Airbnb Automation Platform v2: Enabling LLM‑Driven Conversational AI

Artificial intelligence (AI) and large language models (LLM) are rapidly evolving front‑line technologies that are reshaping how we interact with software. This blog explains how Airbnb upgraded its conversational‑AI platform, the Automation Platform, from a static‑workflow‑driven v1 to a new LLM‑centric v2.

Automation Platform Overview

Automation Platform v1 modeled chat‑bots as predefined step‑by‑step workflows that product and engineering teams could design and manage. While it supported a range of conversational AI products, the workflow‑centric approach suffered from limited flexibility and poor scalability: every new use‑case required manual creation of new workflows and tasks.

Challenges of Traditional Conversational AI Systems

Insufficient flexibility – AI products followed rigid, predefined processes.

Scalability issues – Engineers had to recreate workflows for each new scenario, which is time‑consuming and error‑prone.

Opportunities with LLM‑Driven Conversational AI

Early experiments showed that LLM‑based agents can provide more natural, open‑ended interactions, understanding nuanced user queries and extracting implicit information from ongoing dialogs. However, LLMs still face production‑grade challenges such as latency and hallucinations, making them unsuitable for high‑risk tasks (e.g., claims processing) that require strict validation.

The recommended strategy is to combine LLMs with traditional workflows, leveraging the strengths of both.

Automation Platform v2 Architecture

v2 introduces an LLM application layer that orchestrates user inquiries, context gathering, prompt assembly, LLM inference, tool execution, and final response generation. The blog walks through a concrete example where a user asks, “Where is my next reservation?” The platform collects user ID, role, and chat history, builds a prompt, sends it to the LLM, invokes a tool to fetch the latest reservation, updates the context, and finally returns a natural‑language answer.

Chain of Thought (CoT) Workflow

CoT treats the LLM as a reasoning engine that decides which tools to call and in what order. In Automation Platform v2, CoT is implemented as a workflow that repeatedly:

Prepares context (prompt, historical data, user info).

Requests reasoning from the LLM.

Executes the selected tool and feeds the result back to the LLM.

Repeats until a final answer is produced.

The supporting components include a CoT I/O processor, a Tool Manager (handling payloads, retries, rate‑limiting), and an LLM Adapter for custom integrations.

Context Management

Effective LLM reasoning requires rich, relevant context such as prior interactions, user intent, and itinerary details. Developers can declare static context fields or plug in dynamic retrievers (e.g., help‑article search). The main components are:

Context Loader – connects to various data sources and fetches relevant information based on developer‑defined logic.

Runtime Context Manager – maintains per‑request context, merges new data, and interacts with the context store.

Guardrails Framework

To mitigate hallucinations and jailbreaks, the Guardrails framework monitors LLM outputs and enforces safety, relevance, and ethical constraints. Guardrails can run in parallel and include content‑moderation checks (multiple LLMs) and tool‑level rules that prevent unsafe operations.

Future Plans

The team will continue to evolve the platform, explore additional AI‑agent frameworks, extend the capabilities of Chain of Thought tools, and investigate LLM simulation techniques. These innovations aim to boost the productivity of Airbnb’s AI engineers.

Acknowledgments

The blog lists numerous contributors from product, engineering, and leadership who helped build the platform.