Finance departments have spent the last two years throwing AI at every process they could find, from closing the books to building forecasts. New research from KPMG and PwC shows where that spending is actually paying off, and it is not where most companies expected.
KPMG surveyed more than a thousand senior finance leaders across 20 countries for its 2026 AI in Finance report. Seven in ten said AI is meeting or beating their return on investment targets. But the gains are not spread evenly.
The strongest improvements show up in decision making, forecasting and spotting risk early, the kind of work that depends on judgment. Cost cutting on routine, repetitive tasks shows weaker results. That split is close to the opposite of how most companies pitched AI to their boards.
The original sales pitch was lower headcount costs. The actual payoff looks more like better decisions made faster, which is harder to put a dollar figure on but tends to matter more over time.
The report also found a sharp divide between companies winning with AI and those that are not, and the difference has nothing to do with which software vendor they picked. Firms with strong governance, meaning clear rules for how AI gets checked, approved and corrected, report results three to six times better than firms without it on the same measures. Companies using AI "agents," systems that can act on their own rather than just answer questions, report performance roughly 40 points higher on forecast accuracy and return on investment than companies using AI only to generate answers.
The workforce side tells a similar story. A KPMG quarterly pulse survey found the share of organizations running at least some AI agents in production nearly quadrupled in two quarters, climbing from 11 percent to 42 percent. Meanwhile a survey from the accounting bodies AICPA and CIMA found that 88 percent of finance professionals expect AI to reshape their field within two years, but only 8 percent feel their organization is genuinely ready.
That gap between expectation and readiness is where the risk sits. PwC's response has been to build a formal curriculum pairing AI skills with what it calls human skills: coaching, judgment, agility and reading situations that are not spelled out in the data. That is a tacit admission that the bottleneck was never the AI model itself, it was whether the people around it knew how to question it and act on what it produced.
For any back office function outside finance, the pattern is worth noting. Buying the tool is the easy part. The returns show up for the teams that also fixed their data, built real oversight and trained people to work with the system, not around it.