Emily Lewis-Pinnell, Driving AI Adoption at Evaila.
AI cost management has moved from a back-office concern to a front-page enterprise issue. Uber reportedly burned through its 2026 AI coding budget in four months. Microsoft has reportedly begun pulling back internal Claude Code licenses. The FinOps Foundation reports that 98% of practitioners now manage AI spend, up from 31% two years ago, and the Linux Foundation has announced plans to standardize the economics of token-based AI infrastructure.
The interesting part is that AI has gotten cheaper. Cost per million tokens fell sharply across much of the market. So if the unit got cheaper, why did the invoice climb?
Because total spend is price multiplied by volume, and only the price was falling. Volume climbed faster than most budget models accounted for. Some of that is agents looping through context, tools and retries on their own. But much of it is simpler: people doing more, with heavier tools, than anyone planned for. An engineer who used to type the occasional prompt now runs agentic coding sessions that consume orders of magnitude more. Consumption has stopped behaving like a subscription and started behaving like cloud compute. It scales with the work, and AI-assisted work does not scale in a straight line.
Software development is the preview. The most visible AI cost stories are coming from engineers using agentic coding tools, but the pattern will not stay there. As AI spreads into marketing, finance, customer support, legal and operations, the cost pattern will follow the people doing the work. This was never only about unattended automation. It is also about augmentation: how much work AI is doing, for whom and whether that work is worth what it costs.
We have seen a version of this before, with cloud. Surprise bills gave rise to a discipline: tagging, rightsizing, commitments, chargeback—that earned the name FinOps. AI is walking the same path, faster.
But the analogy breaks in one important place. A virtual machine hour is predictable; you can forecast and reserve it. AI consumption is not. The same task can consume wildly different amounts depending on how an agent reasons through it, how much context it pulls in, how many tools it calls or how a person uses the product. More consumption also does not reliably mean a better result. Cloud taught us to manage capacity. AI requires us to manage how work gets done.
Controlling costs requires more than spend controls. It requires a clear view of which work justifies heavier AI use in the first place.
Define the work before setting the budget.
Before setting a limit, look at where AI is used and ask what a good outcome looks like. For an automated process, that may be concrete: fewer manual reviews, faster ticket resolution, fewer handoffs, lower error rates.
For augmented human work, where some of the fastest-growing spend lives, it is harder. The question is no longer only, “Is automating this process worth the cost?” It is, “How much value is this team creating with AI, which roles and activities warrant heavier use and where is consumption growing without a clear return?”
Measure cost per outcome, not cost per token.
Cost per resolved ticket, per reviewed contract, per proposal or per completed analysis connects spend to work in language finance already speaks. For human use, the unit may shift from the task to the person, team or workflow.
A cheaper model that produces more retries, rework or low-quality output is not cheaper. It only looks cheaper on the line item being watched. This is why AI value management is becoming one of the hardest parts of AI FinOps. Companies can often see the tokens. They cannot always see whether the tokens produced value.
Match model depth to work value.
Classification, extraction, summarization and routing do not usually need a frontier model. Reserve expensive models for genuine complex reasoning, high-stakes judgment support and high-value work.
Put a routing layer in front of usage so model choice is a policy decision, not an accident left by whoever built the first prototype. The goal is not to force every use case onto the cheapest model. It is to make sure the model is appropriate to the work.
Use different guardrails for agents and people.
Agents need hard limits: caps on steps, spend per task, runtime, context size and tool access. They also need human checkpoints before actions that change systems of record, trigger customer-facing communication or create operational risk.
People need a different set of guardrails: approved tools, clear expectations by role, guidance on when heavier models are justified and norms for what good usage looks like. Most companies wrote acceptable-use policies for safety and data. Far fewer wrote them for cost and value, which is exactly the gap the spend is pouring through.
Add the spend layer after the operating model is clear.
Once the work is defined, cloud-style controls earn their place: centralized access, attribution by team, role, product and cost center, caching, batching, procurement aligned to stable demand and provider-level budget monitoring.
But do not confuse alerts with control. Some native provider budgets are soft thresholds that notify rather than stop usage. If the business needs hard limits, enforce them at the gateway or orchestration layer.
None of this requires a transformation program. It requires operating discipline, and it can start in 90 days: inventory and approve tools in the first 30, stand up attribution and a model-tier policy in the next 30, and tie spend to value for the highest-use areas in the final 30. The cadence should include Finance, IT, Security and the business owners who own the outcomes.
Tooling will get easier and faster than it did for cloud. The part no tool solves is judgment: which work should be automated, how heavily people should lean on AI and what any of it is worth. Those are leadership decisions. The bill is only where the cost becomes visible. It was decided well before it was ever spent.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
