Gen AI spending is moving beyond the familiar software model of seats, licenses, and pilots. As AI shifts from copilots to embedded workflows and autonomous agents, one user request can trigger multiple model calls, retrieval steps, retries, orchestration layers, and infrastructure events. A tool that looks affordable in pilot may behave very differently once connected to production systems or allowed to act with less human supervision.
According to Michael Corrigan, CIO of World Insurance Associates, AI introduces a fundamentally different cost model — one that’s usage driven, non-linear, and tightly coupled to business activity. “Success requires shifting from traditional IT budgeting to FinOps-style discipline where consumption, value, and governance are actively managed in real time,” he says.
Here are five ways CIOs can build that discipline before AI costs spiral.
At World, a top 25 insurance broker with about 3,000 employees across roughly 300 locations, AI use falls into three broad categories, Corrigan says. One is broad tools, such as copilots. Another is embedded AI inside SaaS platforms. And the third is bespoke AI built around specific workflows and manual processes.
“The bespoke is the area that’s growing the most right now,” he says. “And that’s where the model, from a cost perspective, has really been shifting from a license seat cost to a token consumption or token burn cost, or even a hybrid.”
WIA
Seat-based pricing is relatively easy to forecast whereas consumption-based AI isn’t. Costs may depend on prompt complexity, output length, model choice, workflow design, and whether the system calls a model once or many times in the background.
World tries to manage that uncertainty by defining the business problem, success criteria, and expected operational improvement upfront. Pilots help estimate consumption before scaling, but Corrigan says they don’t remove the ambiguity.
“We’ll try our best in the pilot to understand what the consumption rate is, what the token burn rate is,” he says. But once a consumption-based workflow goes into production, he adds, an estimate is put into place. That estimate is informed, but still rough.
Elmer Morales, founder and CEO of koder.com, an agentic AI coding startup, says CIOs should think less about headcount and more about workflow mechanics. Agentic AI costs are driven by the number of decisions an agent makes, how often it retrieves external data, how much context it carries, and how many systems it touches.
“CIOs should start by mapping workflows, not necessarily users,” he says. “The relevant variable isn’t going to be the headcount but how many decisions an agent makes per task.”
Pilots can mislead because they often test the cleanest version of an AI workflow. Morales says many enterprises model agentic AI costs around the happy path: the user gives a clear prompt, the system understands the request, the agent completes the task, and the process ends. Production is messier.
“They generally don’t model for situations where the agent is going to need to go back and check its work and redo things,” Morales says. “A lot of times, agents are wrong, either because they hallucinate or they understood the problem incorrectly.”
In an agentic workflow, the system may check its work, call another tool, retrieve more data, or redo a step. While that may improve quality, it also adds cost.
koder.com
The difference between copilots and agents is central. A copilot interaction is often one prompt and one response. An agentic workflow may involve agents moving through a decision tree, executing tasks in sequence or in parallel, and calling sub-agents or external systems along the way. “By the time it’s achieved the original goal, the agent might have made 50 or 100 model calls, compared with a single call for a traditional copilot prompt,” Morales says.
That’s why CIOs should require teams to model the failure path before production, like how many retries are allowed, how much context is resent, which tools can be called, when a human should intervene, and what happens when the agent can’t complete the task.
Traditional FinOps practices still matter, but AI requires more than retrospective dashboards and chargebacks.
According to Pavan Madduri, senior cloud platform engineer at industrial supply company Graigner, looking backward at usage data, as traditional FinOps often does, can be too late. Costs are shaped by prompt design, model selection, agent behavior, orchestration choices, and runtime loops.
“Dashboards or chargebacks, those are historical accounting,” he says. “The money’s already gone.” For AI, he argues, cost controls need to be embedded into the architecture. That includes hard token caps, retry-depth limits, maximum runtime limits, workload prioritization, background-job throttling, and cluster-level controls that prevent runaway consumption.
“The real FinOps means you need to have the cost constraints embedded into your architecture framework,” Madduri says.
Those controls also extend to infrastructure. Expensive GPUs may sit warm between jobs because systems need capacity available when inference demand arrives. Teams may pass huge schemas, databases, or thousands of lines of code into frontier models when a smaller or more focused prompt would do.
Graigner
Enterprises should also adopt event-driven autoscaling, Madduri says. “Use tools like KEDA to scale GPU nodes down to zero the moment inference demand drops, so teams only pay for the windows when the silicon is actively crunching tokens.”
Corrigan says World uses rate limits, spend limits, alerts, and approval gateways for consumption-based tools. When users approach token consumption limits, automated alerts allow IT and the business to review whether the continued spend is justified.
“If it’s not meeting the success criteria we expected, you have to have the control in place to say we’re going to move on or kill that process,” Corrigan says.
CIOs can also reduce AI bill shock by avoiding a default assumption that every task requires the most powerful model available. While some tasks need advanced reasoning, many others don’t. A simple support ticket, log-parsing task, or structured database transaction may be handled by a smaller or cheaper model. A complex architecture decision, legal analysis, or multi-step reasoning task may justify a more powerful one.
“Choosing the right model for the right prompt and right question — that’s where you leverage the maximum from that model, and you can decrease the costing,” Madduri says. “If you default every single call to a frontier model, that’s architectural laziness.”
Morales makes a similar point. Not every step in an agentic workflow requires a top-of-the-line model. Model routing, he says, is the discipline of determining the best model for the task, and providing the relevant context when the model needs it.
According to Jim Olsen, CTO of enterprise software company ModelOp, CIOs should use the least expensive model that can accomplish the business goal. Using the biggest model for everything is easier, but expensive. “It’s like hiring the most expensive engineer to change a few colors in a website’s CSS, or visual styling,” he says. “You wouldn’t do that. You use the appropriate tools for the task.”
For Olsen, the deeper enterprise problem is AI value shock, not just bill shock. Spending $200,000 in a quarter on AI is justified if it produces $2 million in business value. The problem is spending heavily on use cases that don’t generate a meaningful return. “Are you actually getting that return on investment, or are you just blowing tokens for something that’s not delivering the value to your business?” Olsen asks. Tracking token usage by user or department may show who consumed AI, but not whether the consumption mattered.
ModelOp
For most enterprise AI systems, Olsen says costs should be tied back to business use cases. A model may be used for HR document search, customer support, code review, problem resolution, or other functions. Each use case may draw on the same underlying models or agents, but the business value can be very different. That’s why he argues that companies need an AI inventory, a record of which business workflows use which models, agents, providers, workflows, and systems. Without that inventory, enterprises can’t connect consumption to value.
Corrigan takes a similar approach from a governance perspective. At World, new AI ideas go through an intake process. Business users propose improvements, and IT, finance, operations, sales, and business stakeholders evaluate, prioritize, and monitor them from pilot through production.
That may be where the next stage of AI FinOps is heading, toward a clearer understanding of which AI consumption deserves to scale, not just to lower bills. So the question, as Olsen puts it, isn’t whether someone used a million tokens. It’s what are they using them for.