Capgemini warns AI spending is running well over budget

Capgemini has found that 59% of executives say their organisation's AI spending was significantly higher than expected over the past 12 months. The findings are based on a survey of 500 C-suite executives at large companies.

The research highlights growing concern about the cost of deploying AI as companies expand its use across more business functions. It found that 53% of executives now see AI costs as a concern, while only 56% said they were fully confident they could forecast how those costs would change as adoption grows.

Among organisations that exceeded their AI budgets, spending was 24% above expectations on average. Around 21% said they overshot by more than 30%.

The report suggests the issue goes beyond the direct cost of tokens used to run AI models. Executives reported limited visibility into indirect costs such as human oversight, rework and governance, with only a third saying they could confidently estimate those expenses.

Half of respondents said AI spending is approved without enough visibility into the full cost. For companies that went over budget, the consequences often extended beyond a single technology programme, with two thirds reporting effects such as lower free cash flow or reduced discretionary spending.

Cost pressures

The findings suggest finance teams are taking a closer interest in major AI decisions. Capgemini said Chief Financial Officers are becoming more involved in approving substantial AI investments as companies seek tighter control over spending.

Many businesses have already responded by imposing stricter checks. More than half, or 53%, said they had increased expenditure control by tightening budgets, applying broader spending controls across the organisation, or reducing the scale of affected initiatives.

Only one in five said they had rethought the tools or processes involved, including changing which activities should be led by AI rather than staff, or altering model choices and routing.

That pattern suggests many companies are responding first with financial controls rather than redesigning how AI is used. It also points to a gap between enthusiasm for rollout and the operational discipline needed to manage costs once systems move into day-to-day workflows.

Volker Darius, Chief Executive Officer of Capgemini Invent, said: "As AI moves deeper into business workflows, organizations need to stop treating its cost as a technology or procurement issue and start managing AI as a portfolio of economic outcomes. Cost discipline alone will not create a profitable AI estate, just as value creation without cost control will not be sustainable. Organisations need to measure cost, quality and business value together, use the most appropriate model for each task, and account for the human oversight required to deliver outcomes safely. By continuously measuring AI Yield, organizations can make more informed capital-allocation decisions and scale AI with confidence."

Model mix

The research also points to a shift in how companies plan to source AI models. More than three in five executives said their organisation expects to adopt a multi-model strategy over the next 12 months.

Under that approach, businesses would choose models based on factors such as cost, latency, risk and the needs of a particular task, rather than relying heavily on one provider or one large model. Executives also said they expected broader use of lower-cost proprietary models where suitable, along with small or specialised models for specific work.

Nearly three in five organisations said they also planned to increase their use of open-weight models. That reflects a search for greater flexibility over deployment and spending at a time when usage is growing quickly.

Capgemini said token usage had risen 14-fold in six months, despite lower token prices for popular models. The figures suggest falling unit costs are not necessarily reducing overall AI bills, as companies run more tasks, expand access and prepare for wider use of agentic systems.

Value questions

The report found that most organisations still struggle to link AI spending directly to business outcomes. Although two thirds of executives said they were confident they could verify the business value of AI-assisted outcomes, most said they either do not formally measure those outcomes, track spending without linking it to outcomes, or keep the two sets of information separate.

Securing funding also remains difficult. Some 57% of respondents said demonstrating enough value to win approval for AI investment was their biggest obstacle.

Only one in five organisations said they calculate AI Yield, which Capgemini defines as the verified value of an AI-assisted outcome after accounting for the full cost of delivering it safely. The business value cited most often by respondents was employee productivity or released capacity, followed by lower operating costs and faster processes. Less than a third, 32%, cited revenue growth.

The survey suggests the measures companies use to judge AI economics may be changing. While 60% currently rely mainly on total AI spending against budget, only 44% expect that to remain the main priority over the coming months. First-pass success rate and cost per completed task or workflow are expected to become more important as companies look more closely at whether AI produces usable results without rework.

The survey covered executives from organisations with annual revenues above USD $1 billion across 10 countries and 13 industries. Respondents mainly worked in AI strategy and leadership, engineering, value realisation, technology finance and AI cost management.