What your organisation actually spent, and on which model

Monitoring

Playable below - no signup

Most AI spend is invisible until the invoice arrives, and by then the interesting question — which model saw which data — is unanswerable. This is the reporting layer: usage by model, assistant, person and tool, with cost attached and an egress trail behind it.

Try it below

Live demo

A week of usage for a small team

Totals, a daily timeline, and breakdowns by model, assistant, person and tool. The row worth looking at is the local model: real traffic, zero provider cost, nothing that left the network.

Synthetic figures, and they reconcile — per-model calls sum to the total and the timeline sums to the same number. Sample data only.

What it breaks usage down by

What you can answer with it

Three questions that are usually guesswork

What is this costing, per model?

Cost tracks tokens at each model’s own rate, so switching an assistant from an expensive model to a cheaper one shows up as a number rather than a feeling.

Who is actually using it?

Per-person and per-assistant breakdowns tell you whether a rollout landed or whether four people are carrying it. That is a adoption question, and it is usually answered with anecdotes.

Where did the data go?

Every integration call is resolved to a country at log time and non-EEA transfers are flagged, so the answer is a report rather than an assurance.

The local-model row is the argument

In the sample data one model carries roughly a quarter of the traffic at zero provider cost, because it runs on the organisation’s own hardware. That is not a demo trick — it is what the reporting looks like once you route the sensitive work to a model on your own network. The saving is real, but the reason to do it is that those prompts never left the building.

Monitoring that is not itself a leak

Usage records carry the shape of activity, not its content: which model, how many tokens, how long, which tool. Guardrail events record the category and the action taken and never the matched text. A monitoring layer that copied prompts into a log would be creating the exposure it is meant to measure, which is why it does not.

FAQ

Worth knowing

Does the log contain our prompts?

No. Usage records hold model, token counts, duration and tool names — the shape of the activity, not its content. Guardrail events record which category was detected and what was done, never the matched text.

Is the cost figure exact?

It is an estimate from token counts at the model’s published rate, shown alongside billed cost where the provider reports one. Treat it as a good working number for comparing assistants, not as an invoice.

Can we export it?

Yes, and this is usually the point — the figures end up in a board pack or a DPIA. Ops metrics can also be pushed to your own observability stack rather than read from a screen.

Who can see it?

Organisation administrators. It is a per-org view by design: an individual’s usage is visible to the people accountable for the deployment, not to their colleagues.

Run it for a fortnight and read the model column

The first thing most teams find is one assistant on an expensive model doing work a cheap one would do just as well.