Consumption for Automations

This page covers how to read the consumption charts in the Metrics tab and how to use them to make decisions.

The Metrics tab

The tab is built around a small set of controls and a stack of charts.

Controls

• Period — Day, Week, or Month. Decides the granularity of the bars.
• Range — how many periods to show. Defaults: 30 days, 12 weeks, 6 months.
• Automation — All automations, or one specific one.
• Apply — applies your filter choices and re-renders the charts.
• Refresh — forces a fresh fetch from the server.
• Export — exports the current view (see Analyze and Export).

Charts

The default view shows these charts, top to bottom:

• Estimated cost — bars of cost per period, in your workspace's currency. The dominant metric for budget decisions.
• Tokens — input vs. output tokens per period. Useful when cost is unclear and you want to know what is driving it (output tokens are typically more expensive than input tokens on most providers).
• Duration — average step duration per period. Sharp spikes often correlate with provider outages or a long system prompt.
• Job count — number of automation runs per period. Useful for seeing trends in usage.
• Success rate — percentage of jobs that completed successfully. Anything below 95 % deserves a look.

How to read them

Cost

Look for trends, not absolute numbers:

• A flat line at high cost is fine if the workflows are doing useful work.
• A flat line at low cost is fine.
• A slowly creeping upward line means something is growing — usage, content size, or a slowly-bloating system prompt.
• A sudden spike means something changed — a new automation was activated, a knowledge grew dramatically, or a misbehaving loop in an agent.

Tokens

If cost is high but you can't explain why, the token chart often points at the cause:

• Lots of input tokens — the prompt itself, the context memory, or the knowledge retrieval is bringing in too much.
• Lots of output tokens — the model is being asked to write long responses where it should be writing short ones.

Duration

Slow steps are usually:

• An LLM Prompt with a very long input or a slow model.
• A knowledge query against a very big knowledge.
• A network-bound integration (HTTP Request, Database Query) hitting a slow target.

If the duration spikes correlate with provider outages on the same day, the cause is upstream.

Job count and success rate

A successful workspace tends to look like:

• Steady or growing job count.
• 95-100 % success rate.
• Failures concentrated in one or two automations (which are then candidates for review).

A typical investigation

You notice this month's cost is double last month's. Step through:

1. Filter Automation: All, period Month, range 6 months. The increase is visible on the cost chart.
2. Switch Automation to the largest-cost candidate first. Apply.
3. The cost chart for that automation tells you whether it is the driver.
4. If yes, open the automation editor. Look at recent jobs to see what input size each step is processing.
5. Often the cause is a knowledge that grew (longer retrieved context = more input tokens) or a system instruction that someone edited to be much longer.

Recommendations

• ✅ Check Metrics at least weekly while a workspace is new. Surprises are easier to catch early.
• ✅ Use the per-automation filter when investigating, not the global view.
• ✅ When optimizing, measure before and after. Tune one thing, run for a few days, compare.
• ⚠️ Daily granularity is noisy. Weekly is usually the right view for trend spotting.
• ⚠️ The cost shown is an estimate; the provider's invoice is authoritative. They usually match closely, but reconcile monthly if precision matters.
• ❌ Do not optimize aggressively for cost on automations that bring obvious business value. The most expensive automation is the one that should run and doesn't.

What to do next

• Analyze and Export — pull the numbers out for further analysis.
• Models → BYOK — reconcile with the provider's billing.