’s reporting obligations. The COSO-commissioned publication calls for human corroboration proportionate to risk and discusses separating assistance from authoritative decisions in areas of significant impact.

Source context: The COSO-commissioned publication’s illustrative transaction-processing controls include role-defined posting authority, separation of configuration from review, locked configurations and full audit trails for unattended postings. PCAOB AS 2401 notes that consolidating adjustments and reclassifications generally are not reflected in formal journal entries and might not be subject to the entity’s internal controls.

A minimum control package for any material AI-assisted close task

Before relying on the output, retain an evidence record that answers:

• Purpose and owner: What task does the AI perform, and who owns the process and final decision?
• Kind and control: Which kind of AI is involved, does it suggest or act, who can change it, and was it enabled by vendor default or by choice?
• Data and completeness: What were the source systems, extract dates, reporting populations, completeness tests and control totals?
• Method and configuration: What prompts, rules, retrieval sources, model or workflow versions materially influenced the result?
• Permissions: What can the tool read, change, propose, post or approve? Where are limits enforced?
• Independent checks: Which calculations, reconciliations, exceptions, estimates and claims were independently tested?
• Release decision: Who reviewed the result, what defects were found, how were they resolved, and what version was approved?
• Monitoring: What vendor releases, configuration changes, drift, new data conditions or recurring exceptions trigger revalidation?

Several of these items draw on the COSO-commissioned publication, which treats prompts, retrieval connectors and transformation rules as governed configurations with version history and approval workflows, calls for a named owner for each GenAI tool, and suggests that, because GenAI output is probabilistic, multi-metric tolerances may need to replace a single pass or fail test.

For AI built into a licensed platform, some of this evidence sits with the vendor. Ask the vendor how it announces changes to AI features and models, and whether its service auditor’s report covers the AI features you rely on.

These controls can live in existing close workpapers and GRC systems. They do not call for a separate new system merely because an AI tool is involved.

A practical rollout sequence

What existing standards and guidance say

PCAOB AS 2201 describes the period-end financial reporting process and the auditor’s evaluation of its inputs, outputs, IT involvement, journal entries, adjustments and oversight. AS 2401 addresses the risk of fraudulent journal entries and other adjustments, including consolidating adjustments and reclassifications. These are auditing standards, not AI deployment prescriptions for management. They offer a defensible way to locate reporting risks.

Achieving Effective Internal Control Over Generative AI, a publication commissioned by COSO, written by academic and practitioner authors, and released on February 23, 2026, provides practical guidance on adapting the five components of COSO’s 2013 internal control framework (control environment, risk assessment, control activities, information and communication, and monitoring activities) to GenAI. This playbook’s recommendations are CoE | AI Cost & Governance analysis informed by that guidance, not rules issued by COSO or the PCAOB; the publication itself states that it contains general information only.

For SEC registrants subject to 17 CFR 240.13a-14, principal executive and financial officers must sign required annual and quarterly certifications themselves; the rule does not permit signing through a power of attorney or other confirming authority. Under Section 302 of the Sarbanes-Oxley Act, those certifications include the officers’ statement that they have reviewed the report. AI can assist their preparation but does not replace this officer obligation. These requirements are not universal for all finance organizations.

For the same companies, Rule 13a-15(d) requires management to evaluate any change in internal control over financial reporting during each fiscal quarter (each fiscal year, for a foreign private issuer) that has materially affected, or is reasonably likely to materially affect, that control. Whether enabling, reconfiguring or receiving a vendor update to an AI feature in the close is such a change is a judgment for management and its auditors.

The UK audit regulator’s March 2026 guidance lists inconsistency among the kinds of error AI output can contain, alongside hallucinations, omissions, distortions and faulty reasoning. PCAOB staff lists guidance on what it now calls companies’ and auditors’ use of “super intelligence” among the topics it is considering; none has been issued. Whether AI shortens the close or improves reconciliation quality needs measurement in actual deployments; this playbook does not claim it.

Related research

• What We Mean by AI: the four kinds of AI and the questions that matter more than the label
• Treat AI Output Like an Unsupported Journal Entry: evidence, support and completeness
• Review Is a Control Only When It Can Catch the Error: designing review to catch defects
• Your AI Agent Is an Unlisted Signer: delegated authority and access controls
• Internal Control Over Generative AI, Applied to the Close: control environment and governance
• A Chart Is a Claim: financial presentation and disclosure

Expand reliance when the evidence earns it

For many organizations, AI already reaches the close through the software they license. The useful work is to know what kind of AI each step involves, whether it suggests or acts and who can change it, then to decide what to enable, rely on and evidence. The aim is a close where routine work can be accelerated without losing the evidence, authority and accountability on which reliable reporting depends.