Vercel’s CEO Says AI Assistants Are Shifting From Experiments to Production — And That Means New Battles Over Data

Vercel, the hosting company behind a popular developer platform, is quietly handling massive AI traffic as more teams move from prototypes to full-scale deployments, CEO Guillermo Rauch told reporters after the company’s developer gathering in New York. The firm now processes about 6 million deployments each day, with roughly one-half kicked off by code-writing assistants, and routes in excess of one trillion tokens daily through its AI routing layer, Rauch said. He described a turning point: builders are no longer just testing ideas — they’re trying to run real systems reliably and securely.

Rauch said his team learned the hard way that going from experiments to live services exposes a new set of problems. Early this year Vercel deployed many internal autonomous workflows and discovered two clear, high-value uses. The first is automated coding assistants that accelerate development and account for a lion’s share of token use. The second is internal helpers that surface the right company data to the right people at the right time. To manage both, Vercel has developed a set of controls — a natural-language instruction format Rauch calls Eve, and a contained runtime, sometimes described as a sandbox, to limit what data an assistant can touch and what it can export.

Those controls, he argued, are primarily about preventing accidental data leakage. Rauch recounted conversations with executives at large engineering firms who worry that a misconfigured developer tool could send decades of proprietary source code into a third-party training pipeline. “You need to be able to confine an assistant so it can act, but not siphon your crown-jewel data,” he said, describing why the company built more granular access rules and activity logs for every tool call the assistants make.

Inside companies, Rauch said, assistants are turning data-access bottlenecks into productivity gains. He painted a picture of account managers who previously had to wait weeks for new dashboards; today they can ask a private assistant for the handful of customers that added the most seats in the last two weeks and get an instant answer. That shift, he suggested, moves the constraint from human attention to how well data is integrated and audited. And it raises a broader strategic question: many enterprise vendors have built businesses on keeping customer data siloed, but on-demand assistants work better when data flows are more open — a tension that could reshape enterprise software markets.

The CEO also addressed changing relationships with the major AI model providers. He said customers are increasingly thinking in components — model, orchestration, data controls and deployment — and mixing offerings from different vendors. “Some teams are still tied to a single provider, but a lot are treating models as interchangeable building blocks,” Rauch said, noting rising production use of models from several providers, including Google’s Gemini and a number of open models such as DeepSeek and GLM-5.2. He emphasized that teams are optimizing for cost and latency now, and that can tilt choices toward offerings with better price-to-performance.

That modular approach puts Vercel in a competitive posture with the model makers themselves as those companies add hosting and publish-to-web features. Rauch argued the industry is at a crossroads over whether intelligence will be bundled with the execution environment or remain a replaceable module. “We want to be the underlying backbone for many of these applications, with open protocols and interchangeable parts,” he said, framing Vercel’s ambition as providing the plumbing for next-generation developer workflows. Analysts say that contest over openness, data portability and control will determine who wins the enterprise AI stack in the coming years.