Evals are the new PRD, Expedia’s AI chief tells VB Transform 2026

The investable read-through is not "AI at Expedia," it’s that high-stakes AI is becoming a compliance-and-testing problem before it becomes a revenue feature. That shifts value from flashy model demos to the plumbing around evaluation, monitoring, red-teaming, and rollback. For travel specifically, the economic upside is more likely in lower service costs and better conversion on edge cases than in near-term full automation, because the final-click constraint protects the incumbent funnel rather than opening the door to immediate disintermediation.

The broader winner set is cybersecurity and AI-governance tooling, not the app layer. If enterprises are already deploying partially automated workflows while trusting the controls less than the systems they are replacing, the budget eventually has to flow to vendors that can prove detection, isolation, and auditability at scale; that favors PANW/CRWD/ZS over pure "agent" narratives. The second-order loser is the long-tail of startups shipping autonomous workflows without a credible eval stack: procurement will slow, sales cycles will lengthen, and error remediation costs will eat the early productivity gains.

Consensus is probably too bullish on the speed of fully autonomous agents and too bearish on the time-to-value of governance. In the next 1-3 months, the key catalyst is enterprise incident data and whether AI spend starts migrating from experimentation into security/eval tooling. Over 6-18 months, the structural winners are platforms that make agent deployment safer, while the main falsifier for the thesis is evidence that automation can scale materially without an increase in incidents, support cost, or human review overhead.