Quality loop
Answers improve through review
Test AI answers before changing settings. Use conversation feedback and unanswered searches to improve your knowledge, review drafts, and check that saved content is indexed.
Approved knowledge
AI response quality starts with enabled articles, approved policy wording, and source boundaries operators can inspect.
- Enabled articles
- Source boundaries
- Revision history
Feedback and misses
Thumbs-down answers, unresolved intents, top unanswered searches, and stale-knowledge misses become review inputs for the team.
- AI feedback
- Unanswered searches
- Knowledge gaps
Cost and prompt review
Usage review separates model spend, prompt waste, knowledge gaps, and tool-call risk before pricing or ROI claims are made.
- Usage export
- Prompt waste
- Provider cost
Review signals beat unsupported self-learning claims.
| Capability | Status | Detail |
|---|---|---|
| Multi-turn sandbox | Available | Test follow-up questions using recent conversation context without creating a live visitor conversation. |
| Saved answer tests | Available | Save customer questions and rerun them after knowledge or prompt changes. Expected-phrase and decline checks flag answers for review; up to 50 enabled tests run per batch. Passing a check is a screening aid, not a guarantee of accuracy. |
| Candidate settings review | Available | Compare model or instruction changes with your saved settings, inspect the result, then explicitly load and save a candidate. |
| Knowledge index status | Available | Inspect article coverage and queued indexing jobs. Request a rebuild of saved knowledge when an embedding provider is configured. |
| Reviewable gap drafts | Available | Create or reopen a draft linked to an unanswered search. Drafts stay disabled for AI and the gap stays open until your team explicitly resolves it. |
| Knowledge base articles | Available | Enabled articles can ground AI context while article snapshots preserve what changed over time. |
| Article feedback | Available | Visitor and operator feedback can identify stale, confusing, or missing article coverage. |
| Top unanswered searches | Available | Unanswered help-center searches become content-gap signals for future articles and policy fixes. |
| Knowledge revision packets | Available | Revision packets support human drafting, rollback review, and procurement-friendly change evidence. |
| AI response feedback | Available | Feedback on AI answers helps prioritize prompt, article, and escalation review without claiming verified resolution learning. |
| Notion, Google Docs, YouTube, OCR and Q&A sources | Available | Notion syncs pages shared with your integration and re-syncs weekly, or daily on paid plans; Google Docs, Sheets and Slides import by share link; YouTube imports public captions; scanned PDFs and images are read with OCR when the workspace has an OpenAI key, managed or yours; Q&A pairs pin exact answers. |
| Other ingestion connectors | Not available | Not available: private Google Drive folders and Dropbox. |
| Self-learning automation | Not available | Not automatic: the AI does not retrain itself or roll back prompts on its own. Your team reviews flagged answers and decides what to change. |
The quality story is deliberately human-reviewed: use signals to improve approved knowledge before expanding automation.