Skip to main content

AI usage

The AI usage view of Diagnostics shows what the AI was used for on the instance, by whom, and what it cost in tokens. It covers the three ways the AI is reached: the in-app assistant, the auto-map of the mapping editor and Claude connected through MCP.

Open it from Diagnostics → AI usage. The view is granted per role, as AI usage view (Roles).


At a glance​

DiagnosticsErrors, failures and logs of the instance.Integration errorsFindings by rule, by event …Processing logEvery run with its files a …FailuresEvery unresolved failure i …Server logBy severity and moduleAI usageThe assistant and Claude: …7 days30 days90 daysWho, model, tokens and tools — never the question or the answer. Kept 90 days.Turns286Users11Tokens in · 68 % cached812.6kTokens out121.3kTool calls · 9 failed742Average turn7.8 sTurnsBy userBy dayBy channelBy modelBy toolWHENUSERCHANNELPAGEMODELTOKENS INTOOLSOUTCOME2026-11-19 10:42aliceIn-app assistanterrorsclaude-sonnet-5-5403827Answered2026-11-19 10:37brunoClaude (MCP)01Answered2026-11-19 10:20chloeMapping auto-mapclaude-sonnet-5-5582040Cut short2026-11-19 09:58aliceIn-app assistantinvoicesclaude-sonnet-5-5319025FailedPeriod7, 30 or 90 daysBreakdownturns, user, day, channel…A click on a turnopens its tool callsAI usage viewgranted per roleThe period in figuresturns, tokens, tools, timeTurns per dayover the periodThree channelsassistant, auto-map, Claude

What is kept​

Each turn leaves one record: who, when, the channel, the page the question was asked from, the model, the tokens (read, written, read from and written to the cache), the time it took and how it ended. Each tool called during the turn is kept with its name, a summary of its input, its time, the size of its answer, whether it failed and why, and whether it saved a change.

The question, the answer and what a tool returned are never kept.

ChannelOne turn is
In-app assistantA question asked in the assistant of the application.
Mapping auto-mapAn auto-map run from the mapping editor.
Claude (MCP)A tool called by a user's Claude client. No model and no tokens: they are those of the user's Claude plan.

The period in figures​

The period is 7 days, 30 days (default) or 90 days.

TileShows
TurnsTurns over the period.
UsersUsers with at least one turn.
Tokens inEverything the model read, with the share read from the cache (68 % cached), billed at a fraction of the price.
Tokens outWhat the model wrote.
Tool callsTools called, with the number that failed.
Average turnAverage time of a turn.

Below the tiles, a bar chart gives the turns per day.


Breakdowns​

Breakdown switches the list:

ChoiceRows and columns
TurnsThe latest turns: When, User, Channel, Page, Model, Tokens in, Cached %, Tokens out, Tool calls, Tool failures, Time (s), Outcome, Tools.
By user, By day, By channel, By modelTurns, Tokens in, Cached %, Tokens out, Tool calls, Tool failures, Failed turns.
By toolCalls, Failed, Changes, Average time (ms), Average answer (characters).

The list sorts, filters and exports like the other lists of the application.


One turn​

A click on a turn opens it: its outcome — Answered, Failed, Stopped by the user or Cut short (the answer reached its length limit) —, its time, its model and tokens, the page it was asked from, then one line per tool with the summary of its input, its time and the size of its answer. A tool that saved something is marked saved a change; a failed one shows its reason in red.


How long it is kept​

90 days by default. AI usage kept (days), in Settings → Global → AI, sets another duration.


Tips & best practices​

  • Read By tool after opening MCP. The Changes column says which tools saved something, and Failed points at a missing grant or a question the tools cannot answer.
  • A low cached share on long conversations means a prompt that changes at every turn: check the System Prompt and the custom playbooks.
  • Grant the view to those who follow the cost, not only to administrators: it shows no question and no data of the instance.