A third of CFOs have stopped expanding AI: what they're seeing that others aren't

Finance pays the bills for every AI tool in the business, so it's the first function to question the ROI.

More than a third of finance leaders already using AI have no plans to expand it this year, according to a Forrester study of 1,279 finance decision-makers commissioned by Airwallex earlier this year.

At first glance, that sounds like a retreat. Companies have spent the past few years adding copilots, automation tools and increasingly capable models to workflows across finance, marketing, engineering, and operations. The expectation has been that adoption would keep moving in one direction.

The pause points to something more complicated. Finance tends to see both sides of AI adoption at once. It sees the productivity gains promised by new tools, and it sees the invoices, the implementation costs, and the returns that eventually have to appear in a budget, a margin, or a cash-flow statement. That makes finance one of the first places where enthusiasm for AI meets the question of what, exactly, the investment is producing.

Unless stated, all statistics referenced in this article come from "Building An AI-Ready Finance Function", a commissioned study conducted by Forrester Consulting on behalf of Airwallex, June 2026.

AI is exposing problems that were already there

Limited experience running AI-enabled finance processes is the most commonly cited barrier to adoption in the Forrester study, named by 53% of respondents. Another 40% say they struggle to justify further investment beyond short-term productivity gains. For many companies, the technology also sits on top of data that remains fragmented across entities, systems, and geographies.

Finance teams have spent years accumulating systems. Your company might run different accounting platforms across subsidiaries, maintain separate charts of accounts, reconcile entities manually, and rely on spreadsheets to connect information. This fragmentation forces your team to manually stitch together all these data points.

AI is less forgiving when it comes to consolidating data. For example, an employee can learn that the same supplier can appear under three slightly different names in three ledgers. On the other hand, an AI agent trying to reconcile those transactions first needs the inconsistency resolved somewhere in the underlying data.

Fragmented data stays near the top of the obstacle list for that reason. In the Forrester research, 58% of finance leaders in EMEA identify it as a significant technology challenge, compared with 47% in North America.

The skills problem has the same shape. Finance professionals understand accounting, controls, treasury, and the operating detail of their own businesses. Fewer teams have experience redesigning those processes around AI, which means deciding where automation earns its place and where human judgement still does the work.

That redesign calls for someone who can close a month-end and also work with the underlying data models. That combination of accounting judgement and data fluency barely exists in the job market yet, so most teams will have to build it from the people they already employ.

Finance may be asking AI to pay back too soon

Another possibility sits behind the pause: companies may be asking AI to produce an ROI too soon. Deloitte estimates that a typical AI use case takes two to four years to pay back, against roughly seven to 12 months for other technology investments, and only 6% of the executives it surveyed reported a return within a year.

Most technology investments come with a familiar business case, in which a company spends a set amount, expects a measurable benefit, and judges the result against a defined payback period. AI, however, resists that treatment.

A new AI model might cut the time needed to reconcile accounts, prepare forecasts, or answer questions from the business. The value of those hours then depends on what happens to them afterwards. They might let a company grow without adding headcount, close its books faster, or give the team more room for analysis. But none of that shows up as its own line on the income statement.

AI also tends to arrive alongside new systems, data architectures, processes, and roles. The same research found executives struggling to separate the return AI produced from the return produced by everything changing around it. So the answer is greater precision about what finance measures and over what period.

Experience appears to make that easier. In separate research into finance teams, 30% of teams at the earliest stage of AI adoption say they struggle to justify ROI, falling to 21% among the most advanced. The difference is not enormous, but the direction is significant. Companies appear to get better at finding value from AI as they get better at using it.

An AI pause can mean two very different things

There is a difference between stopping AI investment and stopping long enough to make the next investment work.

A finance team might spend a year buying no new AI tools and still make considerable progress: standardising data across entities, reducing the number of manual hand-offs, defining controls around automated processes, and training employees to work with the systems already in place. Another team might simply wait for the technology to improve. A year later, those two teams will sit in very different positions.

The companies reporting meaningful returns tend to have built AI into their operating processes rather than treating it as a standalone software category.

• ServiceNow told the Wall Street Journal CFO Council Summit that it had generated roughly US$355 million in AI-related savings, with about US$125 million reaching the bottom line and the remainder reinvested elsewhere.

• Shopify expects employees to demonstrate why AI cannot perform a task before requesting additional headcount, which places AI within the company’s resource allocation budget rather than the technology budget.

• At Levi Strauss, an agent built internally reduced parts of the order-entry process from days to minutes.

These are three very different companies, using AI in three very different ways. What they share is a technology attached to a specific operating process and a specific outcome. That’s a harder, but more effective way to approach AI than simply asking how many employees use an AI tool.

The data foundation may matter more than the AI model

The debate around enterprise AI keeps returning to the technology: which model performs better, what agents can now do, and the pace of improvement. The questions that decide the outcome for finance leaders aren’t as exciting. Every entity needs to see the same financial information, transactions need consistent coding, and systems need to exchange data without someone having to export a spreadsheet.

Those questions existed long before generative AI. For a company operating across several entities, currencies, and markets, the foundation can begin with something as basic as a consistent view of cash and balances across the business. Standardised data follows, then the slower work of redesigning processes and training the people who understand them.

Plenty of finance teams are right to hit pause on AI until they’re ready to redesign their foundations. But boards should ask what those teams are building while they wait. A team that consolidates data, sets controls and teaches its people to redesign work around AI will have a place to put the next generation of tools. The Forrester findings point in the same direction, with finance leaders asking what their organisations need to change before AI can work at scale and reach its full, massive potential.

Source: Unless stated, all statistics referenced in this article come from "Building An AI-Ready Finance Function", a commissioned study conducted by Forrester Consulting on behalf of Airwallex, June 2026.