Summary

• Pigment surveyed 2,000 CFOs and finance executives across the UK, France, Germany and US.
• Eighty-three per cent said consumption-based AI costs had exceeded expectations, with average overspending around 27%.
• Thirty-two per cent reported increased finance headcount because of AI, while a third had reduced hiring for roles AI can cover.

AI spending is proving harder to forecast than its returns, with 83% of organisations surveyed by Pigment reporting that consumption-based costs have exceeded expectations even as finance executives remain highly confident that the technology is producing value.

The Q3 2026 edition of Pigment’s CFO Index surveyed 2,000 CFOs and finance executives across the UK, France, Germany and the US. Thirteen per cent said consumption-based AI spending had remained within expectations, while around 4% could not measure the cost accurately.

Overspending extends beyond marginal forecasting errors. Pigment calculates the average excess at around 27%, with 38% of respondents more than 25% above expectations and 17% more than 50% over. The median respondent fell into the 11% to 25% range.

Usage-based charging makes AI a more variable technology expense because token volumes, model selection, agent activity and automated workloads can all alter consumption. Costs can consequently grow quickly when systems move beyond controlled pilots and enter ordinary work across a larger employee population.

Confidence remains higher than cost control

Unexpected expenditure has not produced equivalent scepticism about returns. Respondents gave AI an average return-on-investment confidence score of 8.2 out of ten, with almost half scoring confidence at nine or ten and only around 12% choosing six or below.

Many finance leaders therefore appear to regard the excess spending as a control problem rather than evidence that AI itself is failing. Budgeting still becomes harder when a technology regarded as valuable also has consumption that moves materially above forecast.

Expected AI budget growth for the next year has moderated to 15.6%, down from 19.1% in the previous quarterly index. That planned increase is below the survey’s estimated average overspend, increasing pressure on organisations to manage usage more deliberately if their forward budgets are to remain credible.

Governance is uneven despite that financial exposure. Fifty-five per cent of respondents described AI tools as centrally managed or subject to comprehensive controls, while another quarter operated an approved-tools model. Nineteen per cent relied on informal guidance or had no formal controls.

Seniority also affects the view of governance. Sixty-six per cent of vice-presidents and CFOs described controls as centralised or comprehensive, compared with 45% of managers, suggesting that the control environment perceived by senior leaders may not always match day-to-day experience.

Hiring changes before employment does

Finance employment data points towards restructuring rather than straightforward automation-led contraction. Thirty-two per cent of organisations said AI had increased finance headcount, compared with around 20% reporting reductions, while more than a third had changed the scope of existing roles.

Hiring shows a different effect because roughly one third of respondents said their organisation had reduced recruitment for jobs where AI can perform some of the work. Changes in employment may therefore appear gradually through positions that are no longer created rather than through large reductions in the existing workforce.

That pattern can have particular consequences for entry-level finance work. Routine analysis, reporting and data preparation have traditionally provided experience through which employees develop the judgement required for more senior positions, so automating part of that work also requires organisations to reconsider how people build those skills.

Training has not moved at the same pace as role redesign. Organisations can change responsibilities through AI faster than they create reskilling programmes, leaving a potential gap between the work employees are expected to perform and the support available to adapt.

The productivity picture is similarly complicated. Respondents reported spending roughly 32 hours each month preparing dashboards and reports for executives, while organisations describing themselves as more advanced in AI adoption did not necessarily report a lighter reporting burden.

Faster analysis can increase the amount of information produced for executives rather than simply reducing the time involved. Automation may therefore change the volume and sophistication of reporting before it translates into a straightforward fall in working hours.

The CFO Index is vendor-sponsored survey research and measures respondents’ perceptions rather than audited expenditure or productivity, so the results should not be treated as direct financial accounts. They nevertheless expose a consistent tension between confidence, consumption and governance.

As AI becomes embedded in routine finance work, cost management increasingly becomes part of AI governance. Organisations that cannot connect usage with budgets, users and business outcomes may find that a technically successful deployment still creates a material financial-control problem.