AI in the financial close

The promise that AI compresses the financial close from five days to one is, in part, real and in part oversold. The honest answer is that some elements of the close compress materially, others compress modestly, and some elements do not compress at all — because they should not. The CFO who deploys AI in the close without distinguishing the categories will produce either disappointment or a finding.

What the close is actually for

The financial close performs three functions that have to be preserved across any technology change.

• It produces a defensible statement of the entity's financial position at a defined point in time.
• It enforces the controls that make the position reliable — cut-off, reconciliation, journal review, intercompany elimination, foreign exchange translation.
• It produces the analytical content — variance explanation, qualitative narrative — that allows the position to be understood by users of the accounts.

Each function has different AI implications. Some are heavily addressable. Some are not.

What compresses, and why

The activities in the close that compress most materially with AI deployment are the activities that consist of pattern-matching at scale on structured data.

• Reconciliation. Matching transactions between source systems and the general ledger, identifying unmatched items, flagging exceptions. This is what AI does well. A reconciliation that previously consumed two days of analyst time can, with appropriate deployment, run continuously and present only exceptions for human review.
• Journal entry preparation. The routine, rule-based journals — accruals, prepayments, depreciation — are well suited to automated preparation. The discipline shifts to reviewing the rules and the exceptions, not preparing each entry.
• Intercompany matching. The matching of intercompany balances across entities, currencies, and timing differences is structurally similar to reconciliation and benefits similarly.
• Document extraction. Extracting structured data from invoices, contracts, statements, and other unstructured sources is now reliably automated, reducing the manual coding work that previously consumed analyst time at month-end.
• Variance flagging. Identifying movements outside normal ranges, which previously required a human to read every line of the management pack, can be performed automatically with the human focusing on the explanations.

The cumulative effect of these compressions is meaningful. A close that previously took five days, of which three were assembly and reconciliation, can plausibly compress to two days where assembly is largely automated and the human work concentrates on review and narrative.

What does not compress, and why that is fine

Several elements of the close do not compress materially with AI deployment, and the CFO should be honest about why.

Judgement. The assessment of whether a flagged variance is a control failure, a timing issue, or a real business event requires context that the model does not have. The judgement remains human and takes the time it takes.

Estimation and provisioning. The assessment of credit losses, warranty provisions, contingent liabilities, and other estimates is a judgement-heavy area where AI can support the analysis but cannot replace the assessment. The work compresses modestly at most.

Statutory cut-off decisions. The determination of which transactions belong in the period — particularly for complex revenue, multi-element arrangements, or in-flight transactions — requires interpretation of contracts and arrangements that the model cannot fully perform.

Audit committee narrative. The qualitative story of the period — what happened, what it means, what the business is doing about it — is a writing task that depends on judgement, context, and accountability. AI can draft. The CFO authors.

Control sign-off. The certifications that the close has been performed in accordance with the entity's control framework are human accountabilities. They do not compress because they are not about time; they are about ownership.

The control implications

An AI-compressed close changes the control environment in identifiable ways.

The first change is that controls move upstream. Where reconciliation was a control performed at period-end, automated reconciliation that runs continuously is a control performed throughout the period. Exceptions surface in real time, not at the close.

The second change is that the control over the AI itself becomes a primary control. The accuracy of the automated reconciliation depends on the accuracy of the underlying model, the data mappings, and the rules. The control framework must include model validation, monitoring of error rates, and a process for identifying and correcting model drift.

The third change is that documentation requirements increase, not decrease. Each AI-generated output that enters a financial statement must be traceable to its inputs, its model version, and its review. The audit trail expands even as the time spent generating each entry compresses.

The CFO who deploys AI in the close without strengthening the model validation and the documentation discipline will produce a close that is faster but less defensible. That is the worst outcome.

A practical sequencing for the first year

For a CFO planning to deploy AI in the close over the next twelve months, a defensible sequence has the following stages.

• Quarter one. Document the existing close. The activities, the time spent, the people involved, the controls. Without this baseline, the AI deployment cannot be measured.
• Quarter two. Deploy AI in the lowest-judgement, highest-volume activity. Usually this is reconciliation or document extraction. Run it in parallel with the manual process; do not switch over until the model performance is documented over multiple periods.
• Quarter three. Expand into journal entry preparation and intercompany matching using the validation framework established in the first deployment. Build the exception review process before scaling the automation.
• Quarter four. Move into variance analytics, with the discipline that AI surfaces candidates for human investigation, not conclusions for human acceptance.

By the end of the first year, the close should be measurably shorter, the controls should be demonstrably stronger, and the team should be doing work that uses more of the capability the team actually has. The CFO who arrives at year-end with the same close in the same time and a new AI invoice has executed badly.

This piece sits inside the CFO in AI framework. See also agentic FP&A and the continuous forecast and real-time anomaly detection in finance. Lorna writes from practice at IMPT. The verified page records what is and isn't published here.