Treasury AI plans hinge on bank data connectivity, AccessPay says

Corporate treasury teams are being pushed to adopt artificial intelligence, but fragmented bank and payments data remain a constraint on reliable deployment, according to Karen Fagan, head of treasury consultancy service at AccessPay. Fagan said discussions at the recent Association of Corporate Treasurers conference in Liverpool showed that treasurers see potential productivity gains from AI, while remaining cautious about accuracy, governance and data exposure in regulated finance functions.

In an external expert opinion published by Finextra, Fagan said finance leaders are asking treasury teams for AI roadmaps, while treasury management system vendors are promoting tools such as AI-assisted cash-flow forecasting. She said that demand is meeting a practical obstacle: many treasury functions still rely on data held across multiple systems, bank portals and file formats.

Fagan cited examples from treasurers testing AI in limited settings. One treasurer using an AI-enabled cash-flow forecasting tool inside a treasury management system continued to run the established forecasting process at the same time because the newer tool was not yet considered dependable enough on its own. Another was assessing an AI bot that would allow business users to check payment status, while raising concerns about the risk of exposing sensitive information.

Data quality sets the boundary

Fagan argued that treasurers who have begun substantive AI projects tend to start with treasury data rather than the model itself. She described treasury data as often split across systems, manually collected, inconsistently formatted, duplicated and lacking clear ownership, reflecting the complexity of large corporate treasury operations with multiple back-office platforms and banking relationships.

The issue is operational as well as technical. Payment files and bank statements can arrive in different formats depending on the institution, system and process. If those inputs are incomplete or inconsistent, an AI model can deliver a result quickly while still being wrong. Fagan quoted one group treasurer as saying: “If ‘bad data’ is plugged into an AI model, treasury could get a very quick answer that is very wrong.”

She said treasurers are seeking a single reliable view of data and greater automation across processes, but that those goals depend on cleansing, centralising and standardising the underlying information. In her assessment, the starting point is often bank connectivity: automated links between banks and back-office finance systems, including treasury management systems and enterprise resource planning platforms.

How connectivity changes the data feed

Bank connectivity tools sit between corporate finance systems and banks. According to Fagan, that layer can centralise feeds from multiple banks, automate payment and bank statement data flows, and translate data into the formats required by different internal systems. That reduces manual movement of files between bank portals and corporate systems, a process she said remains common among treasurers.

Fagan said such connectivity can improve cash visibility and reduce risks from fraud and processing errors. She also argued that an agnostic connectivity layer can provide the governed, real-time and standardised data feed needed for AI systems to produce outputs that treasury teams can trust.

External research points to rising experimentation in finance. Fagan cited a global McKinsey survey of chief financial officers in late 2025 that found 44% of respondents had identified five or more AI use cases, up from 7% in 2024. She also referred to Gartner research showing that AI agents are beginning to appear in finance workflows.

Fagan said many treasurers still view data control as a prerequisite for broader AI adoption. Near-term use cases are expected to focus on defined finance tasks rather than full operational automation, while future workflows may involve finance staff working with AI agents on repetitive tasks. In her view, those agents will require structured and current data, leaving manual transfers between systems as a barrier to effective use.

This story draws on original reporting from Finextra Research.