AI Spending Ran 27% Over Plan. CFOs Are Budgeting Just 15.6% More for Next Year.

For two years the hard question about AI in finance was whether it would pay off. The question arriving now is simpler and more uncomfortable: what will it cost? New quarterly data shows most companies blew through their AI budgets this year, and the increases CFOs are pencilling in for next year are smaller than the overruns they just absorbed. Unless something changes in how AI spend is forecast and governed, next year's budget is set up to miss the same way.

The Overrun Is Bigger Than the Raise

Pigment's Q3 2026 CFO Index, published October 7, surveyed 2,000 CFOs and finance executives across the United States, the United Kingdom, France and Germany. On AI spending, the result is close to unanimous: 83% say consumption-based costs pushed spending above expectations. Only 13% stayed within budget, and 4% cannot measure their AI spend at all.

The median overrun sits in the 11% to 25% range, but the tail is long. 38% of firms are more than 25% over and 17% are more than 50% over, which brings the average to around 27%. Against that, the expected increase in AI budgets for next year dropped from 19.1% to 15.6%, the lowest figure in the index so far. Pigment reads this as a sign that AI spending growth has peaked. A finance team could read it another way: if usage keeps growing at this year's pace, a 15.6% increase does not even cover the 27% gap already sitting in the run rate.

The pattern is not confined to one survey. A U.S. Bank survey of 1,000 U.S. senior finance leaders, reported by Accounting Today on September 29, found that 51% said spending on AI tools and platforms exceeded their budget over the past year. And SpendHound's survey of 172 finance and procurement professionals found 46% exceeded their AI budget in 2025, against 37% for traditional finance and accounting software. AI is running over at a meaningfully higher rate than the rest of the software stack it sits beside.

Why AI Breaks the Annual Budget

The cause is structural. Traditional software is priced per seat per year, so it can be budgeted once and left alone. Consumption pricing charges for how much the models actually process, which means cost follows usage rather than headcount. CanadianSME's report on the Pigment findings notes that 40% of respondents say unpredictable usage makes AI costs hard to forecast, the most common cost-management challenge in the survey, which was fielded from August 19 to September 21. Pigment's CFO, Susan Phan, described the quarter as "a perfect storm of external political volatility and internal technological disruption."

The forecasting gap is measurable. Earlier this year, CFO Dive reported on Mavvrik's 2026 State of AI Cost Governance Report, which found only 11% of organizations can forecast AI spending within plus or minus 10%, down from 15% in 2025. 43% cited token costs as a top source of unexpected AI spending, and among organizations where an unexpected AI cost materially altered a business decision, 40% required board-level escalation and 33% imposed emergency spending freezes. Forecast accuracy is getting worse as usage scales, not better.

Ownership is part of the problem. SpendHound found 22% of finance and procurement leaders have no single owner for the AI budget, and 57% are not confident they are paying a fair price for AI tools or do not know. Pigment's governance data points the same way: 66% of VP and CFO respondents describe centrally managed or comprehensive AI controls, against 45% of managers, and a quarter of managers say governance is informal or non-existent. The policy set at the top is not reliably reaching the people running up the meter.

Confidence Is High. That Is the Risk.

None of this has dented belief in the technology. Average confidence in AI's return in the Pigment index is 8.2 out of 10, with half of respondents scoring it 9 or 10. In the U.S. Bank survey, 69% said AI investment created commercial opportunities or delivered benefits. The combination of high conviction and weak cost visibility is exactly the condition under which overruns get waved through rather than managed.

Some of the promised efficiency is also slow to show up. Pigment finds the typical finance team spends roughly 32 hours a month producing executive reports, and 30% of the most AI-mature firms spend more than 40 hours a month on it, against 17% of early-stage firms. Faster tools appear to be generating more requests rather than freeing up time. For CFOs trying to justify the next round of AI spend, that is a reminder that the benefit case needs the same rigour as the cost line.

The Playbook: Budget AI Like a Variable Cost

• Re-baseline before you raise: Build next year's AI budget from this year's actual run rate, not this year's plan. With average overruns near 27%, a 15.6% uplift on the original budget starts the year short.
• Forecast on drivers, not seats: Model AI cost from usage drivers such as transactions processed, agents deployed and queries per user, and refresh it monthly alongside the rest of the rolling forecast.
• Name one owner: 22% of companies have no single owner for AI spend. Assign it to a named finance lead with authority over vendor terms, usage caps and the approval queue.
• Set alerts before freezes: Put spend thresholds and usage alerts on every consumption contract so a spike triggers a review, not a board escalation or an emergency freeze.
• Close the governance gap below the C-suite: If managers describe controls as informal while executives call them comprehensive, audit how the policy is actually applied at team level.
• Price the benefit with the same discipline: Tie each AI line to a measured outcome, such as reporting hours saved or forecast accuracy gained, so the renewal decision rests on evidence rather than an 8.2 out of 10 feeling.