AI in Finance: From Pilot to Practice

Key Takeaways

  • Adoption has outpaced realized value: Gartner reports that 63% of surveyed finance organizations found 2025 implementation slower than expected.
  • The pilot-to-production gap is mostly about data, ownership, controls and evidence, not model capability.
  • No binding US standard governs AI in financial reporting as such. Existing ICFR, audit and professional-conduct expectations apply, and COSO issued dedicated guidance in February 2026.
  • Human review operates as a control only when it is specified, performed and evidenced.
  • Accountability for AI-assisted judgment remains with named individuals in the finance function.

In a March 2026 Gartner survey of 204 finance leaders, 63% said AI implementation in 2025 ran slower than expected. A separate Gartner survey from the same period found that 45% of finance AI investment leans toward productivity and 20% toward decision quality. Adoption is not the constraint. Conversion is.

Most large organizations have now run pilots in the finance function: a reconciliation assistant, a drafting tool for disclosure text, a model that flags payroll anomalies. Far fewer have placed such tools inside a controlled process with named owners, retained evidence and ongoing performance review. The difference is operational rather than technical, and it sits within the remit of the CFO, the controller and internal audit.

This article sets out where AI is delivering value, why pilots stall, what production practice requires, and where the governance and regulatory picture stands. It closes with a five-step framework and five questions for leadership.

Where AI Is Delivering Value in Finance Operations

Gartner’s 2026 survey of 160 senior finance leaders found that data extraction, accounts payable and receivable automation, and report creation generally deliver returns within nine to 10 months. These are bounded tasks with clear inputs and checkable outputs, which explains their early traction.

  • Accounting and Close: Document extraction, invoice coding and journal entry preparation are the most mature uses. The value lies in reduced manual handling and earlier visibility of exceptions, not in removing review.
  • Reconciliations: Matching and exception triage suit machine assistance because each match can be traced to source records. The control question is how auto-matched and unmatched items are reviewed.
  • Financial Reporting and Disclosure Drafting: Generative tools are used to prepare first drafts of variance commentary and disclosure text. Drafts need the same tie-out, review and approval as any other, and statements about a company’s own AI use must themselves be accurate.
  • Tax Data Preparation: AI is applied to extracting, classifying and reconciling the data behind provision and compliance work. The GAO reported that the IRS maintained 126 active AI use cases as of June 2025, a reason for preparers to expect analytical scrutiny of submitted data.
  • Payroll, Budgeting and Risk Monitoring: Anomaly detection in payroll, forecasting support and continuous monitoring of control exceptions are commonly described as active uses.

The mix matters. Gartner’s March 2026 survey found that 45% of finance AI investment leans toward productivity and 20% toward decision quality, which may explain why boards perceive limited strategic impact from progress that finance leaders report.

Why Pilots Stall

Pilots succeed under conditions that production does not offer: a curated dataset, an engaged sponsor and tolerance for manual workarounds. The same factors then reappear as obstacles.

Data and Systems: Gartner research identifies data literacy, technical skills and inadequate data quality or availability among the largest barriers to finance AI adoption, and its September 2026 release calls low AI literacy the most significant one. Fragmented charts of accounts, inconsistent master data and legacy ERP configurations raise the cost of every integration.

Ownership and Measurement: A pilot is often sponsored by a technology team or a single finance manager, with no one accountable for performance afterward. Without an outcome defined in advance, whether hours, cycle time, error rate or exception volume, the pilot has nothing to graduate against.

Reliability: Generative output varies, and a result acceptable on a pilot sample may behave differently across the full population.

Talent and Change Reviewers need to know what to challenge, and preparers need to see how their work changes. Gartner’s March 2026 CFO survey ranked building AI talent in the function among the most challenging near-term priorities.

Implications for the Finance Function

Finance roles shift toward specification, review and exception handling. Gartner advises that by 2030 leading finance organizations will be reshaped by AI and automation, with teams moving toward digitally skilled roles that build tools and manage AI-driven workflows. That is a projection, not an observed outcome.

Oversight skills matter as much as technical ones. A reviewer who cannot recognize a plausible but wrong output provides limited assurance, so training should cover the failure modes of the specific tools in use.

Accountability does not transfer to the tool. Where AI assists with judgment, such as an estimate, a reserve narrative or a tax position, the preparer and approver remain responsible for the conclusion, and the record should show what the tool contributed and what the person decided. The IRS OPR guidance takes a consistent view, treating AI output as a draft requiring human review.

Questions for Leadership

(Questions to go over)

  • Which measurable outcome, set before the pilot began, did it achieve, and against what baseline?
  • Which significant accounts or key controls does the output touch, and has the ICFR risk assessment been updated to reflect it?
  • Who is the named owner, and who has authority to suspend the tool?
  • What evidence would an external auditor need to test the process, and is it being retained?
  • What happens to the process if the tool is withdrawn or the vendor changes the underlying model?