Airbnb says AI is helping it ship features faster as it tests a new search function

Airbnb’s internal development workflow now leans heavily on generative AI, with the company claiming that a majority of its code is produced by these tools. This shift has reportedly slashed the time needed to move a feature from idea to production by a sizable margin, allowing the team to ship many more improvements within the same calendar period. For engineers, the practical effect is a tighter feedback loop and a need to incorporate AI-assisted coding into their daily tooling and code review processes.

The latest consumer-facing experiment introduces an AI-enhanced search that interprets natural language queries and presents results in a more visual format. Rather than replacing the existing filter-based search, Airbnb adds a toggle so users can opt into the new experience, preserving the legacy path for those who prefer it. Implementing this requires parallel UI components, routing logic that can switch between query processors, and safeguards to fall back to traditional search when the AI model cannot produce reliable results.

On the support side, Airbnb has expanded an AI chatbot that originated in North America to over fifty languages and is planning voice integration later in the year. The bot now resolves close to half of incoming customer issues without human intervention, which has driven a measurable reduction in support costs per transaction. Engineers maintaining this system must monitor language model performance across locales and ensure escalation paths remain robust for the remaining unresolved cases.

From an operational perspective, the dual-search approach and the multilingual support bot mean that monitoring, logging, and alerting need to cover both AI-driven and conventional pipelines. Teams will have to establish metrics for AI accuracy, latency, and fallback rates to avoid degrading the user experience. The optional toggle also implies that any regression in the AI model can be mitigated by reverting users to the stable, filter-based search while fixes are applied.

The rollout is deliberately limited; the AI search is presented as an experiment rather than a default, indicating that full replacement is not yet expected. Consequently, engineers should treat the new feature as a feature flag that can be disabled if performance or user acceptance falls short. This cautious deployment strategy helps contain risk while the company gathers real-world data on how AI-generated results compare to the established search behavior.