Reviewing results
The Copilot gives you a draft. Your job is to decide whether it is worth acting on — and the interface gives you three signals for that: the answer, its citations, and whether any tool actually ran.
Read the three signals
Section titled “Read the three signals”| Signal | What it tells you |
|---|---|
| The answer | A proposal, in your language, based on the ticket and approved knowledge |
| Citations | Which knowledge articles grounded it — absent citations mean it reasoned alone |
| Activity | Whether the agent read the ticket, searched knowledge, or only talked |
An answer that cites knowledge articles and shows a ticket_read call is grounded in sources you can check. When neither is present, the answer was produced without retrieval or ticket context, so treat it as a suggestion to verify rather than a finding. Generated text is fluent whether or not it is correct; the citations and the tool activity are the signals that distinguish the two.
A four-question check
Section titled “A four-question check”- Does it match the ticket? Wrong symptom means wrong context; say so in the conversation.
- Is it grounded? Cited articles are approved and reviewable; uncited claims are not.
- Would I do this? If the step sounds plausible but you would not run it, do not send it to a requester.
- What would it change? Anything that writes appears as an approval card first — read it as an action, not a suggestion.
The approval card
Section titled “The approval card”When the agent proposes a write, the card names the action and the payload. Three outcomes:
- Approve — the write happens as described, and the audit records you as the decider.
- Reject — the run ends without the write; nothing partially applied.
- Ignore — the run stays Waiting for input; the ticket keeps waiting.
Nothing is written while the card is open. That is the design, not a delay.
Turning a good answer into knowledge
Section titled “Turning a good answer into knowledge”If an answer solved something not covered by documentation, that is a gap worth closing: the fix belongs in the knowledge base, where every technician benefits and the next answer is grounded.
Limits
Section titled “Limits”- Confidence is not correctness. Fluent text can be wrong about your specific ticket.
- Fresh tickets may not be in the index yet; hints improve as ingestion catches up.
- Internal notes are not visible to requester conversations, so an answer there may look thinner than yours.
- Tool output is summarised in the conversation; the full payload is in the audit, not the panel.
Related topics
Section titled “Related topics”AI actions · Overview · Enterprise Audit · Approvals and autonomy