Execs show ‘growing concern’ for AI costs, token consumption

Sponsored by

Header: 59% of professionals spend 11+ hours a week chasing information across systems.†

Copy: This guide breaks down how to connect context across tools and keep work moving.

†Quickbase, 2025

Executives are climbing aboard the AI roller coaster, even if they’re a little concerned about the ride.

A recent survey from global consultancy Protiviti revealed that the C-suite remains concerned about AI-related risks, even as execs continue to focus on onboarding their workforces with the technology.

The survey of 1,330 board members and executives, released on October 8, saw “new risks from implementing AI” rise to the second-biggest concern for the group, falling below only cybersecurity. (The AI risk has been rising steadily in the annual report, moving from tenth position two years ago, to sixth last year.)

Any fear of AI consequences, however, hasn’t slowed execs’ push to “better equip the workforce to realize AI’s value proposition,” which the survey pegs as the number-one priority related to AI deployment (cited by 44% of respondents).

Managers will need to train employees on how to effectively use agents—and how not to use them, Scott Laliberte, global leader for the Emerging Technologies group at Protiviti, told a crowd in New York City on October 6. “Until we get the technology maturation, [employees] are going to be the control that you want to rely on.”

Token gesture. Organizations also remain concerned about leveraging enough value from their AI investments. IT Brew has reported on changing pricing models, and how those modifications have limited IT pros who previously “tokenmaxxed” their AI usage.

With costs rising, executives are more focused on ensuring tokens are used on the right jobs.

“We were seeing developers churn out more and more lines of code. Well, okay, what does that code actually bring you in terms of business value, and what’s the token spend that went along with that? And was that token spend [what] saved you from having to hire and maintain another developer?” Laliberte told attendees. “Absent making that kind of an analysis, you’re just kind of guessing on the value side.”

We caught up with Laliberte after the presentation on how companies can track usage, govern costs, and measure if AI is truly delivering.

Responses below have been edited for length and clarity.

Do you see AI costs as a big concern among executives?

I think it’s a growing concern. I liken it to the early days of cloud, when everybody thought, “We’re going to save all this money and move into cloud,” and then you get the bill, and you realize, “We’re only going to save money if we manage it effectively.”

Do you have any examples of where someone might be having unexpectedly high AI bills?

I think we’re seeing it where organizations create a new app, or they’re using an AI model to create something, or even an agent to do an activity, and they don’t put parameters around it. They let it run over the weekend, and then come back and they’re like, “Oh my God, I just incurred this huge bill,” right? [Execs are] really thinking through what the implications and cost could be of letting that thing run, and putting some limiters and stopgaps in. So, [for example], I’ll let it consume $50,000 worth of tokens, and then it’s going to prompt me to decide if I want to continue or not. We’re just starting to see organizations start to put those types of thoughts and controls together. We call it AI FinOps.

Any other kind of limitations that you’d recommend to keep your bill reasonable?

I think training is a big one. I think utilizing things like model routers, because it’s not just the usage; it’s using the appropriate model for the appropriate task. You can use lower-level models for simplistic tasks, and still get great benefit. You save the higher, more complex, more costly models for those complicated tasks, and really determine whether you need that or not. Is the average user going to be able to figure that out? Some of them, maybe. Hopefully, we’ll start to see a bigger maturation in the model router concept to be able to auto-route tasks to the appropriate model and conserve cost.

What are the characteristics of a good training program for AI?

With AI, first foundational elements that outline the basics of AI—of the security risks and safety risks that people need to adhere to. I like to go back to the existing acceptable use policy, and remind people of the requirements of the acceptable use policy, but then frame it up in the new capabilities of AI.

Step two is to educate them on these safe pathways and tools that they do have…then, I think it’s also [training] department by department on the potential benefits of AI, how they can use AI to achieve productivity gains, specific to their areas.