Airbnb Automation Platform: Architecture and Modules for Conversational AI

Introduction

This article introduces how Airbnb's intelligent automation platform (AP) enhances user experience by supporting conversational AI and customer service automation.

With rapid advances in machine learning and natural language processing, conversational AI has attracted great attention, leading to applications such as virtual assistants, smart speakers, and support chatbots.

Airbnb has developed several conversational AI products to improve experiences for hosts and guests, including a customer service chatbot, a task‑oriented machine‑learning framework, and a guest check‑in assistant.

The article focuses on Airbnb's self‑built, enterprise‑grade Automation Platform (AP) that underpins these AI products.

Platform Overview

AP models conversational AI products as a Markov Decision Process, providing a unified representation for workflows and actions, enabling reuse and rapid iteration through a graphical drag‑and‑drop workflow builder.

Core Modules of AP

AP consists of four main modules:

Event Orchestrator : The event‑orchestration layer that normalizes input/output messages between clients and the workflow engine, supporting channel‑agnostic processing.

Workflow Engine : The "brain" of the platform, managing and executing all supported workflows.

Action Store : The action execution engine that processes action requests from workflows, offering a library of reusable actions for developers.

Flow Builder : A graphical interface (GUI) for constructing workflows via drag‑and‑drop, generating workflow templates for the engine.

Event Orchestrator Details

The orchestrator has three layers: the orchestration layer, the context data layer, and the workflow request layer. It handles three input types—external channel messages, internal asynchronous events, and internal service requests—normalizing them into a unified workflow request.

The context data layer stores all request‑related information, identifies existing sessions or creates new ones, and restores workflow state when needed.

The workflow request layer creates requests for the workflow engine and translates responses back to the original channel format.

Workflow Engine Details

The engine comprises four components:

Session Manager: Manages the lifecycle of workflow executions, persisting state for paused sessions.

Schema Loader: Loads workflow templates (JSON schemas) generated by the Flow Builder.

Workflow Executor: Executes workflows step‑by‑step, invoking actions from the Action Store, handling responses, updating variables, and transitioning between states.

Variable Manager: Stores and manages variables used during workflow execution.

Action Store Details

The Action Store provides a common interface for all actions, ranging from simple data retrieval to complex ML‑driven predictions. It consists of an Action Executor, an Action Information Processor, and an Action Manager, which together handle execution, serialization for the builder UI, and registration of actions.

Flow Builder Details

The Flow Builder offers a drag‑and‑drop UI built with React‑diagrams, allowing users to assemble workflows by linking nodes and configuring transition conditions. It retrieves action metadata from the Action Store, generates workflow templates, and sends them to the Workflow Engine for execution.

Summary

The Automation Platform is a generic, enterprise‑grade system that enables Airbnb’s conversational AI products—customer chatbots, guest check‑in assistants, and automated support—to be developed rapidly and deployed at scale. By abstracting workflow orchestration, action execution, and UI configuration, AP empowers business teams to create AI‑driven solutions without deep engineering effort, continuously improving the Airbnb user experience.