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

The initial honeymoon phase of corporate generative AI adoption—a period characterized by "tokenmaxxing" and unbridled experimentation—has hit a sharp fiscal wall. As enterprises move from low-stakes pilot programs to full-scale production, the underlying economics of Large Language Models (LLMs) are coming under intense scrutiny. Recent reports indicate that major players, including Uber, exhausted their annual AI budgets in mere months, while others are slashing seat licenses for premium tools like Anthropic’s Claude. This shift marks a pivotal transition from the "growth at all costs" mentality of late 2023 to a disciplined search for tangible Return on Investment (ROI) in 2024.

Reflecting on the past eighteen months, the trajectory of enterprise AI has been defined by a rush to integrate tools like ChatGPT and GitHub Copilot to stave off obsolescence. In this early stage, boardrooms pressured CTOs to "do something with AI," leading to internal leaderboards and incentives for high token consumption. However, the lack of a standardized framework for measuring productivity gains meant that many of these initiatives were vanity projects. Now, venture capital leaders like Tiffany Luck of NEA are observing a market correction where the novelty of AI is being replaced by the cold logic of unit economics and bottom-line impact.

The technical mechanics of this friction lie in the unpredictability of inference costs. Unlike traditional SaaS models, where costs are relatively static and seat-based, LLM utilization is consumption-based and highly variable. A single recursive "agentic" workflow can trigger thousands of API calls, ballooning costs in real-time. To mitigate this, companies are moving away from a "one-model-fits-all" approach. Instead of defaulting to massive, expensive frontier models like GPT-4o for every task, engineering teams are increasingly "right-sizing" their compute—using smaller, cheaper models for classification or summarization, and reserving premium models for complex reasoning.

This fiscal tightening has profound implications for the AI competitive landscape. We are entering an era of "LLM pragmatism" where the winners will not necessarily be the models with the highest benchmarks, but those that offer the best performance-to-price ratio. For startups in the AI space, the bar for procurement has risen significantly; they must now prove that their tool doesn't just automate a task, but fundamentally reduces headcount or generates new revenue streams. The decline of internal "AI usage leaderboards" signals that enterprises no longer equate high usage with high value, shifting the power back to procurement officers and CFOs.

Furthermore, the emergence of "shadow AI"—where employees use personal accounts or unsanctioned tools to bypass restrictive corporate budgets—presents a new regulatory and security challenge. As organizations tighten the purse strings, they risk driving innovation underground, creating data silos and security vulnerabilities. This creates a market opportunity for AI governance and cost-management platforms that provide granular visibility into token spending and model performance, effectively acting as "FinOps" for the generative AI era.

Looking ahead, the industry will be watching the next cycle of enterprise renewals. If companies cannot find a way to link AI spend to clear KPIs, we may see a "trough of disillusionment" that slows the pace of adoption. The focus will likely shift toward "agentic" workflows that can perform end-to-end tasks with minimal human intervention, as these offer a clearer path to ROI than simple chat interfaces. The coming year will determine whether generative AI can evolve into a sustainable corporate utility or if it will remain a high-priced experiment for all but the deepest-pocketed firms.