AI’s hidden price tag: Why executives are losing sight of what their models really cost

For many chief information officers, the AI budget conversation has taken an uncomfortable turn. The technology is working, adoption is spreading and boards are asking for more. Yet the invoices keep arriving higher than anyone predicted, and few in the C-suite can say precisely why.

A new snapshot report from the Capgemini Research Institute, published on 12 November 2026, puts numbers on that unease. Based on a September 2026 survey of 500 CXO-level executives at large organisations, it finds that 53% now count AI costs as a concern. More striking is the gap between plan and reality: 59% of respondents say their organisation’s AI-related expenditure over the past 12 months was significantly higher than expected or budgeted.

The scale of the overspend

Among those organisations that overshot, spending ran 24% above expectations or budget on average. Around one in five of them (21%) exceeded their AI budget by more than 30%. Confidence about the future is hardly reassuring either. Only 56% of executives say they are fully confident in their ability to forecast how AI costs will change as adoption scales.

The consequences are not merely theoretical. Two thirds of the organisations that broke their budgets reported negative financial effects, such as reduced free cash flow or cuts to discretionary spending. Unsurprisingly, chief financial officers are taking a closer interest, and are now more involved in signing off major AI investments.

Where the visibility goes missing

The report attributes much of the problem to poor insight into what AI actually costs across the enterprise. Half of executives admit that AI spending is approved without adequate visibility into the full cost. Direct expenses such as tokens are relatively easy to see. The harder figures to pin down are the indirect ones: human oversight, rework and governance, all of which are essential to deploying AI safely. Just a third of executives say they can confidently estimate these costs, and organisations generally lack assurance about the true cost of their existing initiatives.

For CIOs, this should sound familiar. It echoes the early days of cloud adoption, when consumption-based pricing made it simple to start spending and rather harder to understand where the money had gone. The difference now is that the human effort wrapped around AI outputs, from reviewing results to correcting errors, can add a significant hidden layer to the bill.

Tightening controls, then thinking differently

When costs surprised them, most organisations reached first for the brake. Some 53% increased expenditure control, whether by tightening budgets for the affected initiative, introducing organisation-wide spending controls or reducing the scale or scope of the project. Far fewer rethought the underlying approach. Only one in five reconsidered which activities should be AI-led rather than human-led, or changed the solution, models, configuration or model routing.

Looking ahead, however, the direction of travel is clearer. More than three in five executives expect their organisation to adopt a multi-model strategy within the next 12 months, selecting models dynamically according to cost, capability, latency, risk and other business requirements. They anticipate greater use of lower-cost proprietary models where these are sufficient, alongside small or specialised models for particular tasks. Nearly three in five also plan to expand their use of open-weight models.

The implication for technology leaders is that the era of defaulting to a single, most powerful model for every job may be drawing to a close. Matching the model to the task, and routing work accordingly, is becoming a financial discipline as much as a technical one.

Enter AI Yield

The report’s central argument is that spending alone is the wrong thing to measure. Although two thirds of executives say they are confident they can verify the business value of AI-assisted outcomes, most cannot directly connect AI spending with successful results. In some cases they do not formally measure outcomes at all. In others they track costs or usage but not outcomes, or they track the two separately. Proving sufficient value to secure funding is the single biggest obstacle for 57% of respondents.

Capgemini’s proposed answer is AI Yield, a metric capturing the verified value of an AI-assisted outcome after accounting for the full cost of delivering it safely. Only one in five organisations currently calculates it. Of those citing business value, employee productivity or released capacity is mentioned most often, followed by operating-cost reduction and faster processes or cycle times. Fewer than a third (32%) point to revenue growth, including from new products, services or business models.

“As AI moves deeper into business workflows, organisations need to stop treating its cost as a technology or procurement issue and start managing AI as a portfolio of economic outcomes,” said Volker Darius, CEO of Capgemini Invent. He argued that cost discipline alone will not produce a profitable AI estate, and that value creation without cost control will not last. Organisations, he said, need to measure cost, quality and business value together, choose the most appropriate model for each task and account for the human oversight needed to deliver outcomes safely. Continuous measurement of AI Yield, in his view, allows better capital-allocation decisions and more confident scaling.

The metrics are about to change

Executives appear to recognise that their dashboards need updating. Today, 60% primarily manage AI economics by comparing total expenditure with budget. Over the next six to 12 months, only 44% expect to keep prioritising that measure. First-pass success rate is set to become the top-priority metric, with cost per completed task or workflow also rising in importance. Both shift attention from how much AI is consumed to how reliably it delivers a finished result.

For CIOs preparing for more agentic use of AI, where systems take multi-step actions with less supervision, that shift matters. An agent that completes a workflow correctly the first time is cheaper in every sense than one that needs repeated human correction, even if its per-token price is higher.

What it means for technology leaders

The snapshot follows the first issue of The AI Compass, Capgemini’s new paper on AI economics, titled The Era of AI Yield: Mastering Tokenomics to Maximise Enterprise Value. It explores how organisations can look beyond consumption and cost metrics to measure the verified value of AI-assisted outcomes, minus the full cost of achieving them safely. The publication also sets out implications for CEOs and technology leaders, covering capital allocation, governance, model routing, workflow engineering and human oversight.

The message for CIOs is plain enough. With finance chiefs paying closer attention and budgets under strain, the ability to show what each pound spent on AI actually returns will increasingly separate the organisations that scale with confidence from those forced to pull back.