Finance is simultaneously the function with the strongest case for AI assistance and the one where a mistake is least forgivable. The work is frequent, deadline-bound and verifiable — the three properties that make AI assistance safe — but every number carries consequence, which means the workflow has to be designed with unusual care.

This guide covers where AI helps in a finance function, where it must not be used, the controls that make the difference, and how to measure the result. It is part of the Use Cases & Workflows pillar.

In finance, AI drafts the words. The numbers come from the system of record, and a person signs them.

The burden finance carries

The case begins with the workload. In EOG’s synthesis of 2025–2026 research:

  • 51% of the finance week went to manual work such as reconciliation and report stitching (Intuit, May 2026, survey of 2,000 finance leaders).
  • 36% of senior finance staff spend 31–50% of their time on manual data work, and 40% report a close of seven days or more (Financial Education & Research Foundation, September 2026).
  • 57% said time-sensitive action had been missed because of delays in getting data, and 70% reported no single source of truth for critical data (Intuit, May 2026).

Those are the hours and the risks that a designed workflow attacks. None of them is a reason to let a model near the arithmetic.

Where AI helps

Five workflows, all on the language side of the ledger.

1. Variance commentary. Given frozen figures and the prior period, AI drafts a first-pass explanation of what moved. The finance lead verifies the cause, because the cause is the part that requires judgement.

2. Report narrative. The opening narrative of a management or board report is drafting work against a known structure and frozen numbers.

3. Document summarization. Long contracts, policies, supplier terms and correspondence summarized to a structure the finance team defines.

4. Procedure and policy drafting. Standard documents drafted from a brief and an existing template, then reviewed and owned by the finance lead.

5. First-pass analysis framing. Structuring an analysis — what to compare, what to test — rather than producing the analysis itself.

Every one of these keeps the numbers and the judgement with a person.

Where AI must not be used

Four boundaries, and they are not negotiable.

  • Calculations and totals. Never let a model perform or restate arithmetic that will be relied upon. Totals come from the system of record.
  • Journal entries and adjustments. The accounting treatment is a professional judgement with an audit trail.
  • Regulatory or tax positions. Fluency is not authority; these require a qualified human.
  • Confidential financials in unapproved tools. The data rule applies most strictly here. Review the tool’s retention and training position before anything sensitive is entered. See Data Security and Confidentiality in AI Tools.

The boundary is consistent: AI works on language, and the ledger stays untouched.

The workflow

Six steps, with the check placed where the risk is.

1. Reconcile and freeze. Close and freeze the data before any drafting begins. The figures the document will use are fixed. 2. Supply the source. The finance lead provides the frozen figures and the prior period to the drafting step. 3. Draft. AI produces commentary and narrative against those figures, in the standard structure. 4. Verify. A named person checks every figure against the source, every variance explanation against its cause, and every claim in the narrative. 5. Approve. The finance lead signs off, with the version and snapshot recorded. 6. Record. The check, the version and the snapshot are retained, so the report is reproducible.

The freeze is what makes the rest possible: drafting against a moving set of numbers guarantees that the verification step never converges.

The controls that decide the outcome

Three controls separate a finance workflow that works from one that creates risk.

  • One source per metric. A metric dictionary recording the definition, the source and the owner. Where two versions of a figure are in circulation, the drafting workflow surfaces the discrepancy rather than resolving it.
  • A recorded verification. The check must be evidenced, not assumed. This is what makes the workflow defensible in an audit.
  • Separation of preparer and reviewer for anything that leaves the function, even where the same person would otherwise do both.

See Building a KPI & Metric Dictionary and the Human-Verification Checklist.

Measuring it

Four measures, reported monthly.

  • Close and cycle time, against the baseline, including verification.
  • Corrections per report, and the stage at which they are found.
  • Variance explanations challenged, which should fall over time as the drafting improves.
  • Verification recorded, as a proportion of issued reports.

The measure that matters in the first quarter is whether the workflow reduces the senior time spent on assembly. Junior time spent verifying is a reasonable trade; senior time spent reformatting is not.

A worked month-end

A finance team of three produces a monthly management report with a narrative, a variance section and a cash summary. The cycle was consuming eleven hours, of which four were drafting.

  • Close (unchanged). The team closes the books and freezes the figures at day three. This step is not a candidate for AI, and it is where the schedule risk sits.
  • Draft (45 minutes, was four hours). The finance lead supplies the frozen figures, the prior period and the standard structure. AI drafts the narrative and the first-pass variance commentary.
  • Verify (90 minutes). A second person checks every figure against the system of record, confirms each variance explanation against its cause, and reads the narrative for claims that are not supported.
  • Approve and record (20 minutes). The finance lead signs off, and the version and snapshot are recorded.

Net change: around six hours returned per month, with the verification step deliberately longer than it was before. The reconciliation, the judgement and the sign-off are untouched — which is precisely why the workflow is defensible.

Common mistakes

  • Drafting before the freeze. The verification step never converges against moving figures.
  • Letting AI produce a total. The most serious error available in this workflow.
  • Confidential data in an unapproved tool. The data rule exists for this case specifically.
  • No metric dictionary. Two versions of a figure produce two versions of the commentary.
  • Verification not recorded. The control exists in policy and not in evidence.
  • Drafting the conclusion. The narrative can be drafted; the judgement is the finance lead’s.

Frequently asked questions

Can AI be used in finance?

Yes, on the language side: variance commentary, report narrative, document summarization, procedure drafting and analysis framing — with figures taken from the system of record and verified by a named person.

Should AI calculate financial figures?

No. Calculations and totals must come from the system of record. A model can produce a set of numbers that is internally consistent and wrong, and the error is hard to detect in a well-written report.

Is it safe to put financial data into an AI tool?

Only into a tool that has been assessed for retention, training use and access. Confidential financials belong in approved tools, and the data rule should name the categories that are prohibited. See Data Security and Confidentiality in AI Tools.

How do we verify AI-drafted financial commentary?

Check every figure against the frozen source, every variance explanation against the underlying cause, and every claim in the narrative — then record the check and the version.

How much time can AI save in finance reporting?

The honest answer comes from your baseline. The documented burden is substantial: 51% of the finance week on manual work (Intuit, May 2026, n=2,000) and 36% of senior staff spending 31–50% of their time on manual data work (FERF, September 2026).

Does AI help with reconciliation?

It helps summarize and explain reconciliation differences, and it can draft the workings narrative. It should not perform the reconciliation, and the reconciliation itself remains in the finance system.

Will this satisfy an auditor?

Where the workflow is documented, the source is named, the verification is recorded and the version is retained, an AI-assisted drafting step is defensible. Confirm your specific requirements with your auditor or advisor.

What about AI features inside our accounting system?

They can be useful for anomaly flags and summarization, and they carry the advantage of staying inside a system your data already lives in. The same boundary applies: treat flags and drafts as inputs to a person’s judgement, never as the judgement itself.

Next step

Start with one report: freeze the data, draft the commentary with AI, verify against source, and record the check. See Document & Content Production with AI for the workflow detail, or book an AI adoption call to design it with your finance lead.

Sources

  • Intuit Enterprise Suite, Future of Finance 2026 Report (survey of 2,000 CFOs, controllers and VPs of Finance at US businesses over $2.5M revenue, May 2026): 51% of the finance week spent on manual work such as reconciliation and report stitching; 57% missed time-sensitive action because of data delays; 70% report no single source of truth for critical data.
  • Financial Education & Research Foundation / Financial Executives International (September 2026): 36% of senior finance staff spend 31–50% of time on manual data work; 40% report a close of seven days or more.

Figures are cited from their sources and dated. Where a source is a vendor survey, the sample size is stated where published. This article is general information, not accounting, tax or audit advice.