NEA’s Tiffany Luck says enterprises are still figuring out their AI ROI

Last updated: September 14, 2026

Enterprises rushed to deploy generative AI across their organizations in early 2024, only to confront the financial reality of unbounded usage. Uber reportedly exhausted its annual AI budget in a matter of months, while other firms revoked Claude licenses for portions of their workforce and Meta shut down its internal AI usage leaderboard. The pendulum has swung from “tokenmaxxing” — pushing model consumption to its limits — to a sober hunt for measurable return on investment. NEA partner Tiffany Luck, speaking on the firm’s “Founder’s Field Guide” podcast, framed the current moment as a necessary correction: companies are moving from experimentation to accountability, demanding that every GPU hour justify itself against a balance sheet.

The Hangover After the Token Binge

The term “tokenmaxxing” captured the zeitgeist of early 2024. Leadership teams encouraged employees to stress-test large language models across every conceivable workflow, treating compute as an infinite resource. The resulting invoices forced a rapid strategic pivot. Uber’s reported budget overrun became a cautionary tale cited across boardrooms; the ride-hailing giant allegedly burned through its full-year AI allocation in just a few months. In response, enterprises began rationing access — cutting seats for Anthropic’s Claude, imposing hard token caps, and gamifying restraint. Meta’s decision to kill its internal leaderboard, which had ranked teams by AI usage volume, signaled that the metric of success had shifted from activity to efficiency. Luck noted that this cycle mirrors prior infrastructure booms: an initial land grab gives way to cost optimization once the Capex reality sets in.

From Experimentation to Accountability

Luck described a three-phase maturity model now playing out across the Fortune 500. Phase one was broad access — “give everyone a license and see what happens.” Phase two, the current moment, is measurement: instrumenting every workflow to capture time saved, error rates, and revenue impact. Phase three, still aspirational for most, is systematic reinvestment — plowing verified savings into the next layer of automation. She emphasized that the vendors winning renewals are those providing granular observability: dashboards showing cost per task, latency distributions, and quality scores. Procurement teams now demand these metrics before approving seat expansions. The shift has also rewritten vendor roadmaps; model providers are racing to ship cheaper, smaller models and caching layers that reduce redundant inference calls, directly addressing the budget shock their largest customers experienced.

What Small Businesses Can Learn From Enterprise Waste

The enterprise lesson maps directly to smaller operators: automation only pays when it targets real bottlenecks. DigitalEdu deploys AI to handle repetitive work like lead follow-up, scheduling, and data entry, giving teams back 15 to 20 hours per week while keeping costs controlled. Workflow automation saves teams roughly 15–20 hours per employee per week based on DigitalEdu client averages. Small businesses adopting AI automation report approximately 2.5x faster growth than comparable non-adopters, according to industry analysis. Speed compounds that advantage: businesses that respond to a lead within 5 minutes are roughly 21 times more likely to qualify it than those responding in 30 minutes, per industry lead-response research. These figures underscore why targeted automation — not blanket model access — drives ROI. A five-person shop that automates inbound qualification and calendar management captures the same leverage dynamics that enterprises are now scrambling to measure.

The Metrics That Matter for Renewal Season

As 2024 renewal cycles approach, Luck outlined the scorecard CFOs are handing to IT: cost per successful outcome, not cost per token. A customer-support deflection that resolves a ticket without human escalation carries a clear dollar value; a meandering chat session that ends in a handoff does not. She advised founders to instrument three vectors before their next board meeting: (1) unit economics of each automated workflow, (2) quality regression rates when swapping models, and (3) the marginal cost of scaling a proven use case versus the marginal revenue it unlocks. Vendors unable to surface these numbers risk churn, regardless of model benchmark scores. The market is effectively forcing a productization of AI — from “here is a model” to “here is a supported workflow with an SLA and a price tag.”

Implications for the Vendor Landscape

The budget discipline Luck described is reshaping the competitive dynamics among model providers, infrastructure startups, and application-layer companies. Cheaper, specialized models — such as Anthropic’s Claude 3 Haiku or OpenAI’s GPT-4o mini — are gaining traction not because they top leaderboards, but because they fit inside the new cost envelopes. Inference optimization startups selling caching, routing, and distillation services are seeing accelerated sales cycles. At the application layer, vertical SaaS vendors that embed AI into existing workflows — rather than selling a generic copilot — are retaining seats more reliably. Luck predicted a wave of consolidation: point solutions that solve one high-value workflow (contract review, code migration, claims processing) will be acquired by platforms seeking to offer a full suite of measurable automations. For buyers, the message is clear: pilot narrowly, measure ruthlessly, and expand only when the unit economics hold.

The tokenmaxxing era taught the market that raw capability is not a business model. Enterprises are now doing the unglamorous work of matching AI to the specific, repetitive tasks that drain hours and capital. Small businesses that skip the binge and go straight to targeted automation — lead response, scheduling, data capture — capture the upside without the hangover. The winners in the next phase will be those who treat AI like any other capital investment: deploy, measure, reinvest.

For more on this, see DeepMind’s David Silver just raised $1.1B to build an AI that learns without human data.

Related: Obama urges Democrats to have a ‘clear plan’ for AI safeguards, Anthropic CEO outlines plan to ‘pace the frontier’, Mecka AI nears $500M valuation in Sequoia-led deal amid rush for robot training data.

Share this article