Use this when reviewing an AI SRE investigation product. Series: Service Rendered Efficiently. Starter: SRE as service.

Review your own deployment. Do not paste customer identifiers into shared copies of this checklist.

Ten questions

  • Investigations per alert — Do we count digs per alert ID weekly and act when one alert sees double-digit reruns?
  • Reuse-first — Does a recent completed RCA short-circuit Slack follow-ups unless the operator asks to re-investigate?
  • Slack triage board — Do incident cards show scannable KPIs and findings, not only dense logs or hollow Incomplete?
  • Undetermined with findings — Does inconclusive triage still post what was checked?
  • Entry-path honesty — Do we train operators to compare watch links from Slack instead of re-pasting alert text?
  • Correlation gates — If the operator asks to link prior sessions, can the run close without that dig?
  • Completeness vocabulary — Can RCAs say COMPLETE / PARTIAL / FAILED instead of fake “no data” on truncated previews?
  • Budget finalization — When the agent hits a turn ceiling, do findings still render?
  • Measure the expression — Do we gate fan-out on measuring the stamped alert query when title and plane disagree?
  • One-zip handoff — Can support download one debug zip per conversation and find threads by initiator / channel?

Red flags

  • Success measured by frameworks shipped or agent LOC
  • Every @mention launches a full investigate
  • Grounding failures invisible in the primary chat bubble
  • Cold-start dead air blamed on “the model thinking”

FAQ

What is the SRE as service checklist?

Ten yes/no questions to review whether your AI investigation product serves on-call — reuse, honest partials, Slack triage UX, correlation gates, and one-zip handoff — instead of optimizing for engineering credibility.