As companies increasingly adopt AI, they’re finding that the traditional software model of seats, licenses, and pilots no longer applies. Gen AI spending is becoming more usage-driven, non-linear, and tightly coupled to business activity. According to Michael Corrigan, CIO of World Insurance Associates, AI introduces a fundamentally different cost model.
Corrigan says, “Success requires shifting from traditional IT budgeting to FinOps-style discipline where consumption, value, and governance are actively managed in real time.”
To build this discipline, CIOs can start by forecasting AI costs by workflow, not by user. At World, AI use falls into three broad categories: broad tools, embedded AI inside SaaS platforms, and bespoke AI built around specific workflows and manual processes.
Related: Small firms turn to AI for customer help
Corrigan notes that the bespoke area is growing the most, and it’s where the cost model has shifted from a license seat cost to a token consumption or token burn cost, or even a hybrid.
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.
Elmer Morales, founder and CEO of koder.com, 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.
Related: The Unveiling Brilliance: Diamonds and Their Resonance in Engagement Rings
Morales advises CIOs to model the failure path, not just the happy path. Pilots can mislead because they often test the cleanest version of an AI workflow. In production, the system may check its work, call another tool, retrieve more data, or redo a step, adding cost.
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.
CIOs can reduce AI costs by understanding that not every step in an agentic workflow requires a top-of-the-line model. They need to tie consumption to business value. Spending heavily on use cases that don’t generate a meaningful return is the problem. Companies need an AI inventory, a record of which business workflows use which models, agents, providers, workflows, and systems.
Related: The Engagement Ring Proposal: Planning the Perfect Moment
Without that inventory, enterprises can’t connect consumption to value. Corrigan takes a similar approach from a governance perspective, using an intake process to evaluate and prioritize new AI ideas.
This approach may be where the next stage of AI is heading, toward a clearer understanding of which AI consumption deserves to scale, not just to lower bills. As the issue isn’t whether someone used a certain amount of resources, but what they’re using them for.
