NEA's Tiffany Luck Warns AI ROI Remains Elusive as Tokenmaxxing Bills Come Due

Uber reportedly exhausted its annual AI budget in just a few months. Meta killed an internal leaderboard that had gamified AI usage. Several companies slashed Claude licenses for parts of their organizations. These are not isolated anecdotes; they are symptoms of a broader pattern that NEA partner Tiffany Luck calls the central tension in enterprise AI right now. Speaking on TechCrunch's Equity podcast, Luck said companies that spent early 2026 maximizing token consumption are now staring at invoices they cannot reconcile with business outcomes. The CFOs who signed off on open-ended AI budgets six months ago have returned with spreadsheets and questions. Luck's position at NEA, where she invests in the AI application layer and B2B SaaS, gives her a portfolio-wide view of how this plays out across dozens of companies. She sees the same story repeating: enthusiasm outpaces measurement, then measurement catches up with a vengeance.

The speed of the reversal matters. Tokenmaxxing went from C-suite mandate to budget line item under review in less than half a year. That compression suggests enterprises never had clear success metrics to begin with.

Luck identified personal AI agents as a technology that still has not found its footing in enterprise settings. The concept, assistants that autonomously handle complex workflows across applications, generates consistent interest but inconsistent adoption. Enterprises pilot agentic systems, encounter integration friction with existing software stacks, and shelve projects before they reach production scale. Luck's view, as reported by TechCrunch, is that the gap between horizontal model capability and vertical implementation remains stubbornly wide. Agents work in demos; making them work in environments with legacy data, compliance requirements, and idiosyncratic approval chains is a different challenge entirely.

This implementation gap creates openings for startups that can bridge it. Luck has backed companies attacking specific vertical workflows rather than building general-purpose agent platforms. The bet is that narrow, deeply integrated tools will outperform broad ones until the infrastructure layer matures.

The scramble to justify AI expenditure has spawned a category of startups building cost tracking and ROI measurement tools. Luck noted on the podcast that some of NEA's most interesting recent investments sit in this space. These companies sell dashboards that map token consumption to specific business outputs, essentially doing for AI what cloud cost management tools like CloudHealth did for AWS. The market need is acute because native tooling from OpenAI, Anthropic, and Google offers usage data but not business context. A department can see it spent $50,000 on API calls without knowing whether that generated $5,000 or $500,000 in value.

This tooling wave reflects a maturation phase. Early AI adoption treated models like magic; current adoption treats them like software with pragmatically tracked costs. Luck's investment thesis, shaped by her background in e-commerce infrastructure at Amazon and tech M&A at Morgan Stanley, favors companies that make opaque systems legible to finance teams.

Luck also addressed the question of which AI companies are ready for public markets. Her assessment, as conveyed by TechCrunch, is that the bar remains high for pure-play AI companies seeking to price IPOs this year. Public market investors have grown skeptical of growth stories without clear paths to profitability, a shift that particularly affects companies still burning cash on model training and inference. The enterprises that constitute the customer base for these IPO candidates are themselves cutting back, creating a demand-side headwind.

However, Luck suggested that companies with durable vertical advantages, strong unit economics, and demonstrable ROI for customers will find receptive audiences. The froth has come off, but capital is still available for businesses that can show they are not merely riding the AI wave but solving specific, measurable problems. This filters the pipeline considerably.

Luck's broader investment philosophy, detailed in an April Crunchbase interview, centers on vertical AI and what she terms the "last mile" of automation. Horizontal platforms from OpenAI or Google provide general capability; the value-creating work happens in adapting that capability to specific industry workflows with proprietary data, regulatory knowledge, and integration depth. This framework explains her skepticism of undifferentiated agent builders and her enthusiasm for companies that embed deeply in single industries.

Her own career trajectory informs this view. After pioneering CPG e-commerce at Amazon before the Whole Foods acquisition, she moved through Morgan Stanley's tech M&A group and GGV Capital before joining NEA roughly three years ago. That path through operational, financial, and venture roles produces a bias toward businesses that can articulate concrete customer ROI rather than technology-for-its-own-sake. In the current environment, that bias looks prescient.

The coming budget cycle will test whether enterprises retreat from AI or simply get smarter about it. Luck's expectation, based on her portfolio conversations, is the latter but with significant variation. Some companies will cut AI spending sharply after poor initial experiences. Others will reallocate toward proven use cases and better measurement infrastructure. The common thread is a demand for accountability that was absent during the tokenmaxxing phase.

This shift creates both risk and opportunity for the AI ecosystem. Model providers face pressure to demonstrate customer value, not just benchmark performance. Startups must show concrete savings or revenue gains rather than future potential. And enterprises themselves need to develop internal competencies for evaluating AI investments that many currently lack. Luck's message, consistent across her recent appearances, is that the reckoning is healthy even if it is painful. Markets work better when buyers know what they are buying.