Automatic correlation catches a lot on its own, but some incidents need a deeper look across logs, metrics, and traces to explain why something broke, not just what broke. That's what AI Analysis does. It's a separate, on-demand investigation you run from a Problem's page, distinct from the automatic correlation described in Problem Details.

Running an analysis

Click AI Analysis on a Problem that hasn't been analyzed yet, or Re-analyze to run it again, for example after new data has come in. Each run is recorded as a version, shown at the bottom of the panel alongside which AI model produced it, such as "Version 1 · GPT-4o Mini (OpenAI)".

Once an analysis has run, its results appear in a panel with five tabs.

Primary

The main tab shows the AI's best explanation for the Problem:

  • Root cause: a plain-language statement of the cause, with a likelihood label (such as "likely") and a confidence percentage.
  • An explanation paragraph describing the reasoning, referencing the actual metrics involved.
  • AI reasoning: an expandable section with more detail on how the AI reached its conclusion.
  • Evidence: the specific data points that support the conclusion, each labeled with its type (metric or exception), its source (such as APM Analysis or Correlated Issues), and its own confidence score. Evidence tied to an error links straight to the error inspector.
  • Validation tests: concrete steps you can run to confirm the root cause yourself, such as running EXPLAIN ANALYZE on a slow query.
  • Recommended action: what to do about it.

Contributing

Lists secondary factors that made the Problem worse without being the main driver, each with a weight and a category, for example a high-latency Redis call that compounded a slow database query. Use this to understand the full picture, not just the headline cause.

Alternatives

Other root-cause hypotheses the AI considered but ruled out in favor of the primary one. This tab is empty when the AI found one clear, high-confidence cause.

Action plan

A checklist of concrete next steps, split into two groups:

  • AI-generated: steps the AI suggests, each with a priority of Critical, High, Medium, or Low.
  • Team-added: your own steps. Add one with the Add step button and a priority of your choosing. Team-added steps persist across re-analyses, so your team's follow-up work doesn't disappear the next time someone re-runs the analysis.

Check off steps as you complete them. The tab's badge (for example "0/2") tracks your progress.

Prevention

Longer-term recommendations to stop this class of Problem from recurring, such as adding monitoring around a query pattern that keeps causing trouble.

Giving feedback

At the bottom of the panel, use Mark as helpful or Mark as not helpful to tell Atatus whether the analysis was useful. This helps improve future analyses.

Note:

Automatic correlation and AI Analysis can disagree. If automatic correlation hasn't confirmed a root cause, the Detected Root Cause panel shows the AI's conclusion as its "current best hypothesis" instead, clearly labeled as coming from AI Analysis rather than confirmed automatically.