# AI trust & guardrails

> Stop the agent from confidently making things up. The low-confidence guardrail makes it disclaim, escalate, or capture a lead when its knowledge does not support a question — and the Knowledge Gaps report shows every unsure answer by topic so quality never silently degrades.

A support bot that confidently invents answers is worse than no bot — it gives customers wrong information and you find out from lost sales. Bookbag's AI-trust features make sure the agent only answers when it actually knows, and show you exactly where its knowledge is thin.

## The low-confidence guardrail

Every answer is scored for **grounding** — how well your knowledge base supports the question. Turn on the guardrail in an agent's **AI** settings and, when grounding falls below your threshold, the agent will not answer from the model. Instead it does what you choose:

| Behavior | What the visitor sees |
| --- | --- |
| Answer honestly that it is unsure | The agent says it does not have enough information, instead of guessing. |
| Hand off to a human | A polite line plus a handoff to your team. |
| Capture a lead | The agent asks for the visitor's details so your team can follow up. |

You can set the **confidence threshold** (higher = more cautious) and a **custom message**. Trusted Q&A matches always answer — the guardrail only intervenes when grounding is genuinely weak. A guardrailed reply never calls the model, so it never costs a credit.

## The Knowledge Gaps report

Find the report on an agent's **Suggestions** tab. It surfaces every answer the agent was not confident about — low-confidence replies, answers with no sources, and thumbs-down answers — **grouped by topic**, so you can fix what is silently costing you.

- Headline numbers: answers analyzed, low-confidence answers, and your overall **gap rate**.
- Each topic shows a count, downvotes, and average confidence — expand it to read the actual questions visitors asked.
- A threshold selector (strict / balanced / cautious) controls how aggressively gaps are flagged.

> **CLOSE THE LOOP:** Add a source or Q&A pair that answers a high-frequency gap, retrain, and watch the topic drop off the report. This is the fastest way to push deflection up without guesswork.