Grounded answers come from your docs, and a tight system prompt tells the bot to decline rather than guess. The model only has access to the content you scope into the prompt, and you can instruct it to escalate when the grounding doesn't directly answer the question. That keeps the bot honest about what it knows.

Every conversation is logged with the query, the grounding context used, the reply, and any escalation signal. Filter by date, by intent if you tag conversations, or by whether the bot escalated. The recurring patterns are usually three or four queries that account for half the volume.

It hands off via email or webhook, which any modern helpdesk can accept as an inbound ticket. The full chat transcript travels with the handoff, so the agent picks up where the bot left off. No special connector is required - just the helpdesk's existing email-to-ticket or API endpoint.

Yes. Display conditions support time-based and day-based rules so the bot only runs during business hours, or runs a different prompt outside hours that sets expectations about response time. Some teams keep the bot 24/7 and only change the handoff destination after-hours.

Most teams ship a working FAQ bot in an afternoon by pointing it at the existing help docs. The longer work is iterating on the system prompt and watching the first hundred conversations to tune scope, voice, and escalation rules. The infrastructure is the easy part.

Logs feed your prompt and your docs, not a fine-tuned model. That's the right way: refining the system prompt and adding missing help articles produces better, more auditable improvements than fine-tuning, which is expensive and tends to embed bad answers permanently.

Encode escalation triggers in the system prompt - any mention of legal action, accusations of charges, or strong negative language hands off immediately. The bot can also lower the threshold for handoff after a couple of unresolved exchanges, so frustration doesn't compound while the customer types the same question twice.

Don't let the bot answer them. Set the scope and the prompt to refuse anything in the regulated category and escalate to a qualified human, with a clear notice that the bot can't provide that kind of advice. Logs let you audit compliance with that rule retroactively.