Documentation
Vanta AI
Vanta AI answers member questions from a knowledge base you control. When confident it replies directly; otherwise it stays quiet or hands off to staff, so it never guesses.
Vanta AI runs on two model tiers, Vanta V1 (fast, default) and Vanta V1 Deep (used for harder, lower-confidence questions).
Knowledge base
Vanta AI answers only from knowledge you give it. Sources include:
- About and custom instructions you write.
- FAQ and document uploads (Markdown, text, PDF, DOCX, HTML).
- Channel pins imported from a channel.
- Closed tickets, learned automatically when enabled.
- Staff answers, learned from helpful staff replies in watched channels.
- Approved replies, learned back when staff approve an AI draft.
- Website crawl of a URL you provide.
How it decides
- If confidence clears your threshold, it ships the answer.
- If not, it escalates: it posts the draft to your review channel to approve, edit, or reject, or points the member to a ticket.
- If nothing relevant is found, it stays silent rather than guess.
You control the confidence threshold, the retrieval floor, a review channel, staff roles, banned topics, and per-channel tone (Premium). You also choose which channels and ticket types Vanta AI watches.
Tools to keep quality high
- AI insights: answer, deflection, gap, and refusal rates over time, plus most-used and unused knowledge.
- Decision explorer: search every AI decision and mark answers good or bad.
- Retrieval playground: see exactly what Vanta AI would retrieve and whether it would answer or escalate.
- Confidence tuner: see how many more answers would ship or escalate before you apply a new threshold.
- Brain health: a daily score with suggestions to improve coverage.
AI usage draws from a monthly credit allowance set by your plan. You can buy top-up packs if you run out. See Billing.