Read the Admin dashboard¶
Purpose: see whether AI traffic is healthy, and where to look when it isn't, without instrumenting a single app.
Before you start¶
Organisation admins, group admins (their teams only) and organisation viewers (figures without links) have the Admin view. Read your dashboard explains the filters.
Steps¶
- Open Dashboard and choose Admin in Show.
- Choose the Period and what you're Counting.
- Read the figures across the top:
- Calls, with their trend and the share that failed (provider or gateway errors; refusals by policy aren't failures), with See the calls;
- Response time (95th percentile), served calls only, Estimated from the gateway's timing (lower is better);
- Calls without a price: calls with no price yet, never counted as free;
- Calls rescued: a fallback answered after the first choice failed (See fallbacks);
- Outages absorbed: times a provider was down and calls went elsewhere (See provider health);
- Routing rules in force now (See routing);
- Needs attention: what to act on now, as on Home (Open Home).
- Below, in this order: the Inventory, then Outages, then Calls by team over time, Failures and refusals, Failures and refusals by error, Busiest models and Tokens by model.

Failures and refusals¶
A failure is an error from a provider or the gateway. A refusal is a call a policy stopped on purpose: access, routing, a guardrail, a budget or Stop AI traffic. Failures and refusals by error ranks the period's failures by what went wrong (a rate limit, a provider that couldn't be reached, a key not accepted…) and its refusals by why the gateway refused them (Model not allowed, Blocked by a guardrail, Budget reached, AI traffic stopped, and so on), in words, the top five and the rest.
Outages¶
One line per time a model's provider stopped answering in the period: the Provider and model, Down from, Back at (or Still down), Down for and Its calls: how many a fallback chain moved to another model and how many failed while it was down, for example "1,240 moved to another model, 7 failed" (a refusal by a policy isn't a failure). Outages are listed for the whole organisation, whatever the board is narrowed to. No outages means no provider was down in the period.
In the Grachtwerk demo world, the France Central deployment is in a long capacity incident. Maison Ardelle's calls that ask for it are answered from Paris or West Europe through a fallback chain, never outside the EU, so the team keeps working while the outage is listed here.

The inventory¶
People, agents, applications, teams, environments and live keys; active people, keys used, providers and models in use in the period; compliance policies, guardrail policies, region rules, routing rules, budgets and approved destinations; model aliases, and keys that are disabled or expired. Each count opens its list.
Check it worked¶
The calls figure matches Request logs filtered to the same period and scope.