Finance data quality emerges as a barrier to AI in payments

Payments companies are accumulating large volumes of transaction data, but Fynapse executive Ben Catterall said the more pressing issue is whether finance teams can use that information with confidence. In a PYMNTS Summer School discussion, Catterall argued that weak financial data architecture can limit revenue analysis, accounting control and the effectiveness of artificial intelligence projects.

Catterall, global head of solutions engineering at finance ERP provider Fynapse, said payments firms may not gain an advantage by collecting more data alone. The stronger position, he said, comes from preserving the context attached to each financial event, including what happened in the underlying transaction and how it should be reflected in the books.

“Every payment company has tons of data,” Catterall said. “But data without context is not particularly useful.” He added that finance teams need to understand both what the data represents and the commercial activity that produced it.

The challenge has grown as payment channels have expanded. Cards, wallets, buy now, pay later products, subscriptions and app store purchases can each produce different records. Cross-border transactions add currencies, settlement timing, foreign exchange effects and multiple payment gateways. Finance teams still have to reconcile those records and post them to the general ledger accurately.

Legacy finance systems can make that harder, Catterall said, because many were designed for batch processing and summary reporting rather than continuous, transaction-level accounting. In that model, transactions may be grouped and reviewed after the fact, which can hide the drivers of margin leakage or cost increases.

Catterall cited a multinational payments client operating across nearly 18 countries. At an aggregated level, the company’s books appeared to balance. A review of individual transaction records showed foreign exchange spreads averaging about 2%, which Catterall said translated into potential lost revenue of about $2 million a year on roughly $100 million of cross-border payments volume.

For finance leaders, that example shows how detailed records can affect more than reporting speed. Transaction-level visibility can help identify where revenue is lost, where payment fees build up and where operational changes may improve financial performance.

AI depends on the data beneath it

Catterall said the same data problem is slowing artificial intelligence work in financial services. Many institutions have tested AI tools over the past two years, but fewer have moved those projects into live production, he said.

The limitation often sits below the model itself. If an AI system is trained or queried against batched and aggregated accounting data, Catterall said, it can only infer from that narrowed view. It lacks the detailed record needed to explain why a balance changed, how a fee was applied or where an exception occurred.

Catterall said research indicates that 46% of AI proofs of acceptance do not reach production because poor data quality constrains their usefulness. He argued that finance data should be treated as core infrastructure rather than an operational byproduct of payments processing.

Fynapse’s approach, according to Catterall, is to capture financial events as they happen rather than rebuild the record at period end. The aim is to give finance teams continuous visibility into margins, payment costs and business performance while transactions are taking place.

He said richer data can be used across treasury, pricing, forecasting and risk functions, provided the records retain the transaction details behind the accounting entries. During the PYMNTS discussion, Catterall cited T-Mobile as a Fynapse user processing 200 million journal lines an hour in real time.

As payment methods proliferate and AI takes on a larger role in finance operations, Catterall’s central argument was that reliable automation depends on finance-grade data: individual transaction records, preserved commercial context and timely availability to the teams making decisions.

This story draws on original reporting from PYMNTS.