Recognize True Costs

A pattern is playing out across industries right now that nobody wants to talk about out loud. Companies went to their boards and CFOs with aggressive AI cost targets. The math was clean on the slide. AI would bridge the gap, cover the savings, redeploy the headcount. Leadership approved it. Some companies moved fast enough to lay people off before the technology was ready to fill the role.

Now the bills are coming due and the savings aren’t there. A senior advisory executive who works across dozens of organizations describes senior leaders who are afraid to be the first one in the room to say it isn’t working. Nobody wants to be the CFO who admits the ROI case was built on assumptions that didn’t hold. So the bluffing continues until it can’t.

Part of what nobody modeled is what AI actually costs to run at scale.

A director of advanced analytic operations at a major BevAl company describes token consumption as something her organization learned about the hard way. Contracts were signed, tools were deployed, and then the real cost of usage at scale became clear. Teams were not trained on how to optimize what they were consuming. Preferred vendors came with exclusive arrangements, but different teams had legitimate reasons to use different tools, and each came with its own cost structure. The financial picture got complicated fast.

A senior global intelligence executive at a major CPG company puts it plainly: the energy and compute cost of running AI at scale is becoming a reality check for leadership that thought AI was essentially free once the license was purchased. It is not. And when the token bills arrive alongside an ROI case that hasn’t yet proven out, projects get killed.

Another global director of marketing and commercial intelligence at a major beverages company frames the industry dynamic honestly: the whole sector is still flying on hype. At the POC stage nobody demands a real ROI number. But token consumption becomes a genuine issue when you scale, and the industry as a whole has not yet worked out how to prove the math.

The fix is not complicated. Model the full cost of AI before you make the commitment: licenses, compute, tokens, retraining, and the time required to get people using it effectively. Then build the ROI case against that number, not against a sanitized version of it. A business case that holds up under scrutiny will survive a budget cycle. One built on assumptions nobody stress-tested will not.

AI is not cheap to scale. The companies that knew that going in are the ones still scaling.