Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience

Airbnb’s new CTO, Ahmad Al-Dahle, came from Meta, where he led generative AI and the rollout of Llama over 2023 to 2025. At Airbnb, he’s trying to turn a travel company into what he calls an AI-native one. The plan is simple enough to say and hard enough to do: use AI inside the company to move faster, then turn that speed into better products for guests and hosts.

The biggest change so far is in how software gets built. Instead of the old sequence of requirements, design mockups, engineering handoffs and testing, Airbnb has pushed teams to work directly from prototypes. Al-Dahle says that shift helped a lot with product development, because the code itself becomes the thing people reason about. He says 60% of Airbnb’s code is now AI-authored, the company has shipped nearly 80% more features and improvements year over year, and the average engineer’s pull-request throughput is up about 1.6x.

Support was the first place AI showed up for customers, and it remains one of the most sensitive. Al-Dahle says roughly half of Airbnb’s support tickets are now handled entirely by AI, in line with the company’s latest quarterly figure of nearly 45%. The company leans on synthetic data before systems go live, and it is careful about what it leaves to humans, especially when safety is involved. That caution matters. Support is the kind of place where a clever demo can turn into a customer nightmare very quickly.

Airbnb is also using internal systems to move new services out the door faster. Everest, its internal context graph, uses LLMs, embeddings and AI-based retrieval to connect knowledge across the codebase. That helped Airbnb launch grocery delivery and airport pickups, two partner-driven services that look similar enough for learnings to transfer. In its quarterly filing, Airbnb said groceries took eight or nine months to build, while airport pickups took about six weeks. Al-Dahle says that is the payoff of giving generalists enough context to work across specialist parts of the system.

The company is not betting on one model, either. It uses frontier and open models together, does most of its post-training and reinforcement learning on open models, and says it has at least 10 customized models in production. For coding, it prefers the strongest frontier model it can get. For search, it wants smaller, faster models. And the next frontier, in Al-Dahle’s view, is asynchronous agents that can spin up from events, triage incidents and help run on-call work. That is a very Airbnb answer to the AI moment: less grand theatre, more machinery.