What does artificial intelligence really cost when an experiment becomes a production service used across an enterprise? Token consumption is easy to see, but Greg Holmes argues that it may represent only a small part of the final bill.
In this episode of Business Tech Perspectives, I speak with Greg, EMEA Field CTO at Apptio, an IBM company, about AI economics, technology business management, and the growing challenge of connecting infrastructure spending to business value. Greg works with CIOs, CFOs, and technology leaders across EMEA, where conversations that once centered on cloud cost and FinOps are increasingly focused on a simple question: is AI worth what the organization is spending?
Greg describes how teams can move surprisingly far through an AI pilot before finance sees the full consumption pattern. The token bill may then reveal a larger operating model involving data storage, data movement, cloud infrastructure, governance software, security, and the people who built and maintain the service. In examples discussed during the interview, he says direct AI costs can account for around 20 percent of the total. That figure should remain attributed because every workload and enterprise will have a different cost structure.
The measurement problem becomes harder when businesses ask for a single return-on-investment formula. Greg argues that each initiative needs a clear unit of work. A service desk system might be measured by tickets resolved, an insurance system by claims processed, and a document service by files completed accurately. Leaders can then compare the value of that completed work with its full production cost rather than treating token usage as the result.
We also discuss agentic AI, where autonomous activity can consume resources continuously and unpredictably. Greg recommends guardrails such as a spending ceiling per agent, automatic checks when consumption rises, and model tiering. A cheaper model may complete the straightforward 80 percent of tasks successfully, leaving an expensive frontier model for work that needs greater capability. This approach gives engineers room to test ideas while limiting the risk of applying the costliest system to every request.
Another concern is behavior. Greg shares examples of organizations that encouraged employees to prove they were using AI by consuming their full token allocation. Engineers responded to the metric and built tools that used tokens because appearing low on the usage chart looked bad. The organization received the behavior it rewarded, even when that behavior had little connection to a useful business result.
That lesson matters for the relationship between finance and engineering. Finance wants predictability, while technologists need room to test new ideas. Technology business management and FinOps can give both sides a shared language based on services, unit costs, and outcomes. Greg argues that visibility creates the basis for trust, allowing a team to discuss whether a service is worth funding rather than fighting over an unexplained bill.
We also examine the belief that AI will inevitably become cheaper. Models and computing methods may become more efficient, but growing demand, data center construction, energy constraints, and the creation of additional workloads could keep total spending high. Greg’s advice is to make a production service financially worthwhile from its first day rather than depending on future price reductions to rescue a weak business case.
Is your organization counting AI tokens, or can it explain the full cost and the business result produced by every AI service? Listen to the episode and share how your finance and technology teams are measuring AI value today.