Skip to content

AI insights

For the helpdesk team

AI insights is an experimental surface that asks a large language model to comment on the helpdesk’s recent activity — “what trends do you see this month?”, “which tickets look like they belong to the same problem?”, “where is the team spending the most time?”. The model returns text suggestions; humans decide whether to act.

  • /ai/insights — the index. A list of generated insights, newest first.
  • Clicking one opens the show page — the full insight body plus the prompt that produced it and the data window it considered.

Two paths:

  • Scheduled — a nightly job runs each insight kind against the previous day’s data. Backlog summary, resolution-time trends, category drift, agent throughput.
  • On-demand — an admin with platform::ai::generate can trigger a fresh generation against the current data. Useful for end-of-week reports.

The LLM provider, model, and prompt for each insight kind are configured in AI settings.

Each insight has:

  • Title — what the model says it found.
  • Body — the model’s commentary, usually 200-400 words.
  • Generated at — when the insight ran.
  • Source window — what date range and what data slices were considered.
  • Prompt — the actual prompt that ran.
  • Model + provider — which LLM produced this.

The prompt and source window let you read the insight critically. “Tickets resolved 23% faster this month” is more interpretable when you know whether the window was 30 days or 7.

  • Aggregate signals“the network category is twice this month vs last” — these are derived from queries the prompt described. Mostly reliable.
  • Causal claims“because the new Wi-Fi rollout” — the model is guessing. Don’t act on these without checking.
  • Specific predictions“will breach SLA at 3pm” — verify against the actual SLA timer for the named ticket; the model can hallucinate.
  • AI insights cost money — every generation hits a paid API. The scheduled job has cost guard-rails; on-demand generation is gated on a permission to prevent runaway spend.
  • Provider failures (rate-limit, model unavailable) show as a banner on the index. The previous insight stays visible; the new generation just didn’t land.
  • Don’t paste insight text into reports without reading it. The model can be wrong, and your name is on the report.