Expedia's AI Principles: Scaling Trust and Value in the Agentic Era

Expedia Group, drawing on years of extensive experience and billions of AI predictions across its travel platforms, has established a comprehensive framework of machine learning (ML) and artificial intelligence (AI) principles designed to ensure its AI systems deliver lasting business value, operate safely, and scale responsibly. This proactive strategy comes as the industry transitions from predictive AI to more autonomous, "agentic" AI systems that converse, reason, and increasingly take action on behalf of travelers, raising new expectations for reliability and accountability.

The Shift to Agentic AI

The inherent challenge, as articulated by Xavi Amatriain, Expedia Group’s Chief AI and Data Officer, is that "velocity without discipline and strategic direction is a liability, not an asset." The company emphasizes that true success in AI isn't just about a model working once, but about establishing robust systems that continuously operate, scale across diverse teams and use cases, and consistently improve over time. As AI takes on more autonomous roles, making decisions on a traveler's behalf, the principles behind these systems become paramount.

Expedia's deep experience spans a wide array of AI/ML applications, including personalization, recommendations, fraud prevention, and customer support, now extending into generative and agentic AI experiences. This extensive background informed the development of their guiding principles, which aim to define how AI systems are measured, designed, governed, and operated across the entire organization, with a core goal of ensuring business value, scalability, and safe operation.

From Principles to Practice: Agentic Release Tollgates

Translating these principles into practical action, Expedia has introduced "Agentic Release" tollgates. These are a series of recommended and, in some cases, mandatory checks that teams must clear before launching any agentic AI feature. These tollgates integrate requirements for clear ownership, risk-based governance, thorough evaluation, safe rollout procedures, and continuous monitoring directly into the software development lifecycle. The ultimate goal is to embed these expectations into every stage of AI system design, evaluation, approval, launch, and monitoring.

Three Pillars for Responsible AI

Expedia's framework is structured around three core pillars: Outcomes, Design, and Trust. The Outcomes pillar mandates that every ML effort must directly tie to a key business outcome or traveler experience metric, prioritizing return on cost. Technical optimizations, while useful as midpoints, are not considered end goals. It also requires rigorous offline and online evaluation for all models, with offline predictions reliably predicting online performance.

The Design pillar focuses on building scalable systems that extend beyond individual teams. This means building on shared, platform-wide foundations for core capabilities, treating data as a first-class product with clear lineage, reproducibility, and reliable SLAs. The emphasis is on prioritizing generality, ensuring that learnings and assets can be reused across various teams and brands, rather than optimizing purely for local performance. Manual business rules are minimized and sunset to avoid maintenance debt.

Finally, the Trust pillar addresses responsible operation at scale, ensuring the company can stand behind its AI. Each model demands defined ownership encompassing business, product, AI, and operational aspects, ensuring accountability for outcomes and prompt incident response. Governance is applied proportionally to risk, meaning a customer-facing model impacting millions requires a significantly higher bar for review and human oversight than an internal tool. Expedia also designs for fairness, privacy, transparency, safe rollout, rollback mechanisms, and continuous monitoring to adapt to shifting data and performance.

Setting a Standard for AI Governance

These deeply integrated principles are not merely guidelines; they define what Expedia is willing to ship and how it stands behind its AI solutions. This robust framework is crucial in an era where AI systems increasingly make real decisions for travelers and partners. It positions Expedia to responsibly navigate the complexities and ethical considerations of autonomous AI, fostering long-term success and user confidence by building AI that lasts.

Xavi Amatriain is scheduled to provide further details on Expedia's architecture and "blueprint for building autonomous agents for high-stakes transactional systems" at VB Transform 2026 on July 14.

FAQ

Q: What is the core difference between AI that "just works" and AI that "lasts at scale," according to Expedia?

A: According to Expedia, AI that "just works" may perform adequately once, but AI that "lasts at scale" is built within systems designed to continuously work, scale beyond individual teams, and consistently improve over time. It requires discipline and strategic direction beyond mere velocity.

Q: How is Expedia adapting its AI development for the rise of "agentic AI"?

A: Expedia is preparing for agentic AI, which can converse, reason, and take actions for users, by implementing "Agentic Release" tollgates. These are mandatory checks and recommendations integrated into the software development lifecycle, ensuring clear ownership, risk-based governance, thorough evaluation, and safe deployment for these more autonomous systems.

Q: What are the three main pillars guiding Expedia's AI development principles?

A: Expedia's AI development is guided by three main pillars: "Outcomes," which focuses on linking AI efforts to clear business value and rigorous evaluation; "Design," which emphasizes building scalable systems on shared foundations with robust data practices; and "Trust," which covers clear ownership, proportional governance, ethical considerations, and continuous monitoring for responsible operation.