How LATAM Airlines Built Intelligent Agents in Aviation | Interrupt 2026

Nico Venegas and Claudio Urbina Lara from LATAM Airlines — the largest airline in Latin America — walk through what it actually takes to run AI agents at scale in a 3–5% margin industry where every interaction is either value created or value lost. They cover the architecture of LATAM Concierge, a LangGraph-based B2C travel agent with 4,000 daily users, and the hard lessons learned from production: a 15% cost reduction through architectural restructuring, and how a 13% out-of-scope rate turned out to be a product gap, not a model failure. They also introduce Compass, LATAM's internal pipeline for turning millions of unstructured agent conversations into a structured knowledge graph in BigQuery — and why, at this scale, the agent is no longer the product. Chapters: 0:00 What 6,000 passengers in the air right now means for your agents 0:47 LATAM by the numbers: 87M passengers, 3–5% margins, 31 cents of every dollar on jet fuel 2:04 Why extracting intelligence from agents is a different challenge than modeling flights 2:40 Every interaction is value created or value lost 3:51 Why you need a platform before you can build agents: introducing Cosmos 4:12 LATAM Concierge: the B2C travel agent built on LangGraph 4:54 The tool-per-agent architecture and how the supervisor stays in control 5:23 How LangSmith made architectural evolution possible 6:06 What passengers are really telling you when they ask about a restaurant 6:49 Lesson 1: How restructuring the architecture cut costs 15% 7:33 Lesson 2: The 13% out-of-scope problem that turned out to be a product gap 9:10 What questions you can only answer across all conversations, not just one 9:36 Introducing Compass: turning unstructured conversations into a knowledge graph 10:36 The Compass pipeline: parser, mapper, modeler, and ontology registry 11:18 Two examples — UX research interviews and legal contracts 13:09 Bottlenecks, BigQuery Graph, and the architecture decision to ditch Spanner 13:56 The vision: connecting agent graphs across the full passenger journey 15:01 The flywheel: from agent conversations to analytical improvements 15:32 Three takeaways: scale, unstructured data, and why constraints are an advantage 16:30 The agent is not the product anymore — the intelligence across all of them is Resources: → LangGraph: https://www.langchain.com/langgraph → LangSmith: https://www.langchain.com/langsmith → LangChain Academy: https://academy.langchain.com