When a question failed because of the AI provider, such as a workspace's own Anthropic key being rejected, the assistant could show a long block of web page code where its answer belonged. The explanation was never shown.
It now says what went wrong and where to fix it, for example that an admin can update the key in Settings → Assistant. If the reason cannot be read at all, it asks you to try again in a minute instead of showing raw code.
This does not fix a rejected key. An admin still needs to update it.