Skip to main content
When an answer looks wrong, the useful question is not “why is the AI bad” but “what did it actually have in front of it”. The explainability panel answers that for a single turn.

Opening it

In a conversation, expand Why this answer on any agent message. Nothing is fetched until you open it. The panel does not slow down loading a conversation, and a workspace that never uses it pays nothing for it.

What it shows

Which specialist answered

Whether the turn went to qualification, knowledge, booking or handoff. A knowledge question that got routed to booking explains a lot of otherwise baffling answers.

What was retrieved

The knowledge entries the agent found, with how closely each matched, each one linking to the entry itself. This is where you see that the right answer exists but a worse entry outranked it.

Model output versus delivered

Where a guardrail rewrote the reply, both versions are shown along with the reason. This is how you tell a bad answer from a good answer that got blocked.

Timing

When the turn ran.

Honest gaps

The panel says what it does not know rather than filling in plausibly:
  • Retrieval that was never recorded stays unreported. The panel shows what the turn actually persisted, and does not re-run the search to reconstruct it, because a re-run answers a different question than the one the customer asked.
  • A turn with no knowledge lookup says so, rather than showing an empty list that looks like a failed search.
  • Where the specialist is inferred from timing rather than a hard link, the panel says so on the row.

Feedback

Each answer carries a thumbs up and thumbs down. A thumbs up records immediately. A thumbs down asks which of five things went wrong: You can add free text. This feedback feeds quality reporting and the knowledge gap miner, so a thumbs down on an unanswerable question turns into a ranked gap rather than disappearing.

Using it well

  • Before rewriting a knowledge entry, check what was actually retrieved. Often the entry is correct and something else is winning the match.
  • Before loosening a guardrail, check whether the customer saw the model’s answer or a replacement. A workspace that thinks the agent is unhelpful is sometimes a workspace with an aggressive block action.
  • Before reporting a routing problem, check which specialist took the turn.