Skip to content
plugboard.appRequest a demo

Running the desk

AI activity

Settings, AI activity (/admin/ai-activity). Requires audit.view. Every model call the school has made, and every suggestion the AI has put in front of somebody, searchable and exportable as CSV.

No prompt is kept, and the call log holds no text. Three tables back this page:

  • Calls: one row per model call, counts only: the feature, tier and model, token counts, whether it succeeded and a short error kind, the duration, and who asked (a person, the system, or an API key).
  • Suggestions: what the AI proposed, a one- or two-sentence reason, the ids of the records it used, and what a person did with it (accepted, dismissed, or left alone). What the AI proposed can be model-written text, such as a reply draft or a one-line reading, kept with real names restored, so this table can hold personal information.
  • Embeddings: computed from ticket, article and document text for finding similar records. No text, only ids, and each is re-read through the reader’s own permissions.

Learning and flows, last 30 days gives the AI calls behind continuous learning as one total, split on a bar into embeddings of tickets, repairs and documents as they are saved, decisions for triage, routing, requests and flow questions, and drafting chat calls from Describe a flow. Under it, the count of flow ideas offered, which are read from the desk’s records with no AI calls. The panel is left out when there is nothing to count.

Two tabs list the records:

  • Model calls: when, the feature with its model and tier under it, who asked (a person, Plugboard, or an API key) and whether it was On request or Background, the Usage in tokens in and out, the record it was about, and Failed with the reason when a call got no answer.
  • Suggestions: when, the feature with its kind, model, confidence and reason, the record, who decided and when, and the status: Shadow, Shown, Accepted, Edited, Dismissed, Applied on auto, Undone or Out of date.

Features carry the names they have in AI settings, such as Ask panel commands and Ask about this record.

Rows are grouped by day, newest first. Display groups them by feature instead, and the page remembers the choice you made. Beside the list, one press filters by outcome or status, by the last day, 7 days or 30 days, or by any feature the log holds; the counts there appear once every matching entry has loaded. The filter row has the same feature, outcome or status and date filters, and a search. Export CSV downloads what the filters match. The page and the export show token counts, never money. The page says how long model calls and suggestions are kept, which follows your retention policy.

Shadow mode keeps each answer and suggest mode asks somebody to accept or dismiss it. This page shows the pattern across the school: which features are firing, how often somebody dismisses what the model proposed, and what technicians did with its suggestions. It is not a per-ticket history.