AI in Finance Has Entered Production. What Should CFOs Do Next?

Artificial intelligence (AI)The use of computer systems to simulate human intelligence processes like learning, reasoning, and decision-making—often applied in AP automation for data capture, fraud detection, and workflow optimization. B is already changing finance work. The more useful question for CFOs is whether it is changing the right work and whether the organization can explain, measure, and control the result.

That is the message I take from a new survey Edenred PayEdenred Pay is the market leader in B2B payments automation. conducted among accounts payable (AP)The amount a company owes to suppliers for goods and services received but not yet paid. and finance leaders. Respondents report meaningful productivity improvements and plan to invest further, particularly in fraud prevention and payment automation. But their deployments remain limited, and they want people accountable for the decisions that carry financial consequences.

Here are five takeaways I believe deserve a place on every CFO’s AI agenda.

1. Moving from pilot to production is a management challenge

Forty-three percent of respondents are running AI pilots, and another 43 percent have limited production deployments. Broadly embedded use remains uncommon. That is progress, but a promising demonstration is different from a dependable finance process.

In AP, an AI-generated recommendation may rely on an invoice, purchase order (PO)A formal document issued by a buyer authorizing the purchase of goods or services under agreed terms. R, supplier record, approval rule, and payment instruction held in different systems. A tool may perform well on a sample of clean invoices, then encounter missing fields, conflicting supplier names, or an unusual approval path in production. The CFO’s question should be: What happens to those exceptions, who owns them, and can we reconstruct the decision later?

Scale the workflow only after the answers are clear. Define data ownership, integration points, escalation paths, access permissions, and evidence requirements alongside the technology itself.

2. Productivity gains are real, but the business case needs a baseline

Improving employee productivity is the leading primary objective for finance AI, cited by 41 percent of respondents. Seventy-two percent report productivity improvements of 11 percent to 25 percent. Those are encouraging results. They do not, by themselves, tell a CFO whether an investment has reduced total cost or improved working capital.

An invoice may take less time to capture while still waiting for days for approval. An alert may identify a potential duplicate while creating more review work than it saves. To understand the net effect, measure the whole process before and after deployment: manual touch time, exception volume, invoice cycle time, payment timing, error rates, and staff capacity. Include the effort required to investigate AI-generated flags and maintain controls.

This discipline matters because 58 percent of respondents cite difficulty proving return on investment as a barrier. A clear baseline turns an impressive demonstration into a defensible investment decision.

3. Put AI where cash and control meet

Fraud prevention and payment automation lead respondents’ planned AI investments. When asked which capabilities could create the most value, 61 percent selected fraud detection, 50 percent payment optimization, 48 percent predictive forecasting, and 41 percent automated exception handling.

These priorities connect more closely than they may appear. An unusual supplier-bank change can affect payment risk. A delayed invoice approval can alter a cash forecast. A payment method recommendation is useful only if it respects supplier preferences, policy, and available cash. CFOs should look across invoice receipt, supplier records, approvals, payment execution, and reconciliationThe process of matching financial records—such as payments and invoices—to ensure accuracy in accounting and reporting. when selecting use cases.

The opportunity is to bring relevant information to the person making a decision at the right moment. For example, AI could identify an invoice that looks similar to one already paid, assemble the supporting records, and route it for review before payment. The value lies in the controlled action that follows the detection.

4. Keep authority with accountable people

Finance leaders are willing to use AI recommendations, but they expect verification. Seventy-six percent say they are somewhat confident in those recommendations and verify important decisions. Another 23 percent verify most recommendations. Seventy-five percent believe fraud investigations should always require human approval, while 62 percent say so for invoice approvals and 55 percent for payment approvals.

That should shape the business rules for an AP system. AI can extract information, compare records, flag anomalies, recommend next steps, and prepare an exception for review. Authorized employees should retain authority over the consequential decisions specified by company policy. Rules should make clear which transactions can follow a standard path and which require additional scrutiny.

An audit trail records what information was available, what the AI recommended, what a reviewer decided, and who approved the final action. Oversight becomes practical when the evidence travels with the transaction instead of requiring a later reconstruction.

5. Repair the foundation before asking AI to do more

The biggest obstacles respondents identified are data quality and availability (65 percent), difficulty proving ROI (58 percent), security and compliance (55 percent), and integration with existing systems (53 percent). A more capable model cannot reconcile inconsistent vendor records or clarify an undocumented approval policy on its own.

Start with one process. Map its data sources, handoffs, exceptions, decision rights, and current performance. Clean the records that matter, connect the relevant systems, and specify how employees will review and correct outputs. Then choose a use case with a measurable outcome and expand as the evidence supports it.

The greatest potential is when intelligence is connected to the invoice-to-pay workflow: from receiving an invoice through approvals, supplier payments, reconciliationThe process of matching financial records—such as payments and invoices—to ensure accuracy in accounting and reporting., and reporting. That connection gives finance teams the context to act on insights and the controls to do so responsibly.

The survey points to a practical path for CFOs. Invest in AI where it solves a specific finance problem. Measure the entire workflow. Preserve human accountability for consequential decisions. And treat trustworthy data and connected processes as part of the investment, not as cleanup work to be addressed later.  That is how AI moves from an interesting pilot to a finance capability leaders can rely on.

Explore the full findings: