Accounting & Finance AI 2026: close faster, prove the numbers

88%

Expected impact

In AICPA & CIMA's 2025 survey of 1,446 global finance and accounting leaders/managers, this share expected AI to be the most transformative technology trend over 12-24 months.

8%

Very well prepared

Share in the same AICPA & CIMA survey who felt their organization was very well prepared to manage the AI trend.

46%

Skills gap

Share identifying generative AI as the most significant skills gap for their teams in that survey. These are perceptions, not an audit of proficiency.

The close is a chain of evidence

A close is not a stack of narratives. Transactions enter subledgers; reconciliations tie them to the general ledger; controllers investigate differences; entries are approved; reporting packages are reviewed; the business explains what changed. AI is most useful where a person already has to compare records or summarize a supported variance. It is least useful when it turns an unsupported number into a polished explanation.

Example: A controller asks an assistant to draft a variance narrative only after it receives the approved prior-period and current-period ledger extracts plus a map of known one-time items. Each sentence links to a workpaper line.

Choose a reversible first workflow

Begin with a task that does not post to the ledger: classify incoming support requests, prepare reconciliation checklists, compare two policy versions, or draft variance questions. For journal entries, payment instructions, tax positions, and external reporting, require explicit authorization and independent checks. The system should propose a question or draft, not quietly change the books.

Example: An accounts-payable team flags possible duplicate invoices for review. A human compares vendor, date, amount, purchase order, and receipt before any hold or payment decision.

Keep source and period attached

Every figure needs an entity, account, period, currency, source system, extract time, and owner. The assistant may compare or explain, but it must not fill a missing field by guessing. A strong output says which data was excluded and what remains unresolved. Version records matter when the underlying close package changes after the draft is produced.

Example: A multi-entity finance team labels trial balances by company and currency. The tool refuses to combine them until conversion and consolidation rules are supplied.

Audit evidence and the reviewer

An auditor needs sufficient appropriate evidence, not just a plausible answer. The PCAOB's technology-assisted analysis amendments emphasize how electronic information is evaluated and used. The finance team should preserve source files, extraction rules, transformations, exceptions, and approvals so an auditor can reproduce the path. Reviewers should challenge outliers and omitted populations, not merely sign a generated summary.

Example: For a revenue sample, an auditor traces selected transactions from ledger to contract, invoice, delivery, and cash evidence. AI can organize the packet; it does not decide whether recognition criteria were met.

The economics of finance AI

Model the unit of work: reconciliations, invoice exceptions, variance packages, or audit requests per period. Subtract license, integration, review, correction, and controls costs. Released hours are capacity until overtime, contractors, hiring, or an external fee actually falls. Faster close can have strategic value, but avoid assigning a dollar figure without an observed decision or financing consequence.

Example: A team handles 400 monthly reconciliations. AI saves six preparation minutes but adds three review minutes: 20 net hours of capacity at that volume. That is not a 20-hour payroll reduction unless staffing or paid work changes.

A 90-day close pilot

Days 1-20: choose one entity and close task, audit data definitions, and record two recent baselines. Days 21-45: run on prior periods, then compare generated output to approved workpapers. Days 46-70: use reviewed drafts in a live close, with an exception log and a manual fallback. Days 71-90: decide whether accuracy, reviewer load, close timing, and auditability improved together. Do not expand because the prose reads well.

Example: The controller, process owner, internal audit, information security, and a staff accountant review ten randomly selected outputs each week.