Intuit Develops AI-Powered Disaster Recovery Assistant Using Amazon Bedrock

Disaster recovery at scale presents significant operational challenges, especially for companies like Intuit that manage millions of users across multiple AWS regions. To address this, Intuit enhanced its existing disaster recovery system, EWOK, by integrating an AI-powered agent called EWOK Agent built on Amazon Bedrock.

EWOK standardizes failover execution across various infrastructure components, reducing recovery times from hours to approximately 20 minutes. However, decision-making during failovers still relied heavily on experienced engineers' tribal knowledge, particularly during sensitive periods like change-freeze windows.

The new EWOK Agent introduces an AI-driven layer that interprets plain-language requests, such as "failover payments-gateway in production," and orchestrates the recovery process. It leverages Amazon Bedrock's foundation models, which provide access to numerous AI providers through a single API, enabling flexible model selection and updates without rearchitecting the system.

The architecture comprises four layers: a consumer layer for engineer interaction, an agent layer that manages model selection and reasoning, a skill layer defining operational procedures, and an execution layer that interfaces with EWOK APIs to perform the actual recovery steps. The system encodes recovery knowledge as structured skills, which are compiled into model-understandable tool specifications, allowing the foundation model to reason about and select appropriate actions.

This setup ensures that the model only decides what to do, while the deterministic execution layer handles how to do it, maintaining clear boundaries and auditability. The use of Amazon Bedrock Guardrails adds security and safety, ensuring data privacy and preventing unintended actions.

Intuit's approach exemplifies how enterprise disaster recovery can be augmented with AI, enabling faster, safer, and more auditable failover processes. The pattern described is adaptable to other operational automation scenarios, emphasizing the importance of structured skills, model pluggability, and bounded reasoning loops.