Running a team

Usage & cost

What consumes AI capacity, where to see it, and how to bring it down without giving anything up.

Every workspace meters what it spends. Find it under Usage.

What counts

Anything that makes a model think:

  • Assistant turns — including the tool calls inside one turn, and the work it does to keep a long conversation coherent.
  • Worker runs — on demand and scheduled.
  • Workflow AI steps — the steps that classify, summarise or decide.
  • Voice — transcription and speech.
  • Calls from outside — when an external MCP client triggers work in your workspace, that workspace pays for it.

Metering is per workspace, so a shared account working across three clients can tell which client cost what.

Reading the usage page

You get input and output volume broken down by where it came from, over the period you choose. The useful question is rarely “how much in total” — it’s “what changed”. A flat line that suddenly steps up almost always maps to one new scheduled job.

Bringing it down

In rough order of effect:

Check your schedules. A worker running hourly over a growing dataset is the single most common cause of a jump. Ask whether it needs to run hourly, and whether it needs to look at everything each time.

Right-size the model. A worker that sorts records into four buckets doesn’t need the strongest model available. Set a cheaper one on its definition. Keep the strong model for work that involves judgment.

Narrow the brief. “Review the invoices” makes a worker read everything. “Review invoices created since your last run, over $10,000” makes it read a handful.

Prefer workflows to workers for fixed sequences. A deterministic step costs nothing to think about. Only the genuinely fuzzy steps need a model.

Graduate your hot types. A graduated type queries faster, which means fewer tokens spent on retrieval detours.

During the beta

CogniWork is free while it’s in public beta and there are no charges. Usage is still metered and shown, so you can see the shape of your consumption before pricing exists — and so we can tell you in advance what a workspace like yours would cost. See the Beta Terms for what “beta” commits us to and what it doesn’t.

Next