How Expedia built scalable AI systems with lasting business value

Expedia’s journey with artificial intelligence spans over a decade, evolving from basic predictive models to today’s agentic systems that handle complex traveler decisions. Along the way, the company learned a hard lesson: rapid deployment doesn’t guarantee long-term success. Without deliberate design, governance, and measurement, AI systems risk becoming brittle, expensive to maintain, and unable to scale beyond their original teams.

The turning point came when Expedia realized that principles alone weren’t enough. Publishing guidelines is straightforward. Embedding them into daily workflows, automated checks, and measurable outcomes—that’s where durable value is created. This transformation required rethinking how models are evaluated, launched, and monitored, with a sharp focus on business impact rather than technical perfection.

From aspirational rules to enforceable standards

Like many enterprises, Expedia began by drafting AI principles aimed at safety, transparency, and scalability. But publishing a document is only the first step. The real work lies in turning those principles into operating procedures that teams follow instinctively.

To bridge this gap, Expedia introduced Agentic Release tollgates—structured checkpoints that enforce accountability before any AI feature reaches travelers. These gateways translate abstract values like clear ownership and risk governance into concrete requirements. Some checks are automated and integrated directly into the software development lifecycle, ensuring teams can’t bypass critical safeguards.

Over time, these expectations are becoming second nature. New AI systems are now designed, evaluated, and monitored with these principles in mind from day one, not bolted on as an afterthought.

Measuring what actually moves the needle

A model’s performance on a technical benchmark rarely translates to real-world success. Expedia’s approach flips this assumption by anchoring every AI effort to a business outcome or traveler experience metric.

• Link models directly to business impact: Every machine learning project must improve a key metric like conversion rates, fraud reduction, or customer satisfaction. Technical improvements are merely intermediate steps, not the final goal.

• Balance value against cost: Before scaling a model, teams must prove it delivers measurable returns that justify its development, training, and operational expenses. Complexity without clear ROI is a liability.

• Start simple; justify complexity: Expedia favors strong baselines—existing models, heuristics, or off-the-shelf solutions—over custom architectures. Specialized models are only adopted when simpler approaches fail to meet performance thresholds.

• Require dual validation: No model skips from offline testing straight to live deployment. Every system must pass both offline evaluations and real-world A/B tests, with offline results reliably predicting online performance.

This data-driven discipline ensures teams focus on outcomes that matter, not vanity metrics that inflate short-term gains.

Building systems that grow beyond their creators

A model that works brilliantly in one team often fails when deployed across multiple brands, regions, or use cases. Expedia’s solution? Shared foundations and reusable assets.

• Standardize core capabilities: Centralize common building blocks like data schemas, feature pipelines, and evaluation frameworks. Specialized models can extend these foundations, but isolated stacks create technical debt that compounds over time.

• Treat data as a product: High-quality models depend on high-quality data. Expedia enforces robust pipelines, lineage tracking, and reusable features with documented ownership and service-level agreements. Teams rely on these shared assets, knowing they’ll improve consistently.

• Prioritize reusability over local optimization: When two approaches perform similarly, Expedia chooses the one that can scale benefits across teams and use cases. This compounding effect accelerates innovation and reduces redundant work.

• Phase out manual rules: Hard-coded business rules are treated as temporary fixes, not permanent solutions. They’re documented, reviewed regularly, and replaced with automated logic whenever possible to avoid maintenance bottlenecks.

• Design for traceability: Every aspect of a model—from training data to deployment versions—must be documented and recoverable. This ensures teams can debug issues months later and transfer ownership without losing institutional knowledge.

Trust is earned, not declared

Deploying AI isn’t just about technical feasibility. It’s about accountability. Expedia assigns clear ownership for every model across its lifecycle: a business owner, a product owner, an AI owner, and an operational owner. These roles may overlap, but the responsibilities are explicit.

Who ensures the model performs reliably at 3 a.m.? Who responds when drift is detected? Without defined accountability, systems become orphaned, and problems surface without clear resolution paths.

Trust is also built through transparency. Teams must justify complexity, document decisions, and demonstrate consistent performance. This level of rigor isn’t optional—it’s the foundation of scalable, agentic AI at Expedia.

As AI systems take on more decision-making power, the standards will only get stricter. The companies that succeed won’t be the ones with the flashiest models, but those that build systems designed to last.