How do I read a service's Usage tab?
A service's Usage tab answers two different questions, which is why it has two sub-tabs. License usage scores whether each seat is earning its keep. Telemetry shows what each person actually consumed, in tokens and cost.
Use the first to decide who keeps a licence, and the second to understand how a tool is being used.
In this guide
Before you begin
You need Admin or Manager access.
The tab only appears on services that have usage data or a usage provider. On everything else there is nothing to show and the tab is absent — not a permissions problem.
License usage
One row per licence:
| Column | What it shows |
|---|---|
| Employee | Who holds the seat. Search also matches their work email |
| Plan | The plan the seat sits on |
| Score | Usage health from 0 to 100 — a dash means nothing has been scored yet |
| Latest metrics | The raw counts from the most recent period the score was built from |
Low scores are the seats worth reclaiming. Latest metrics is there so a score is never a black box: you can see the numbers behind it before acting on somebody's access.
Scores come from telemetry and are recalculated daily. Reclaiming seats is covered in How do I find and reclaim unused licences?.
Telemetry
One row per person, for the period you select:
| Column | What it shows |
|---|---|
| Employee | Who the usage is attributed to, matched on the email in the telemetry |
| Tokens | Input plus output for the period — cache is counted separately |
| Input tokens | Sent to the model, excluding anything read from or written to the cache |
| Output tokens | Generated in response |
| Cache read | Served from the prompt cache — far cheaper than fresh input |
| Cache creation | Written into the cache for later reuse — priced above regular input |
| Reasoning | Spent thinking before answering, where the model supports it |
| Cost | Approximate spend at API-equivalent pricing |
| Models | Which models that person called |
| Last active | Most recent day with recorded usage in the period |
Three tiles sit above the table: Active users (with how many of your licensed seats that represents), Total tokens broken into in and out, and Estimated cost.
Filter by Time period — 30d, 3m, 6m, 12m or All — by model, and by Attribution. Search matches employee, email or model.
Reading the cache columns
Cache read tokens are much cheaper than fresh input, so a high cache share is a sign a tool is being used efficiently, not a sign of heavy spend. Judging cost by total tokens alone will mislead you: two people with identical token counts can differ several-fold in cost depending on how much came from cache.
The three kinds of row
Attribution separates them:
| Kind | What it means |
|---|---|
| Matched employee | Tied to a person in Velory — the normal case |
| Unmatched seat | The email in the telemetry matches no employee. The raw email is shown instead. Badged Unmatched |
| Former employees | An anonymised bucket for people who have left. Badged Anonymized |
Unmatched rows are worth clearing: fix the address on the employee record and matching resolves itself as new usage arrives.
Note that the Active users tile does not count the anonymised bucket, so it can read lower than the number of humans whose usage is in the table. That is deliberate — the tile counts people you can act on.
Good to know
Why is the cost different from my invoice?
It is an approximation at API-equivalent pricing, useful for comparing people against each other and not for reconciling a bill. A seat-based or monthly plan charges the same regardless of consumption. Where a tool reports its own cost Velory uses that; otherwise it derives one from model pricing, and the tile says which.
Why is the Telemetry tab empty when License usage has data?
They come from different places. License usage can be filled by a usage report from the vendor, while Telemetry needs the tools on your machines to be reporting. Set that up separately — see How do I set up AI usage telemetry?.
Somebody has a licence but no telemetry row.
They have not used the tool in the selected period, or their machine never got the configuration. Widen the period to All before concluding anything; a quiet fortnight is not an unused seat.
Can I export this?
The licence tables elsewhere in Software export to CSV with the health score included. The telemetry breakdown is a read-only view.
Does a high token count mean somebody is doing more work?
Not reliably. Token counts reflect how a person works — long context, large files, extended thinking — as much as how much they do. Treat the numbers as a prompt for a conversation rather than a productivity measure.
Was this article helpful?