The CFO of a distribution company should build the company's first finance AI agent as one named application, exception handling in accounts payable for the invoices the ERP's own match flags, owned by the accounts payable lead and paid from the finance budget, and should stop funding it if accounts payable's own numbers miss their written targets after three month-end closes.
Gartner, which describes itself as a business and technology insights company, published its 2025 Gartner AI in Finance Survey on November 18, 2025, from 183 chief financial officers (CFOs) and senior finance leaders asked in May and June 2025. Of those 183, 59% reported the use of AI in their finance function, and among respondents that had implemented AI, knowledge management was the most common use case (49%), followed by accounts payable process automation (37%). The release counts AI of every kind and mentions neither agents nor any industry.
PwC, which sells audit, tax and consulting services, separately surveyed 308 US business executives between April 22 and April 28, 2025, and published the results on May 16, 2025; the page's opening calls it a May 2025 survey of 300 senior executives. The page names no industry for them and reports no figure for finance executives. Respondents expressed the highest trust in AI agents for data analysis (38%), and trust fell for higher-stakes activities like financial transactions (20%). Payment release should therefore stay with the controller.
The first application should cover invoices that cite a purchase order; invoices with none should stay with the accounts payable specialists. Each morning the ERP's own match, within the controller's tolerances, should decide which new invoices are exceptions. The application should then read each exception against its purchase order and receiving record and, in an exception queue of its own, give each a reason, attach the documents and draft a note to the vendor, which a specialist sends, edits or discards once the owner below has confirmed the cause. Inside the ERP, IT should limit it to reading. A price variance or a freight charge missing from the purchase order goes to the buyer, a quantity variance or missing receipt to the receiving supervisor, and a duplicate to the accounts payable lead, who owns the queue.
The CFO should pay an outside builder a fixed fee from the finance budget for an application the company owns and carry the running cost, what the AI model charges for each invoice read, as a monthly line in the accounts payable cost center. Invoices carry vendor bank details, so the model account should sit in the company's name, under a contract that bars training on them.
Before go-live the accounts payable lead should pull from the ERP the invoices and exceptions per day, to size the specialists' hours, and three numbers for the three prior month-end closes and, as a seasonal baseline, for the months a year before the next three closes: the share of the month's exceptions cleared in the ERP within a number of working days the lead sets, the exceptions open on the last day of the month, and the median days from an exception's entry to its approval for payment. The CFO and the accounts payable lead should write a target for each against both baselines, larger than the spread across the closes pulled: a higher share, fewer exceptions, fewer days. The controller should keep the tolerances unchanged through those closes; if any of the three misses its target after them, the CFO should stop funding the application. An operating partner can ask any portfolio company running purchase orders through an ERP for the same three numbers.
Gartner's release adds that after a pilot launches it still takes time to realize significant gains, with 91% of respondents reporting low or moderate impact initially.
