The Tokenmaxxing Hangover: Why Enterprise AI ROI Is Still a Moving Target
Earlier this year, a new word quietly took over Silicon Valley boardrooms: tokenmaxxing. The concept was simple — push AI usage as hard and as far as it could possibly go, flood employees with access, and trust that the productivity gains would follow. CEOs championed it. Vendors loved it. And for a brief, optimistic window, it felt like the obvious path forward for any enterprise serious about competing in the age of artificial intelligence.
Then the invoices arrived.
Uber reportedly burned through its entire annual AI budget within just a few months. Some organizations quietly clawed back Claude licenses from entire departments. Meta shut down an internal AI usage leaderboard that had been designed to gamify adoption. What began as a race to embrace AI at scale was now bumping hard into a very old-fashioned problem: no one could clearly prove it was worth the money.
Tiffany Luck, a partner at leading venture capital firm NEA, has been watching this tension unfold from a front-row seat — and her perspective offers some of the clearest thinking available on why enterprises are still struggling to make sense of their AI investments.
The Gap Between AI Activity and AI Value
One of the core problems Luck identifies is the difference between AI activity and AI value. Enterprises became very good at measuring how much AI their employees were using — tokens consumed, queries submitted, features accessed. What they did not become good at was measuring whether any of that usage was actually making the business better.
This is not a trivial distinction. Traditional software ROI is relatively straightforward to calculate. You license a CRM, you track sales cycle improvements, you attribute revenue. AI doesn't work that way. Its value is often diffuse, embedded in dozens of small productivity nudges across an organization, none of which individually appear on a financial statement.
When a developer uses an AI coding assistant to shave twenty minutes off a debugging session, where does that value appear? When a customer service rep uses a generative AI tool to draft a response 30% faster, how is that captured? The honest answer is that most enterprises don't have the measurement infrastructure to know — and without that infrastructure, ROI conversations stall.
Why Enterprises Are in an Awkward Middle Stage
What makes the current moment particularly difficult for enterprise AI adoption is that most large organizations are caught between two extremes. They have moved well past the early experimentation phase — the era of small pilot programs and isolated proofs of concept. But they have not yet reached the stage of deep, workflow-integrated AI that delivers measurable, systemic returns.
This middle stage is expensive, messy, and often politically uncomfortable. Budgets are real and growing. Skepticism from CFOs and boards is growing alongside them. The pressure to demonstrate returns is intensifying exactly at the moment when the infrastructure to measure those returns is still being built.
Luck's view from the VC side underscores something important: the companies that are figuring this out are not necessarily the ones spending the most. They are the ones being most deliberate about where AI is applied, how workflows are redesigned around it, and what success is defined as before the spending begins — not after.
The Tokenmaxxing Lesson: More Is Not Always More
The tokenmaxxing episode is already becoming one of the defining cautionary tales of the 2025 enterprise AI moment. The logic behind it seemed sound: if AI tools improve productivity, then more usage should yield more improvement. Maximize the input, maximize the output.
But this reasoning ignored a few critical realities. First, AI tools are not uniformly valuable across all tasks and all roles. A writing assistant that transforms a marketing team's output may do almost nothing for a finance team running structured data workflows. Blanket adoption incentives push AI into corners where it adds little, while the bill climbs regardless.
Second, and more fundamentally, AI usage without process redesign rarely unlocks the deep value that justifies enterprise-scale spending. If an employee uses AI to do the same job they've always done, just slightly faster, the productivity gain is real but modest. The transformational returns come when entire workflows are rethought around what AI can do — and that kind of change requires deliberate organizational work, not just license distribution.
What a More Mature Enterprise AI Strategy Looks Like
The enterprises navigating this well tend to share a few common characteristics. They identify specific, high-value use cases before deploying AI broadly. They build measurement frameworks — however imperfect — that can track impact at the workflow level rather than just the activity level. They treat AI adoption as an organizational change management challenge, not just a technology procurement exercise. And they stay close to feedback from the employees actually using these tools day to day.
None of this is glamorous. It lacks the energy of a company-wide tokenmaxxing push. But it is far more likely to produce the kind of durable, defensible ROI that justifies continued investment and earns the confidence of skeptical finance teams.
The Bigger Picture: A Necessary Maturation
What Tiffany Luck and NEA's perspective ultimately points toward is something the enterprise technology world has seen before with every major platform shift: an inevitable maturation from hype-driven adoption to value-driven integration. It happened with cloud computing. It happened with mobile. It is happening now with AI.
The companies that will emerge strongest from this period are not the ones that spent the most during the tokenmaxxing era. They are the ones willing to do the harder, slower, more deliberate work of figuring out where AI actually fits — and building the systems to prove it.
That work is less exciting to announce at an all-hands meeting. But in the long run, it is the only AI strategy that genuinely pays off.