NEA’s Tiffany Luck: Enterprises are still unsure of AI ROI, and budgets are tightening

Tokenmaxxing looked unstoppable earlier this year. The basic idea was simple: CEOs told employees to push AI usage as far as it would go, then let the work itself prove the value. For a brief moment, it felt like every team could sprint, every workflow could be turbocharged, and every internal experiment could turn into a competitive edge.

Then the bill came due. TechCrunch reports examples of that shift from enthusiasm to accountability: Uber reportedly blew through its annual AI budget in a few months, some companies cut Claude licenses for parts of their org, and Meta killed its internal leaderboard. That combination is the tell. When the incentives are “use it more,” people experiment. When the incentives become “show the ROI,” the org starts slicing costs, tightening access, and rethinking what it actually needs.

This is the tension NEA's Tiffany Luck is describing. In her comments to TechCrunch, she says enterprises are still figuring out their AI ROI. Translation: many businesses are still moving from AI as an activity to AI as a measurable business outcome. That shift is harder than it sounds because ROI is not just about whether the model works. It is about whether it works in your specific processes, with your specific data, at your specific scale, with acceptable risk, and for a cost profile that does not explode the moment usage spreads.

Why this matters right now is that the AI spending conversation is changing shape. Early on, AI value was easy to talk about because it lived in demos, pilots, and productivity anecdotes. Tokenmaxxing, at least in spirit, encouraged breadth over discipline. But budgets are finite and procurement is patient in the way startups are not. When usage scales across an entire organization, even a small inefficiency becomes expensive, and “we tested it” stops being a justification.

The Uber and Meta examples underscore a broader pattern. Uber's reportedly rapid budget burn in a few months suggests that demand can outpace planning when adoption spreads faster than finance can forecast. Meta killing its internal leaderboard suggests leadership is also reconsidering how to drive usage. A leaderboard might motivate experimentation, but it can also accidentally optimize for the wrong metric: activity volume, not business impact. Cutting Claude licenses for parts of an organization, meanwhile, is what cost discipline looks like when teams have different value levels and the company decides to stop subsidizing every experiment equally.

This is where Luck's point lands for executives: if enterprises are still figuring out AI ROI, then the organization is in a transition period where measurement systems, governance, and incentives are still catching up. Boards and CFOs want crisp answers. The business wants agility. That is a classic mismatch, and AI has made it more visible because the marginal cost of usage is real, and the output can be hard to validate at the speed experiments demand.

There is also a second-order implication for leadership teams: “AI ROI” is not one number, it is a stack. Even if a use case generates measurable productivity gains, executives still have to account for implementation costs, integration, change management, and ongoing model or license costs. If different departments see different results, corporate-wide policies become awkward. That is likely why the reported actions are uneven: budgets get cut in certain places, licenses are reduced in parts of the org, and leadership changes how it measures internal adoption.

Luck's framing also hints at the organizational politics of AI. When ROI is uncertain, the debate shifts from strategy to evidence. Who owns the ROI question? Who is responsible for proving it, and how soon? In that environment, tokenmaxxing can turn into a scapegoat, and experiments can turn into budget arguments. The companies that manage this phase well are the ones that treat AI like a portfolio. They experiment, but they also instrument results and stop treating every outcome as equally valuable.

For executives watching this unfold, the lesson is not that AI failed. The lesson is that the early adoption model collided with the reality of spending, measurement, and governance. If enterprises are still figuring out their AI ROI, then peers should expect more procurement friction, more pilot-to-production gatekeeping, and more internal re-scoring of which AI workloads earn continued investment. The strategic stakes are straightforward: the next wave of value goes to teams that can convert AI usage into reliable, trackable outcomes without detonating budgets.