Airbnb’s CTO outlines internal AI transformation and agent strategy

Ahmad Al-Dahle, who previously led Meta’s Llama models, joined Airbnb as CTO in January to spearhead the company’s transition into an “AI-native” organization. Al-Dahle described this as an “inside-out AI” approach: using AI internally to accelerate product development before deploying those capabilities to enhance the guest experience.

Airbnb’s customer support is one of the first areas where AI has been deployed at scale. Al-Dahle reported that roughly half of support tickets are now resolved purely by AI, aligning with the company’s Q2 results indicating nearly 45% automation. To ensure reliability, Airbnb generates synthetic data to test systems before production deployment and deliberately avoids automating high-stakes issues such as safety concerns.

The company’s internal context graph tool, Everest, has been instrumental in accelerating the launch of new services like grocery deliveries and airport pickups. Everest uses LLMs, embeddings, and AI-based retrieval to build and query a graph that captures organizational context. This tool enabled the airport pickup service to launch in about six weeks, while grocery deliveries took eight to nine months, and allowed generalist developers to work across specialized parts of the codebase.

Airbnb operates as a “multi-model company,” deploying at least 10 customized models in production. The company evaluates models along a Pareto frontier balancing cost, performance, and latency, tailoring selections to specific use cases. For coding tasks, Airbnb prefers frontier models due to the high cost of defects, whereas latency-sensitive search relies on smaller, specialized models. Al-Dahle noted that post-trained smaller models can sometimes outperform frontier models for narrow use cases.

Airbnb has developed an internal agent called AirChat, which integrates organizational context via MCP. Al-Dahle indicated that async agents represent the next major shift in AI deployment at the company, signaling a broader move toward agentic workflows.