Finance posted the fastest rise in AI agent use of any function this year. The same survey shows the agents only work on their own when the workflow around them is already connected, and a new bank tool for cash application shows what that connection looks like in practice.
Key Takeaways
Finance spent most of the last two years as the cautious function on AI agents, and for a reason: an error in a sales email is embarrassing, while an error in the general ledger is a restatement. That caution is fading fast. Agent use in finance departments grew faster this year than in any other part of the business. But the same data shows that buying agents and getting them to work without a person stepping in are two different achievements, and the difference has less to do with the agent than with whether the process it sits in is connected end to end.
Zapier's 2026 AI agents survey, its second annual poll of enterprise leaders, found the share reporting AI agent use in finance departments rose from 24% in 2025 to 41% in 2026. That 17-point gain was the largest of any function tracked, ahead of human resources (30% to 43%), operations (47% to 59%) and sales (26% to 37%). Across the business, 71% of organizations now have agents in production, up from 40% a year earlier, and 90% of leaders say they are certain or likely to increase agent investment over the next year.
The work agents are doing is not exotic. The most common uses in both years were document analysis and summarization, and data extraction and entry. In finance terms, that is reading remittances, invoices and statements and keying what they say into another system. The leading barriers are the ones a controller would expect: 27% of respondents cite data privacy, security and regulatory compliance risk, and 24% cite accuracy, reliability and errors.
Zapier's own CFO, Ryan Roccon, told CFO Dive that earlier models "weren't always accurate or reliable enough for finance," and that accounting makes a good starting point because many of its processes are rules-based and their outputs can be checked. His control is blunt: "Nothing hits the general ledger ever without a human clicking approved."
The more revealing number is how much agents actually do on their own. Only 30% of leaders say their agents complete more than 60% of assigned tasks without human intervention. With 71% running agents in production, most organizations have deployed something that still hands a large share of its work back to a person.
The survey isolates what separates the two groups, and it is not the model. When more than half of an organization's agents are part of broader automated workflows, 77% of leaders report that their agents complete more than 60% of tasks unaided. When half or fewer are connected, that falls to 8%. An agent dropped into a manual process inherits every gap in it: the missing reference, the file that lives in someone's inbox, the exception nobody has defined. Zapier's recommendation is to build human review checkpoints into the workflow itself, such as routing billing disputes to a person, rather than relying on someone to notice when an agent has gone wrong.
Accounts receivable is where that lesson is easiest to see. On November 2, J.P. Morgan Payments launched Receivables Matching, which performs three-way matching across incoming payments, remittance information and a client's open receivables data, and combines matching, exception management and reporting in one tool. Clients can adopt it without replacing their existing technology. Launch client Grimco, a distributor of sign, graphics and printing supplies, reduced manual review time by more than 60% and saw about 95% fewer errors requiring manual intervention.
The third leg of that match is the important part. Michelle Conklin, who runs the bank's receivables and public sector business, told CFO Dive that matching against the business's own AR file "was always the missing piece." Until now, clients matched only the payment and its remittance data. Clients upload an Excel file of expected payments, and the bank checks that payments arrive as expected and are applied correctly, without accessing the client's AR system. About 9,000 of the bank's clients use its receivables platform, and the service carries a fee the bank declined to disclose.
The pressure behind it is real. In Versapay's 2026 Cash Flow Clarity Report, a Wakefield Research survey of 400 U.S. and Canadian finance leaders, 69% said late customer payments rose over the past year and 74% said their teams spend significant or moderate time each week chasing them. 78% said unexpected AR issues are forcing changes to capital investment, hiring or borrowing plans. Half expect to use AI to automate repetitive AR tasks, and 82% plan to increase AR automation investment over the next 12 months. The Grimco result suggests where that money works: on the data an agent or a matching engine needs to make a decision, not just on the decision itself.

Guide
J.P. Morgan's matching tool works alongside existing systems because the ERP was never where cash application got solved. This guide explains why ERPs fall short as the sole foundation for AR and what finance needs on top of them to automate and cut DSO.
Download
Guide
Agents ran unaided far more often when they sat inside connected workflows. This guide covers the continuous accounting, real-time reconciliation and integrated reporting that give finance that connected foundation before agents are layered on.
Download
Report
71% of organizations have agents in production, but only 30% see them finish most work alone. This report looks at how technology, process design and organizational alignment move finance and operations teams from intention to execution.
Download
In a Redwood survey of 300 leaders, 82% say existing systems, not strategy, are the bigger obstacle to scaling AI.
Fully automated accounts payable rose from 7% to 31% in a year, but only 21% are very confident they would catch a fraudulent vendor change.
A survey of 270 CFOs finds teams spend 26% of the working week verifying or correcting AI output.