| Where the draft comes from |
A prompt, an uploaded file, or an unsanctioned paste. The model will still write when it does not know. |
Retrieval from approved knowledge, then a cited draft. Low-confidence questions stop for a person. |
| Can you show it to a buyer or auditor |
You cannot prove which source, which version, or who said it was allowed. Invented citations are a known failure mode. |
Source, owner, and confidence sit on the answer before it is external. |
| Who is allowed to say it |
One chat for the intern and the CISO. No permission model, no expiry, no owner. |
Permission-aware retrieval. Sensitive or unpublished claims do not answer like public ones. |
| When the answer is uncertain |
The model keeps typing. Review happens later, in email, if it happens at all. |
Uncertain or sensitive questions route to the named SME in Slack or Teams. The decision is recorded. |
| Reuse next quarter |
Last week’s thread is not a system of record. Stale and conflicting language repeats. |
Owners, review rules, and an audit trail of what changed. Corrections improve the next response. |
| The actual deliverable |
A block of text in a chat window. |
A completed questionnaire in the buyer’s format, or a sourced answer in Slack / sales prep, with a trail. |
| Deal and call context |
The model does not know the opportunity, the last call, or the CRM fields. |
Tribble Brain can shape the answer with CRM and call context when sales is asking. |
| Data handling |
Pasting an RFP into a consumer chat is its own incident. |
Answers come from systems you already trust, under the review path you already run. |