The AI Budget Paradox: Why Traditional IT Budgeting Fails for AI

The adoption of AI is forcing finance departments to fundamentally change course. Whereas IT budgets consisted for decades of predictable fixed costs, AI introduces unprecedented volatility. Now that vendors are shifting their pricing models from fixed licences to 'pay-as-you-go', fixed contracts are giving way to variable spending based on token consumption. CFOs are currently trying to rein in these unpredictable streams of invoices, but at its core, that only treats the symptoms.

This defensive reflex creates a strategic problem: you calculate what AI costs, but not what it delivers. By treating AI purely as an expense, you sidestep the real question: how do we transform these new, variable IT expenses into a lasting, measurable return on the profit-and-loss statement?

Why CFOs struggle with the cost side of AI

The CFO's initial focus on cost control is entirely rational. Many organisations have, over the past twelve months, fallen prey to the 'pilot trap'. Within the isolated AI test environment of a proof of concept, the costs of AI look negligible. But as soon as such a model is integrated into core processes and scales up to tens of thousands of interactions in production, the financial reality check follows: infrastructure costs explode.

A cost-control strategy is a crucial first step. But as soon as it becomes the only objective, a real danger emerges. By focusing solely on containing AI costs, the organisation fixates on driving down invoices from technology vendors. In doing so, it loses sight of the actual value those investments are meant to create.

The risk of a one-sided focus

While the CFO puts up financial guardrails, a blind spot emerges for real value. Many AI business cases currently lean on the assumption that 'saving time' automatically equals 'making money'. Although research by the National Bureau of Economic Research (NBER) shows that AI lets employees work faster, in practice that speed does not yet translate into lower costs or higher revenue. The hours saved are often absorbed or simply not put to use for revenue-generating activities.

Managing for value: the shift to AI ValueOps

AI is becoming a permanent, fixed layer in our IT architecture. The pitfall is that these costs rise without a solid business case behind them. To correct this, we must not see these expenses as an 'unavoidable IT burden', but as costs that have to prove their value in operational KPIs and ROI.

This shift from blind cost to lasting value management (ValueOps) rests on three pillars:

• The yardstick: from 'time saved' to hard numbers. 'Saving hours' is not a result but an activity. A recurring AI line item on the budget is only justified once it directly leads to cost reduction or greater productivity elsewhere.
• The owner: the business pays, not IT. The bill for AI consumption still too often lands with IT, and that is no longer tenable. Managers who deploy AI for their team also come to own the costs. Treat AI usage as an operational expense: like staff costs, these expenses belong directly on the profit-and-loss statement of the relevant department.
• The feedback loop: managing for KPIs in real time. We need reporting that sets the costs of AI directly against business results. If AI costs for a specific operation are rising while average handling time is not falling, and output is not increasing, that is waste. That is the right moment for the CFO to step in.

Questions for the boardroom

Leading organisations will not distinguish themselves merely by controlling the costs of AI, but by their ability to translate those expenses directly into measurable business value. The following strategic questions offer a starting point for the boardroom dialogue:

• The return question: Are we managing for facts or for assumptions? Is the success of our AI investments directly visible in our P&L figures, or do we remain stuck in the 'time-saving trap', where the claimed efficiency shows up nowhere in the business results?
• The accountability question: Who controls the spending? Do the bill for AI consumption and the responsibility for the return sit with the business units deploying the AI, or do we place them collectively with IT, so that the incentive for strict return monitoring in the operation disappears?
• The growth question: What do we do with the capacity we free up? Are we ensuring that the time AI frees up is actually deployed for new growth and revenue-generating activities, or do we quietly let those gains 'evaporate' into existing operational overhead?