Your AI pilots have a budget line now. Can you explain it?

There is a meeting happening in enterprises across Europe this quarter, and it follows the same script everywhere. Finance has noticed that spending on AI providers, negligible eighteen months ago, is now a line worth examining. The question they bring to IT is not hostile, but it is precise: what are we paying for, who is spending it, and what did we get?

For most organizations, the honest answer is uncomfortable. AI adoption happened team by team, each with its own API key, often on a personal or departmental card. The result is spending that is real, growing, and almost perfectly opaque.

Why the usual tools don't help

Cloud cost management taught us to tag resources and allocate by account. Model providers offer little of that. An invoice tells you which organization spent, rarely which application, and never which business outcome. Token-based pricing makes forecasting harder still: the same feature can cost five times more in a month when users paste longer documents.

The unit of spend is also unlike anything finance is used to. A cloud instance costs the same whether it does useful work or sits idle; a model call costs in proportion to the words that pass through it, which means the bill moves with user behavior in ways no capacity plan predicts. A single feature going viral inside the company, or one team adopting a chattier prompt, can reshape the invoice before anyone files a change request.

You cannot allocate what you cannot see, and you cannot see spending that leaves the building through forty separate API keys.

Consolidate the path, and the numbers follow

The organizations that answer the CFO's question well share one architectural choice: model access flows through a single governed layer. Once that is true, attribution stops being detective work. Every call carries the identity of the team and application that made it; consumption rolls up by cost center as naturally as any other metered service.

• Budgets become enforceable: a team approaching its monthly envelope is throttled or alerted, not discovered in arrears.
• Provider negotiations improve: consolidated volume is leverage that forty scattered keys will never be.
• Waste becomes visible: retry storms, oversized prompts, and abandoned experiments show up as patterns, not surprises.

There is a compounding benefit to consolidation that goes beyond accounting. Once every call runs through one path, the same place that measures spend can reduce it: caching repeated questions so you do not pay twice for the same answer, routing a request to a cheaper model when quality allows, and catching the runaway integration before it runs. Visibility is the prerequisite; the savings follow from it.

The CFO's question is a gift

One Waygrid customer, a European retailer, found that a quarter of its model spending came from a proof of concept everyone believed had been switched off in the spring. The finding paid for the governance project several times over. The CFO's question, it turns out, was a gift.

It is worth reframing the whole conversation this way. The finance team asking where the AI money goes is not an obstacle to AI adoption; it is the thing that makes sustained adoption possible. Spend nobody can explain gets frozen the first time it draws attention. Spend that rolls up cleanly by team, with waste visible and budgets enforced, gets renewed, because the people who sign for it can see what they are buying.