CFOs want AI returns—but their data-agent buying test starts with control

Finance has acquired a second AI job

Finance leaders are being asked both to use AI and to police how the rest of the company uses it. In a Salesforce-commissioned survey of 865 senior finance leaders, nearly three-quarters said their role had expanded in the past year, with managing AI use and expansion the most-cited reason. Half said they are now primary decision makers in their company’s AI strategy.

That shift matters for enterprise data agents. A natural-language interface may be demonstrated to analysts, but finance is increasingly involved in deciding whether the system’s access, controls and claimed returns are acceptable.

The demand signal is real—and self-reported

The survey describes a finance function under operational pressure. Sixty-five percent of respondents said their companies manage more than one revenue model, while 71% said they sell through more channels than a year earlier. Sixty-seven percent said their teams still complete at least one in five workflows manually, often through spreadsheets, and respondents estimated that as much as 40% of their work could be automated.

Among respondents already using AI, 90% reported positive return on investment. More than 90% of those using AI agents reported benefits in time savings, productivity, cost savings and forecast accuracy, according to Salesforce.

Those figures are useful as a demand signal, not as an independent performance benchmark. The results are self-reported, Salesforce commissioned the research, and the published summary does not disclose product-level costs or measured before-and-after outcomes. The double-blind survey was fielded May 4–15, 2026, across France, Germany, Japan, the United Kingdom and the United States.

Control is part of the product

The strongest buying signal may be in the blockers. Forty-six percent named security as a top obstacle to expanding AI use. Governance concerns and integration with existing systems were each cited by 42%.

For text-to-SQL and other conversational-data products, the practical implication is that answer quality cannot be the only acceptance test. A finance buyer also needs to know which identity authorized a query, what data the agent could access, how an answer can be traced back to source systems, and where a human approval remains in the workflow. Those requirements are an inference from the survey’s security, governance and integration findings; the survey did not specifically test text-to-SQL products.

The near-term opportunity is therefore less “replace the spreadsheet with chat” than connect governed analysis to a revenue process that still crosses contracts, usage records, billing systems and recognition rules. Vendors that can show that control path clearly have a stronger finance case than those presenting fluent answers without an auditable operating model.