Expedia Runs 900 Billion AI Predictions a Year. Here Is What It Learned.

Quick Facts

• Expedia Group runs more than 900 billion AI predictions per year across 350 integrated models.
• The company’s AI fraud prevention saved over $2 billion in attempted fraud in the past year alone.
• Expedia’s SVP of Data and AI warns against jumping straight to AI agents, citing risks of bypassing existing workflows and removing humans from decisions.

Expedia Group has spent years building AI infrastructure at a scale most software companies will never reach. The lessons its executives have drawn from that work cut against much of the current hype around AI agents.

The travel platform now runs more than 900 billion AI predictions annually across 350 models embedded throughout its marketplace. That infrastructure sits on top of more than 70 petabytes of historical and real-time travel data, covering over 3 million properties, 500-plus airlines, and 168 million loyalty program members.

Data Before Models

CTO Ramana Thumu says data quality is the starting point, not a secondary concern. “For any AI journey to be successful, it starts with the data at its core as a foundational building block,” Thumu said. His team focuses on building trusted data assets with clear lineage that both business analysts and AI models can use.

Former CTO Rathi Murthy framed the scale of the data problem in travel: “There are around 1.2 quadrillion variables across hotels alone when you factor in room types, dates, and prices.”

Classical ML Still Runs the Business

Expedia’s leaders are direct about one thing many AI vendors obscure: generative AI has not replaced traditional machine learning. Predictive models, ranking systems, time series forecasting, and structured data pipelines still power personalization, fraud detection, and inventory optimization across the platform.

Generative AI changes interfaces. It does not replace the underlying model infrastructure that has run for years.

Models Learn the Wrong Things

Shiyi Pickrell, SVP of Data and AI, described a specific failure pattern Expedia encountered. AI models would learn the patterns of property IDs incorrectly, producing erroneous outputs. “Once they learn a pattern, they form their own [predictions],” Pickrell said. The fix required more rigorous guardrails and enhanced fact-checking built into the model pipeline.

The Agent Warning

Pickrell’s sharpest advice is aimed at teams moving too fast toward agentic systems. “There’s so much hype on agents, and I want to caution ourselves. If you just jump straight to agents, you could not be mindful of existing flows or you may take the human out of the loop,” she said. “There is no compromise on responsible AI.”

That warning carries weight given Expedia’s investment in the space. The company has already built an internal generative AI playground where employees across all roles can test roughly 19 large language models. The platform logged more than 10,000 sessions and produced nearly 200 AI agents internally.

Fraud at Scale

One of the clearest returns on Expedia’s AI investment is fraud prevention. In travel, fraud goes beyond stolen payment credentials. It includes fake reviews, false property listings, inappropriate content, and misuse of vacation rentals. Expedia’s AI systems flagged and blocked over $2 billion in fraud attempts in the past year.

Consumer Products

On the product side, Expedia launched Romie, an AI travel assistant trained on a mix of in-house and OpenAI models that integrates with iMessage and WhatsApp for group trip coordination. At its 30th anniversary event in 2026, the company also announced an Activity Planner, an AI Compare tool on Hotels.com, a Property Expert Q&A feature, and natural-language search on Vrbo.

CEO Ariane Gorin said the company is “experimenting aggressively” across major AI platforms to keep Expedia brands visible in generative search results and functional with agentic browsers, even as AI-driven booking volume remains small today.

The Broader Takeaway

Expedia’s experience points to a discipline that scales regardless of company size: invest in data infrastructure first, maintain classical ML pipelines, build guardrails before expanding model capabilities, and treat agents as an advanced layer built on top of proven systems rather than a shortcut past them.

Read more: What billions of AI predictions taught Expedia before the age of AI agents