Expedia Group, a global travel technology leader, has shared the strategic guidelines it developed after processing over one billion AI predictions. Xavi Amatriain, Chief AI and Data Officer, explained how these experiences shaped a framework for building AI systems that are not only effective today but also sustainable at scale. The gap between a model that works in a lab and one that performs in production is vast, and Expedia's lessons are valuable for any company looking to integrate AI responsibly.
Measure real outcomes, not just technical metrics
The first principle focuses on aligning AI metrics with business objectives. Expedia insists that every model must improve a concrete outcome for the traveler, not just a technical indicator. Evaluation must be both offline and online: no model is deployed broadly based solely on offline tests, and no model skips directly to A/B testing. A robust combination is needed to predict real-world impact. Complexity must be justified: start with strong baselines such as existing general models or simple heuristics, and move to specialized models only when simpler options genuinely fail. This avoids unnecessary operational costs and maintenance debt.
Sponsored Protocol
Design systems that scale beyond the teams that build them
A working model is just the beginning. The real challenge is making its value extend beyond the team that developed it. Expedia promotes shared platform-wide foundations, avoiding isolated stacks. Data must be treated as a first-class product, with robust pipelines, clear lineage, and reusable features. When two approaches have similar performance, favor the one whose learnings can be reused across teams, brands, and use cases. Manual business rules should be minimized and reviewed regularly to prevent technical debt. Reproducibility and traceability are mandatory: every decision, configuration, and version must be documented to enable debugging and handoffs without losing institutional knowledge.
Sponsored Protocol
Govern trust with clear ownership and risk-based assessment
The bar for deployment is not just "does it work?" but "can we stand behind it?" Trust is built throughout the model lifecycle. Expedia assigns explicit ownership roles: business owner, product owner, AI owner, and operational owner. Even if not four different people, responsibilities must be clear. Models must adhere to approved standards and governance processes, with release tollgates for agentic systems. Governance should be proportional to risk: a model affecting pricing for millions of travelers demands far more rigor than an internal tool. Design for fairness, privacy, and transparency from the start, not as an afterthought. Rollouts must be progressive, with rollback paths and circuit breakers ready. Once live, continuous monitoring of quality, drift, latency, and cost is essential to adapt to data shifts.
Sponsored Protocol
Expedia's lessons confirm that scalable AI requires a balance of speed and discipline. As Amatriain stated, "principles define what we are willing to ship and how we stand behind it." These standards are already embedded in software development processes, with automated requirements in release tollgates. For further reading, see how Anthropic expands Claude Cowork to mobile and web, another case of agentic AI. Also, scalability is key for DeepSeek's proprietary chip plans. For broader context, refer to the Wikipedia page on Expedia Group. Xavi Amatriain will share more details at VB Transform on July 14, 2026.
