How is Generative AI actually changing the way SREs handle on-call rotations and incident response?

All answers to this question.

AI is a powerful assistant but a terrible pilot right now. In 2024, we started using LLMs to summarize long Slack threads during outages and to suggest possible fixes based on historical post-mortem data. This has cut our "Mean Time to Acknowledge" (MTTA) significantly. However, I wouldn't trust an AI to automatically execute a script that wipes a production database based on a "prediction." The real value is in data correlation—spotting a pattern across logs that a human would miss in the heat of the moment.

   Answered 2025-07-10 by Megan Roberts

Megan, do you find that the AI suggestions are accurate enough to be useful, or do you still find yourself manually double-checking every single log entry it references?

   Answered 2025-07-12 by Kevin Murphy

• Kevin, manual verification is still 100% necessary. AI can hallucinate a log line or a configuration setting that doesn't exist. We treat AI like a "junior engineer"—it can do the grunt work of searching through thousands of lines of code, but the senior SRE still needs to sign off on the actual repair plan. It’s an efficiency booster, not a human replacement for the time being.

     Commented 2025-07-13 by Megan Roberts

The best use for AI in SRE is definitely writing the first draft of the post-mortem. It saves so much time after an incident when you’re already exhausted.

   Answered 2025-07-15 by Charles Reed

• I agree, Charles. Documentation is the part everyone hates, and AI is surprisingly good at organizing a timeline of events if you feed it the right metadata.

     Commented 2025-07-16 by Deborah Lewis