Instacart Develops Blueberry, an AI Assistant for Incident Investigation

Instacart, a grocery delivery service, has developed an AI assistant called "Blueberry" for engineers. This tool is designed as a support system to help on-call engineers who respond to system incidents investigate the root causes more quickly.

In modern software services, unexpected problems occurring in production environments (systems that actual users access) are not uncommon. When such "incidents" occur, responsible engineers must manually dig through various data called logs and metrics to identify the root cause of the problem. This investigation work is time-consuming and relies heavily on experience and intuition, making it a significant burden for engineers. Blueberry was created precisely to address this "burden of investigation."

Blueberry's mechanism combines three elements: AI agents, operational data, and historical incident records. An AI agent refers to an AI program that autonomously collects and analyzes information toward a given objective. Blueberry can collect and analyze information from multiple angles by running multiple sub-agents in parallel. It also incorporates MCP (Model Context Protocol), a mechanism for integrating external tools, enabling seamless data exchange with various internal systems.

Analysis results are displayed on "Slack," a business chat tool that engineers use daily. Blueberry references historical records of similar incidents and generates "hypotheses about root causes" to present to engineers. Importantly, final judgment and remediation remain the engineer's responsibility, with Blueberry serving to organize and provide information that supports that judgment.

This design reflects the philosophy of positioning AI not as the primary actor in complete automation, but as a "supporting role" that draws out human expertise. Because engineers can receive results within the Slack platform they already use, the tool can be integrated into operations while reducing learning costs associated with new tools—a notable advantage from a practical standpoint.

For consumer-facing services like Instacart, the recovery speed when incidents occur directly impacts user experience and business loss, making investigation time reduction a response to actual business challenges. This initiative can be understood as an example of how engineering organizations are leveraging AI to improve operational efficiency.

Regarding quantitative results—how much Blueberry can actually reduce investigation time and against what types of incidents it is effective—no publicly disclosed figures are available at this time. As results from real-world deployment are shared in the future, Blueberry could serve as a reference case for engineering organizations facing similar challenges.