Inside-Out AI: Why Airbnb Starts In-House

Airbnb set a goal of becoming an “AI-native company.” Al-Dahle breaks it down into two phases of work.

The first phase is internal: engineers and product teams get AI tools that shorten the path from idea to release.

The second phase is customer-facing: guests get new search, booking, and support experiences built on the same components. The Latent Space article mentions an internal tool called Everest.

According to the authors, Airbnb used it to accelerate the launch of a new product.

Guests never see Everest, but it determines how quickly the next feature reaches them.

His experience with Llama is valuable because Al-Dahle knows how much infrastructure stands behind a single model response.

In CIS business, an AI project often starts this way.

Management sees a demo, the team builds a website chatbot in a month, and the bot confidently answers the ten questions from the presentation but gets confused by the eleventh.

An external pilot without an internal foundation tests only the interface. Data, integrations, and controls remain untested.

The customer is the first to notice a bot's mistake, and management shuts down the project before the team has time to find it.

The budget is spent, the metric has not moved, and getting the next AI project approved becomes harder.

A customer who receives an incorrect order-status update will most likely go to a competitor.

That is why the company tests AI with its employees while building three layers that the customer product will later need. ### Data layer

The model needs access to accounting systems, the catalog, and documents. At KT.Team, we connect 1C, PIM, and ERP to models in two ways.

The first is MCP, an open protocol that lets an LLM call functions in enterprise systems.

The second is RAG: the model answers using passages retrieved from the knowledge base.

If this setup already works in an internal manager assistant, the customer bot gets ready-made infrastructure. ### Control layer

Requests to the model pass through the LLM & Security Gateway.

The gateway masks personal data, blocks prompt injections, logs requests, and tracks spending by team.

Through it, the team switches between GigaChat, YandexGPT, Qwen, and Claude without rewriting applications.

Internal use cases give the gateway real traffic and usage data before external users connect to it. ###

The AI-native development process layer covers the engineering cycle between “code written” and “working in production”: test generation, reviews, documentation, and incident analysis. A team that has mastered this cycle ships customer-facing AI features faster because each one passes through a proven pipeline.