FinOps grew up around a specific problem: cloud spend that scaled faster than anyone’s ability to track it, spread across teams, regions and services with no shared owner. The discipline that emerged, shared visibility, accountability and continuous optimisation, solved that problem well enough that most mature engineering organisations now treat it as standard practice.
AI spend is running into the same problem today, and in many organisations it is running into it faster than cloud spend ever did. The question worth asking is not whether AI needs its own version of FinOps. It is how much of the existing discipline transfers directly, and where it needs to be extended.
What Classic FinOps Got Right, and Where AI Breaks the Model
Classic FinOps assumes relatively stable pricing, tied to compute, storage and data transfer, with predictable scaling based on usage. Teams could reasonably forecast a quarter’s cloud spend by looking at growth trends, because the underlying unit economics did not change week to week.
AI spend does not behave this way. Token pricing varies enormously between providers and even between models from the same provider, new models ship every few months at new price points, and a single prompt engineering change can shift the cost of a feature overnight without touching infrastructure at all. The unit economics themselves are far less stable, which means a FinOps approach built entirely around cloud pricing habits will miss most of what actually drives AI cost.
The New Metrics AI FinOps Needs
A handful of metrics matter for AI spend that rarely came up in a classic cloud FinOps conversation. Cost per request, broken down by model, tells you whether a feature’s economics are sustainable at scale. Blended cost per million tokens, tracked across every model a team touches, reveals how much of the total spend could be avoided simply by routing more traffic to cheaper models. And the ratio of premium model spend to total spend is often the single most revealing number in the whole practice, since it usually shows a business spending far more on its most expensive model than the actual mix of task difficulty would justify.
Building an AI FinOps Practice
Visibility First
Nothing else in this list works without this step. Every AI provider used across the business needs to feed into one place that shows spend by model, by team and by project. Without that baseline, every later stage of a FinOps practice is guesswork.
Allocation and Chargeback
Once visibility exists, spend can be attributed back to the team or product line generating it, the same way cloud FinOps allocates compute cost to the services consuming it. This is what turns AI spend from an abstract line item into something a product owner actually feels responsible for.
Forecasting AI Spend
Because AI pricing shifts more often than cloud pricing, forecasting needs to be revisited more frequently too, ideally monthly rather than quarterly. A forecast built on last quarter’s model mix can be badly wrong within weeks if a team adopts a new model or a provider changes its pricing.
Governance and Guardrails
Budgets, alerts and approval workflows, standard tools in cloud FinOps, apply just as directly to AI spend. The difference is that AI guardrails also need to cover model selection policy, since the choice of which model handles a given task is often a bigger lever on cost than raw volume. Some cost governance platforms, WrangleAI among them, now build this kind of model routing directly into the governance layer, so a policy decision about which model handles which task gets enforced automatically rather than depending on every engineer remembering it.
Who Should Own AI FinOps
In most organisations today, AI cost sits in an uncomfortable gap. Engineering owns the technical decisions that drive it, finance owns the budget it affects, and neither side has full visibility into what the other is doing. The organisations making real progress on this are the ones that have explicitly assigned ownership, often to an existing FinOps team with a mandate extended to cover AI spend, rather than leaving it to emerge informally.
AI spend is not going to shrink, and the pricing landscape underneath it is not going to stabilise any time soon. Extending FinOps discipline to cover it now, rather than waiting for the bill to force the conversation, is the difference between a planned cost and a recurring surprise.
