Skip to content

AI triage & reply copilot

For the helpdesk team

Two AI features sit on the ticket worktop. Both depend on AI being enabled for your workspace and respect the monthly usage limit; both strip personal details before anything reaches the AI provider.

When a ticket is raised, Keystone classifies it in the background — predicting ticket type, department, priority and sentiment against your trust’s actual types and departments, with a confidence score and a one-line rationale.

What happens next depends on confidence:

  • High confidence (the workspace threshold, 0.85 by default) — suggestions are auto-applied, but only to fields a human hasn’t already set. It fills blanks, never overwrites. If a requester picked a ticket type via the catalogue, the AI leaves it alone.
  • Lower confidence — nothing changes automatically; the suggestion waits on the worktop for you.

On the ticket you’ll see the suggestion with two buttons:

  • Accept — applies the suggested type, department and priority to the ticket.
  • Dismiss — leaves the ticket untouched, but records that you disagreed. That’s a feedback signal that feeds accuracy reporting, so dismissing a wrong call is worth doing rather than ignoring.

Triage runs best-effort: if the AI has a bad day the ticket is unaffected — it just won’t carry a suggestion.

When you’re composing a reply, press AI assist. The copilot returns three things to lift from:

  • Up to three knowledge-base articles relevant to the ticket.
  • Up to three similar resolved tickets (matched on meaning, not just keywords). You see their subjects and can click through — the copilot never exposes raw body text.
  • An optional drafted reply, grounded in the articles and similar tickets above.

The draft is a starting point — paste it in, then edit. It’s written to be a friendly first-line reply and is told to say so plainly when it can’t ground an answer rather than invent one. If AI is disabled or over quota, the article and similar-ticket lists still appear; only the drafted reply is skipped.

Both features run the ticket subject and description through redaction first — staff names, emails and similar identifiers become anonymous tokens before any AI call. The structured context the model sees carries aggregates and definitions, not personal data.