AI Spending Is Outpacing ROI, Industry Leaders Warn
July 29 (Hypepotamus) - Over a coffee a few months ago in Midtown, a CEO told me they had stopped plans to hire another junior-level developer. Instead, they were seeing how far the team could get with everyone’s new favorite co-worker: Claude Code. But what happened three months later? They’d spent over $55k on increased AI spend. Their job posting for that developer position had planned a hire for $85k for the year.
That CEO, who wished not to be named for the story, is not an outlier. Giant enterprises like Uber have reported astronomical budget spend as AI becomes more integrated into their team’s workflows. And if companies with enterprise-scale finance teams are struggling to track AI costs, startups with far thinner margins face an even steeper climb.
Make every text, call, and lead count. CallRail connects the dots between your campaigns, conversations, and customers.
The AI Bill Does Not Capture the Full Cost
Daniel Ruke, who runs a Florida-based marketing agency that has worked with brands like Disney, Marvel, and Epic Games, told Hypepotamus that he spends about $5,000 each month on AI tools and tokens. But that isn’t the costly part about implementing new AI features.
“The real bill is hours,” Ruke said.
The more deeply a company relies on AI, the more employee time it may need to dedicate to reviewing outputs, identifying errors and repeatedly prompting the system to correct its work. Because AI can deliver inaccurate information confidently, those mistakes may not become obvious until an experienced person reviews the result.
“If my token bill is $5K, the labor wrapped around correcting the machine is worth multiples of that,” he said. “Nobody budgets for it, because it never shows up as a line item. It shows up as your afternoon.”
Maija Ehlinger
Ruke’s suggestion to other businesses trying to navigate AI spend? Set clear boundaries.
“Decide the exact outcome before anyone prompts,” he said.
He also recommends limiting how many times a team retries an unsuccessful task. “If it is wrong after three tries, that model is not the tool for that job, stop feeding it” he said. Other practical steps include making small revisions within an existing chat rather than repeatedly regenerating an entire document, assigning lower-cost models to routine work and reserving more expensive models for tasks that require greater reasoning or judgment.
The final safeguard is human review, he added.
“Keep the judgment human,” Ruke said. “Cheap models for grunt work, expensive models only where the call actually matters, and an experienced person reading everything before it ships.” And remember, he added…tokens are revenue for AI companies, so they have a built-in reason to keep you (and your team) typing.
Agentic AI Changes the Cost Equation
AI spending becomes harder to predict when companies move beyond employee chatbots and begin deploying autonomous agents.
A traditional chatbot responds to a single prompt. An agentic workflow may repeatedly call a model, reconsider its output and pull additional context before completing one task. Each step consumes more tokens, while the expanding context window adds to the bill.
“The nature of agentic AI is that it compounds on itself,” said Jon Winsett, founder and CEO of Atlanta-based IT sourcing optimization firm NPI, which was founded in 2003. Depending on the workflow, Winsett said, an agentic session can consume as much as 30 times the tokens of a typical chat interaction.
That shift has caught some large companies off guard. Winsett said he’s known enterprises that have received monthly AI bills approaching $1 million without being able to clearly identify which teams, tasks or use cases generated the expense…or whether the work delivered a meaningful return.
To try to curb runaway AI costs, it is common for enterprise companies today to have AI Centers of Excellence and broader governance committees. What they often lack, Winsett said, is an AI-focused FinOps function responsible for tracking consumption, assigning costs and controlling usage.
Without that layer, employees and automated workflows frequently default to the newest, most expensive, frontier AI model. Winsett said NPI has seen cases in which roughly 95% of an enterprise’s usage runs through a premium model, even though most routine tasks could be completed by a less expensive model or an open-source alternative.
Moving the same task to a lower-cost model can help keeps costs down, he said. Companies can find additional savings by negotiating different contract terms and improving things like caching and context-window management. Together, those technical changes can cut AI spending forcing enterprises to abandon or rethink their AI initiatives.
“[This] level of spend is not sustainable,” Winsett said. “There’s going to have to be a reckoning.”
See more Hypepotamus stories in your search results
Featured Photo by Suryadhityas / Unsplash