A food and beverage manufacturer whose CFO is weighing AI for invoice processing should first remove duplicate and inactive suppliers from its vendor master and fix the payment terms on the rest, because the AI matches every invoice against that master.
Gartner published on November 18, 2025 a survey of 183 CFOs and senior finance leaders. Among the respondents that had implemented AI in their function, accounts payable process automation was the second most adopted use, at 37%, and the release names inadequate data quality and availability, with data literacy and technical skills, as the largest obstacles to AI adoption across all organizations. KPMG, whose member firms sell audit, tax and advisory services, published on May 11, 2026 a survey of 1,013 senior finance leaders at organizations with annual revenues of at least 250 million US dollars (500 million in the United States) in 20 countries, in March 2026: 36 percent of organizations cite data quality as both their top barrier and opportunity. Technology and financial services are 58 percent of that sample; the release names seven of its 13 sectors; Gartner gives no sector breakdown.
Karat Packaging, a specialty distributor and select manufacturer of disposable foodservice products, not a food or beverage manufacturer, reported in its 10-K for 2022, filed March 16, 2023, a material weakness in which management did not design and maintain effective activity level controls over the procurement process, including controls over vendor master file changes, the approval and matching of purchase orders, invoices and receiving documents, and authorization of vendor payments. Its 10-K for 2023, filed March 15, 2024, reports those process weaknesses fully remediated as of December 31, 2023, with a company-wide approval matrix for cash disbursements, while still reporting material weaknesses in entity-level and IT general controls. Invoice AI works inside that same control set: it reads the supplier, the remit-to address and the terms from the master to match an invoice to its purchase order and receipt, so a supplier set up twice gives it two records to choose between.
The cleanup itself can use AI. A paper in the proceedings of the International Conference on Extending Database Technology, March 2025, tested large language models on entity matching, deciding whether two records describe one real-world entity, and found that the best models require no or only a few training examples to perform comparably to earlier models fine-tuned on thousands, and that GPT4 can generate structured explanations for matching decisions.
At a two-plant food manufacturer, the accounts payable manager exports the vendor master with, for each record, the legal name, the remit-to address, the tax identification number, the payment terms and the date of the last invoice. The matching model groups records that read as one supplier, an ingredient broker set up under its legal name and again under its trade name with a second remit-to address. The accounts payable manager decides each group, purchasing confirms the terms against the contract, and the controller approves every merge, inactivation and change of terms through the same approval matrix that governs payments.
In the first month the export, limited to records with an invoice in the last 24 months, and the matching run produce the list of proposed groups and record the current count of failed three-way matches for one supplier group, packaging. In the second, every group is decided, the terms are confirmed with purchasing and the changes are posted. In the third, invoice AI runs on that group, and the count goes to the CFO at month end.
KPMG's survey also found that only 29 percent of organizations track where AI adoption fails.
