Answer Review & the Learning Loop¶

Your agent learns from what your team does. This page explains the three Phase-7 surfaces in plain English so an owner can pick them up without reading the code.

The loop in one diagram¶

Visitor question
       │
       ▼
   AI tries
       │
   ┌───┴───┐
   │       │
confident   not
   │       │
   ▼       ▼
Visitor    Pending question
sees       (your team picks it up)
answer        │
              ▼
          Team's answer → Visitor
              │
              ▼
        Becomes a Candidate fact
              │
              ▼
        You approve / edit / merge
              │
              ▼
        Next time, AI answers it confidently

That whole loop has four configurable knobs and three management surfaces.

The knobs¶

Confidence preset (`Configure → Settings → Confidence`)¶

Five positions on a slider. Drives whether your agent answers on its own or routes the question to a teammate:

Setting	What happens
Always answer	Agent replies on every message, even when it isn't confident. Fastest experience for visitors; you accept that some answers may be wrong.
Low / Medium / High	Agent answers when its confidence clears the threshold; otherwise it routes the question to your team via Pending Questions. Medium is the balanced default.
Always review	Every reply waits for a teammate to confirm. Highest quality; slowest visitor experience. Worth it for high-stakes domains (legal, medical, financial).

The agent auto-calibrates the threshold over time once it has ~30 interactions of history, so "Medium" on a brand-new agent and "Medium" on a six-month-old agent both mean roughly the same thing for that agent's specific traffic — they just use different underlying numbers.

Facts vs documents balance (`Configure → AI Brain → Facts vs documents`)¶

A slider that controls how much your agent leans on curated facts (answers your team approved) versus raw documents you uploaded. Default 0.6 slightly favours facts. Disabled until your platform admin enables curated-facts retrieval.

The management surfaces¶

Pending Questions (`Configure → Pending questions`)¶

Questions your agent couldn't confidently answer. Open one, type the right answer, click Submit:

The visitor gets your reply (via the chat, the widget, or — if they verified their email — a notification).
A Candidate fact is automatically created from the question + your answer, ready for you to approve.

Working as a team? A small lock badge appears next to each row when a teammate is editing the answer (your 10-minute claim). You can "Take over" if they walked away — their submit will be blocked.

Facts (`Configure → Facts`)¶

Everything your agent knows that came from documents, the SME pickup loop, or your manual additions. Three statuses:

Candidate — waiting for you to review. Approve, edit, retire, or merge.
Live — actively used in retrieval. The agent quotes from these.
Retired — taken out of retrieval. Kept in history for audit.

A "Possible duplicate" badge appears on Candidates that look very similar to an existing Live fact. Open the row to compare, then either merge them or click "Not a duplicate" to dismiss the flag.

Answer Review (`Configure → Answer Review`)¶

When a teammate sends an AI draft as-is or edits it, you can confirm the change was needed — or flag it if it wasn't. The page shows the AI's draft and your teammate's answer side-by-side. Three verdicts: Change was needed, Draft was fine, Flag for coaching.

On Plus and higher plans there's also a workspace-wide inbox at /answer-review/ — every AI draft across every agent you own, in one place.

The confidence badge¶

When the AI was answering but only by a hair, the chat (and the widget) shows a small italic text under the message:

Medium confidence — the agent's retrieval was just above the abstention line. Worth a sanity-check.
Low confidence — only possible if you're in "Always answer" mode (no abstention). The agent answered, but its retrieval was weak.

The badge is hidden for confident answers, for human-typed replies, and for greetings where no retrieval ran. The goal is honest signal on borderline cases, not metadata noise on every reply.

Who can do what¶

Per-agent roles, granted by the owner via Configure → Team. Adding someone fires an email so they know about the new role and where to find it.

Role	Can do
Owner	Everything. One per agent.
Knowledge Manager	Curate facts (approve / edit / retire / merge). Review SME tweaks. Manage team.
Subject Matter Expert (SME)	Answer pending questions. View facts. Cannot curate.

Need help?¶

📧 Email: support@brilio.ai 📚 Technical Docs: docs.brilio.ai/technical