Lawrence Jones from Incident.io @ AIE Europe: building an AI SRE

The podcast discusses the application of AI in Software Reliability Engineering (SRE), focusing on developing tools to manage complexity in production systems through automated root cause analysis and incident investigation. Key challenges include scaling investigations across large numbers of customer accounts and ensuring AI can monitor performance metrics like accuracy and degradation. AIs effectiveness is highlighted with a reported 85-90% accuracy in root cause analysis for well-configured systems, though non-deterministic issues such as logging failures, misconfigurations, and external data inconsistencies pose variability in results. The discussion emphasizes the need for human expertise to complement AI, particularly through runbooks and analysis playbooks that provide contextual reasoning and structure data for meaningful insights, addressing AIs limitations in mathematical analysis and context-based interpretation.

The role of organizational context is critical, with AI tools requiring internal infrastructure and historical data to avoid spurious conclusions, especially compared to generic tools lacking domain-specific knowledge. Log telemetry challenges are also addressed, emphasizing the need for structured summarization to prevent insights from being lost in overwhelming volumes of data. Product integration focuses on enabling direct access to AI-driven investigations via desktop apps linked to tools like Court or Codex, with an emphasis on confidence scoring and alignment with incident management workflows. Examples include AI detecting undocumented system behaviors, such as a 750ms timeout in a telecom providers documentation, and resolving issues faster than manual teams. The potential of ambient analysis tools to identify unpredictable patterns and anomalies is highlighted, alongside plans for the upcoming AI Incident System (AIS), which aims to broadly detect previously undetected issues and streamline collaboration through centralized data sharing during incidents.