Ahmad Al‑Dahle joined Airbnb as chief technology officer (CTO) in January. Before that he led generative AI at Meta and was involved in launching Meta’s open‑source Llama models between 2023 and 2025. At Airbnb, his mandate is to turn the company into an "AI‑native" organization by first embedding AI into internal workflows to accelerate product development, and then applying those capabilities to customer‑facing experiences — an approach described as "inside‑out AI."
Why move from Meta to Airbnb?
Al‑Dahle said he shifted because, at Meta, teams understood the model improvement flywheel and how to enhance model capabilities across generations. The next challenge, in his view, is deploying models at scale. At Airbnb he aims to change how people work with these tools and to deploy them in production in ways that add value to core user experiences.
Scale and measurable changes
Al‑Dahle cited several figures to illustrate Airbnb’s current AI adoption:
- Approximately 60% of the company’s code is now AI‑authored or AI‑assisted.
- Airbnb has shipped nearly 80% more features and improvements year over year.
- Average pull‑request throughput per engineer is about 1.6× higher.
These metrics are presented as evidence that integrating AI into development processes has materially increased productivity.
Process change: prototypes and code as primary artifacts
A fundamental shift has been organizational: moving away from a sequential product requirements → Figma design → engineering → testing handoff model, toward teams working directly with prototypes. Product, design and engineering teams now iterate on working code and prototypes instead of relying on intermediary artifacts such as long requirements documents.
Al‑Dahle said this reduction in handoff time is among the largest gains traditional software companies need to make to adopt AI effectively, and that the change provided substantial tailwind to Airbnb’s product development.
Customer support: the first customer‑facing AI use
Customer support was the first outward‑facing area to receive AI. Al‑Dahle described support automation as the hardest area to deploy due to high stakes and severe consequences for mistakes.
- Roughly half of Airbnb’s support tickets are now resolved solely by AI; the company’s Q2 results place that figure at nearly 45%.
- To ensure safety and robustness, Airbnb generates extensive synthetic data to test models and agents before production rollout: build a model and agent, then generate a battery of synthetic cases prior to going live.
- The company deliberately leaves some ticket types, such as safety‑related issues, to human handlers, so while agents solve about 50% of tickets, they are selective about which tickets they automate.
Everest: an internal context graph that sped up new services
Airbnb developed an internal organizational context graph called Everest that uses LLMs, embeddings and AI‑based retrieval to represent and query codebase and organizational context. Everest helped accelerate the launch of two services introduced this year: grocery deliveries and airport pickups.
- According to Airbnb’s Q2 report, the grocery delivery service took eight to nine months to build, while airport pickups took roughly six weeks; learnings from the grocery project were recorded in Everest and reused.
- Everest also reduces the need for specialist knowledge on some projects; the context graph allows generalist engineers to navigate specialist code areas.
Both services were highlighted by Airbnb CEO Brian Chesky at the company’s 2026 Summer Release in May.
Multi‑model strategy and model selection
Al‑Dahle described Airbnb as a "multi‑model company," running a variety of models in production. The company mixes frontier and open models, but conducts most of its own post‑training and reinforcement learning on open models. Airbnb deploys at least 10 customized models for production use.
Model choices are evaluated on a Pareto frontier of cost, performance and latency, with different trade‑offs for each application. For each use case the company runs specific evaluations using production‑sampled queries and edge cases. For example:
- For coding tasks, where defects are expensive, Airbnb prefers the strongest available frontier model despite higher latency or cost.
- For search, which is latency‑sensitive and high‑volume, they opt for smaller, specialized models.
Al‑Dahle also said that for narrow tasks, smaller post‑trained models sometimes outperform frontier models by being faster, cheaper and further adapted to the task.
Asynchronous agents: the next shift
Airbnb runs an internal agent called AirChat that contains organizational MCP context. The company is also moving toward asynchronous agents that run in containers and are triggered by events. These agents can automate on‑call triage: when a monitoring alert fires, agents spin up, triage and may even close incidents if they detect flaky alerts, while a human engineer can review a PR the agent proposes.
Al‑Dahle envisions large numbers of asynchronous agents helping to automate marketplace monitoring for fraud, trust violations, quality and software defects.
Preserving engineering craft
Al‑Dahle argues companies should organize around outcomes and missions rather than feature development to succeed in the age of AI. He also worries about junior engineers developing judgment, since AI now performs much of the work. Senior engineers earn judgement through years of shipping and operating systems, including mistakes.
To address this, Airbnb requires that every engineer be able to explain the work an AI produced for them: even if AI generated a pull request, the engineer must explain what they built. This requirement is intended to preserve high standards in interface design, architecture, unit testing and integration testing.
Conclusion
Under Ahmad Al‑Dahle’s leadership, Airbnb is applying an "inside‑out" AI strategy: embedding AI into internal development to speed product iteration, then using those internal capabilities to power customer‑facing services. The company’s internal tools (like Everest), its multi‑model deployments, and measurable productivity gains suggest the approach is reshaping how Airbnb builds and operates its marketplace and products.
