Not long ago, companies were building internal leaderboards to reward employees who used AI the most. Some even tied AI adoption to promotion prospects. The message was clear: use AI for everything, all the time. Now, those same companies are scrambling to stop employees from burning through AI budgets on routine tasks.
The shift is sharp and revealing. According to TechCrunch, consulting giant Accenture has been working to prevent employees from draining its token reserves on simple jobs like converting PDFs into presentation slides. This is the same company that reportedly warned staff they would risk missing out on promotions if they did not use AI enough.
It is a remarkable reversal, and Accenture is far from alone. Across the industry, a pattern is emerging: companies pushed hard for AI adoption, racked up unpredictable costs, and are now asking whether any of it was worth it.
The details come from leaked audio of an internal Accenture meeting featuring Justice Kwak, the firm’s agentic AI strategy lead. His words sum up the mood at a lot of companies right now. “We’re hitting this inflection point where AI is becoming material to the cost structure,” Kwak said. “Spend is becoming very unpredictable; and leadership, especially at the CFO, COO, and CIO level, are still asking the question of whether they’re getting value from what we’re spending on in the context of AI.”
That question, coming from the C-suite, changes everything. When AI was still new and exciting, finance teams were willing to treat it as an experiment. Now that the bills are real and recurring, executives want to see returns. And in many cases, they are not seeing them.
The core problem is straightforward. AI tools charge by the token, which is roughly a chunk of text processed by the model. When employees use those tools constantly, even for small tasks, the costs add up fast. A team of hundreds of people asking AI to reformat documents, draft emails, or summarize reports can burn through a budget that was sized for something far more strategic.
This is the dynamic that gave rise to what some are now calling “tokenmaxxing,” the practice of using as many tokens as possible, often encouraged by management eager to show AI adoption metrics. The problem is that high usage does not equal high value. A lot of that spending produced output that a junior employee could have handled in minutes.
The broader industry is feeling this too. Recent weeks have brought what analysts are calling an “AI selloff,” with some AI-dependent businesses taking a hit on markets. Memory chip makers, which supply the hardware that runs these models, have been among the hardest hit. The signal is hard to ignore: investors are starting to ask the same questions that CFOs inside companies are asking. Is this spending producing results?
The AI industry has reached a point where novelty is no longer enough. For years, the pitch was about potential: what AI could do, what it might become, how it would change everything. That pitch worked long enough to drive massive investment and adoption. But potential does not show up on a balance sheet.
What companies are left with now is a more grounded set of questions:
- Which tasks actually benefit from AI, and which ones just feel like they do?
- How do you measure the value of AI output against its cost?
- What controls do you put in place to stop casual, low-value usage from eating into budgets meant for higher-stakes work?
None of these questions have easy answers. Building token budgets, setting usage policies, and training employees to think about AI costs the same way they think about other business expenses is a significant operational lift. But that is where things are heading.
The era of treating AI as a blank check is closing. What comes next looks a lot more like regular software procurement: justify the cost, track the value, and cut what does not pay off.
