Zum Hauptinhalt springen

Introduction

The Metrics tab shows you what your workspace is doing — how much, how often, how expensive, and how reliably. It exists for two reasons:

  1. Cost control — what is each automation actually costing in AI provider tokens?
  2. Operational awareness — which automations are slow, which fail, which are quietly idle?

This section explains what is measured, how to read the charts, and how to export data for further analysis.

What is measured​

Every step of every automation produces metrics. The platform records:

  • Token count — input + output tokens for each model call.
  • Estimated cost — tokens multiplied by the provider's per-token rate at the time of recording.
  • Duration — how long the step took, in milliseconds.
  • Success / failure — did the step complete cleanly?
  • Per-job total — sums of the above across all steps of one run.
  • Per-automation total — sums across all runs in a time range.

The Metrics tab with its controls (period, range, workflow filter) and charts.

Time granularity​

You choose how to slice the data:

  • Day — bars for the last N days.
  • Week — bars for the last N weeks.
  • Month — bars for the last N months.

The range slider next to the period selector controls N. Defaults are sensible — 30 days, 12 weeks, 6 months.

Filtering by automation​

By default Metrics shows everything in the workspace. Use the automation filter to scope to one. This is the typical workflow:

  1. Notice high overall cost.
  2. Filter to each automation in turn and find the outlier.
  3. Open that automation, identify the offending step, and either tune the prompt, switch to a smaller model, or change the schedule.

What is not measured​

  • The provider's actual bill. The platform shows an estimate based on published rates. The provider's invoice is the source of truth for actual cost — usually the two agree within a few percent.
  • Embedding cost on knowledge refresh. This is captured separately from automation cost. See Knowledge → Refresh.
  • Storage cost. The AI Kit does not charge for storage. If you run on-premise, your storage cost is whatever your own infrastructure charges you.
  • End-user time. Time spent by humans interacting with the Inbox is not a metric the platform captures.

How accurate is the cost estimate​

Token counts come directly from the provider's response. The rate per token is configured per model and updated when providers change pricing. The estimated cost is therefore as accurate as the rate configuration — usually accurate, occasionally off by a few percent during transition periods.

For absolute billing precision, reconcile with the provider's invoice once a month.

What to do next​