An inbound support draft is a proposed reply to a customer-initiated request. An AI system can summarize the request, retrieve approved facts, and prepare the response, but the draft is not a completed support outcome until it is reviewed or passes a narrowly defined automatic-send policy.
How the workflow works
A new email, chat, or ticket enters the helpdesk. The system identifies intent, customer context, language, sentiment, and relevant help content. It prepares a draft with citations or internal source references. A person reviews it, or a policy allows automatic sending for a tested low-risk intent.
What makes a good draft
The draft answers the actual question early, uses current policy, avoids claims outside the knowledge base, asks only necessary follow-up questions, and records the source. Tone matters, but factual grounding and a clear next step matter more.
Cases that need a person
Escalate refunds, chargebacks, legal threats, safety issues, harassment, account ownership disputes, security incidents, unusual discounts, contractual interpretations, and any answer below the confidence threshold. The system should not hide uncertainty with polished language.
Useful metrics
Measure acceptance without edits, material correction rate, time to first useful response, reopen rate, escalations, and customer effort. A faster draft that causes more follow-up is not an improvement.