Resolve card, fraud, and account issues on script

No, and that's a hard rule, not a polite preference. The agent's instructions refuse to recommend a product, interpret a customer's circumstances, or comment on whether something is a good idea. It explains what a product costs and who qualifies against your published criteria, then hands to a human the moment the conversation turns into advice, with the full transcript attached.

Every action that touches an account is identity-bound. Freezing a card, opening a dispute, or reading a transaction only fires after your verification step passes; an unverified session gets general policy answers and nothing else. You define the checks, and which cases your policy says a person must handle.

Yes, through Custom Actions: point one at any REST endpoint your core, card processor, or KYC provider exposes. The agent fills variables from the conversation, sends the request, and parses the response. Attach a Widget when the customer needs to pick a transaction or confirm a card action inline in chat.

Auto-Retrain refreshes website crawls, Notion sources, and remote storage every 24 hours on Standard and Pro. For terms and fee schedules you upload as PDFs, retrain after you publish the new version, usually whenever the document changes.

Yes. Each conversation is stored with the sources the answer was drawn from, the actions that fired, and the timestamps, in the Activity feed. Escalations carry the same record into your helpdesk, so a reviewer can see exactly what the customer was told and why.

One agent handles the languages the model supports, auto-detected per conversation and answered in the same language on every channel. Lock the set in the agent config if you need replies scoped to the languages you're licensed to service.