Train the agent: rules, qualification, handoff, safety

You train the AI Lead Agent from two places. In Leads → Configure, the Training tab links every surface that teaches the agent — business facts, playbooks, your service catalog, and rollout dashboards — while the Agent Rules, Qualification, Handoff, and Safety tabs show the guardrails that always apply. In the Leads inbox, labeling AI drafts Good, Bad, or Edited is itself training: enough high-quality labels unlock the automatic modes per channel.

Before you start

• The AI Lead Agent requires the Phone Enterprise plan.
• Editing business facts and playbooks requires an owner or admin login; labeling drafts works for anyone who can work the inbox.

The three ways the agent learns

The Training tab sums it up: verified business facts and playbooks are injected into every AI prompt, the service catalog limits what the agent may claim you offer, and your feedback labels unlock higher automation modes. Safety rules always win over training.

The tab links each surface:

• Business facts & persona — weekend availability, free-estimate policy, emergency service, service-call fee, the do-not-promise list, and the agent persona. The agent treats these as verified facts and deflects on anything not set. Edited in Settings.
• Playbooks — your rules, by intent — free-text guidance per customer intent (pricing questions, urgency, objections). A matching playbook is injected into the AI prompt and overrides the built-in industry playbook.
• Service catalog — the agent only claims you offer services that exist as active Job Types or catalog service items. Keeping those accurate IS training.
• Label AI drafts (Good / Bad / Edited) — every label is training signal; "50 labeled runs at 80%+ quality unlock auto-qualify for a channel".
• Rollout readiness, Booking shadow review, and Agent quality — dashboards for per-channel promotion progress, simulated bookings, and skip/handoff/completion rates.

Label drafts in the Lead Inbox

Under every AI draft in a conversation you'll see three buttons:

1. Good — the draft was right as written.
2. Bad — the draft was wrong or off-brand.
3. Edited — opens an inline editor pre-filled with the draft; fix the wording and click Save edited. Your edited version is the strongest training signal.

Each draft accepts exactly one label. If someone already labeled it, you'll see "Feedback already recorded." and the buttons disappear.

Shadow Mode — train without customer risk

The Shadow Mode panel on the Sources tab toggles per-channel shadow generation: the AI writes replies but never sends them. They appear in the Shadow Review panel of the Lead Inbox, marked "Not sent to customer", with the same Good/Bad/Edited controls. Shadow data accrues per channel and unlocks the controlled auto-qualify rollout once the channel meets the gate.

What the read-only tabs tell you

Agent Rules lists the enforced conversation rules — "Ask one question at a time.", "Do not quote prices unless verified tenant data exists.", "Do not promise availability or arrival time unless verified." — with the note: "Rules shown here are enforced by the Lead Agent guardrails. Editing is coming later."

Qualification lists what the agent collects: Service type, Service address or ZIP, Urgency, Photo when useful, Contact name, Preferred day/time for non-urgent leads — and the urgent rule: "Urgent, emergency, ASAP, and today requests skip normal scheduling and go to dispatcher review." Custom qualification fields are coming later.

Handoff describes the human-review behavior: "Urgent leads pause AI and require dispatcher review.", "Mark Reviewed records that a dispatcher saw the lead.", "Resume AI clears the handoff and lets AI continue on the next customer message.", and "Close Lead closes the conversation without creating a job." — see take over a conversation.

Safety lists the always-on constraints — "No raw webhook creates a job.", "Bookings require explicit customer confirmation.", "Booking execution uses backend re-checks before creating a job." — with the note: "These controls are always enforced."

Troubleshooting

• Auto qualify or Auto book won't save for a channel — the training gate isn't met yet. The save error shows your labeled-run count and quality score against the requirement. Label more drafts (Shadow Mode helps you build volume safely).
• The feedback buttons are gone under a draft — that draft was already labeled (one label per draft), or the message isn't an AI draft.
• "Feedback already recorded." — another teammate labeled the same draft first. The recorded label stands.
• The Shadow Review panel says "No shadow drafts for this conversation." — Shadow Mode is off for that channel, or the AI has not evaluated the conversation yet. Turn Shadow Mode on from the Sources tab and wait for the next inbound message.