Explain panel

The Explain panel answers the question every governance review eventually asks: what actually happened when Zoe produced this reply? It shows the recorded trail for one message — which systems were touched, which documents were cited, which guardrails applied, and who approved what.

The panel is for administrators of your organization. If you are a member rather than an administrator, the Explain control does not appear; ask an administrator to review a reply with you.

How to open it

1. Find the reply you want to examine in the conversation.
2. Select the Explain chip below it. The chip appears only once a reply has finished — it is not there while Zoe is still writing.
3. The side panel opens with the heading Why this answer.

Use Refresh to reload the panel — worth doing after you approve a pending action, since the approval becomes part of the trail. Close dismisses it. Open full replay takes you to the whole conversation's audit replay in the Customer Portal.

What the panel shows

At the top, a badge states the recording status: Recorded on the audit trail, Recorded — some content has been erased where a redaction has been applied, or Nothing recorded for this reply. Conversations from before the audit trail existed are marked as such rather than reconstructed.

Below that, the events are grouped into sections. A section only appears when it has something in it:

• Trigger — the user turn that produced this reply.
• Model calls — each call, with the channel, step number, tokens in and out, latency, and an expandable list of the tools that were offered to the model at that step.
• Tools — every tool call, with the connector it belongs to, whether it was read-only, whether it succeeded, and how long it took. A tool call that proposed a change carries a drafted chip.
• Sources — the document passages cited in this reply, with file, page and a link where available.
• Skills — the procedures Zoe loaded for this turn.
• Delegations — named agents Zoe consulted, and the questions it asked them.
• Memory — anything Zoe proposed to remember, and whether it was saved.
• Guardrails — policies that fired, with severity, direction, and whether output was masked.
• Approvals — the decision on any governed change, who decided it, and how.

Metadata is open, content is revealed

Field-level metadata is visible as soon as the panel opens. The content behind an event — the system prompt, a tool's arguments and result, the text of a delegated question, the proposed change — sits behind a Show control, with the notice Revealing content is recorded in the audit trail.

That is literal: each reveal is itself written to the trail with your identity. This is by design, so a record of who looked at what is as durable as the record of what Zoe did. Reveal what your review needs, not everything on the page.

When to use it

Reach for the Explain panel when:

• A reply looks wrong and you need to know whether the cause was a stale source, a system that returned bad data, or the model.
• Someone asks which systems Zoe touched on a particular request.
• You are reviewing an approval after the fact — what was drafted, who approved it, and whether it executed.
• You are checking that guardrails behaved as configured on a sensitive conversation.
• You are assembling evidence for an audit and want the per-message detail before opening the full replay.

Limits worth knowing

Replies produced during a voice call show a notice instead of per-turn detail: voice tool calls are recorded against the conversation as a whole, so open the full replay for those. If the panel reports that the audit ledger is not ready for your organization yet, the trail is still being prepared — try again shortly. Admin only means the account you are signed in with is not an administrator of this organization.

For the approval flow this panel documents after the fact, see Records and approvals.