AI log summaries

<Note>AI log summaries are only available in Prefect Cloud.</Note>

An AI log summary is a short, AI-generated description of a flow run's logs, such as Failed due to a KeyError in the extract_data task. Instead of scrolling through raw logs to work out what happened, you get the gist in one line.

Summaries are generated on demand and cached, so requesting the same flow run's summary again returns the cached result rather than calling the model a second time.

Enable AI log summaries

AI log summaries depend on AI features being enabled for your account. An account administrator enables them under Settings > Account Settings > Controls by turning on the Marvin AI toggle in the Data processing section. This setting is account-wide and applies to all workspaces.

While AI features are disabled, no summaries are produced and the model is never called.

View a summary in the UI

Open a flow run in the Prefect Cloud UI and select the Logs tab. Click the AI summary icon in the top-right corner of the log panel, next to the log search, to generate a summary of the run's logs. The summary appears inline above the logs. This is useful for investigating a single run after the fact.

Include a summary in automation notifications

You can also surface the summary automatically in notifications sent by an automation, so an on-call recipient sees the likely explanation inline without opening the UI.

Reference the flow_run_log_summary variable in a custom notification template on an automation triggered by a flow run event:

Flow run {{ flow_run.name }} entered state {{ flow_run.state.name }}.

{%- if flow_run_log_summary %}
AI summary: {{ flow_run_log_summary }}
{%- endif %}

A few details to keep in mind:

• Opt in by referencing the variable. The summary, and the underlying call to generate it, is only produced when your custom message references {{ flow_run_log_summary }}. The default notification body does not include it.
• Account setting applies. If AI log summaries are disabled for your account, the variable renders an empty string and no summary is generated.
• Graceful fallback. If the summary cannot be generated (for example, during a model provider outage), or the triggering event has no associated flow run, the variable renders an empty string rather than failing the notification.

Wrapping the variable in an {%- if flow_run_log_summary %} block keeps the notification tidy when the summary is empty, so recipients do not see a dangling AI summary: label.

Further reading

• Templating notifications with Jinja
• Sending notifications with automations