Product · AI

Drafts with receipts.

Every suggestion is generated from an evidence pack, scored for confidence, and validated before you see it. You approve, edit, or discard — the AI never sends behind your back unless you explicitly raise the floor.

Where suggestions appear

Inbox composer — one click to generate a draft for the latest visitor message; edit inline before sending.
Auto-send (optional) — above a confidence threshold you set (default 80%), drafts can send without review. Off by default.
Visitor deflection (optional) — before an agent picks up, grounded answers resolve common questions automatically; low confidence escalates to a human instead.
MCP — agents and coding tools can generate a suggestion for analysis; MCP output is advisory-only and never sent to visitors.

How a draft is built

Evidence pack — HAL collects the conversation thread, matching knowledge base content, and relevant project facts.
Draft pass — the model writes a reply grounded in that pack.
Validation pass — a second, stricter pass checks each claim against the evidence, adjusts confidence, and flags escalation or clarification needs.

Confidence scoring

The score on every draft is a decision signal, not decoration:

No knowledge base grounding → confidence is capped low; the draft is a starting point, not an answer.
Prompt-injection patterns in the visitor message → confidence is slashed and the attempt is flagged.
Policy-sensitive topics (billing, refunds, security, legal, outages) → extra scrutiny and human review nudges even when grounded.

Improving quality

Feed the knowledge base — grounded drafts are the whole ballgame. Start with your ten most-asked questions.
Let agents edit — every edit is a signal about what the knowledge base was missing.
Turn on auto-send only when the edit rate drops — the threshold exists so you can raise it as trust grows.

EU AI Act posture

Visitors see an AI disclosure in the widget and can opt out of AI answers entirely. HAL never trains models on your data, and human oversight is the default — auto-send and deflection are opt-in. See the EU compliance page.

API

POST /api/conversations/{id}/suggestion/generate
POST /api/suggestions/{id}/accept | /reject | /edit
GET  /api/ai/usage    # metering and quota