Support inbox triage with drafted replies
A shared inbox gets sorted, deduplicated and routed before anyone opens it, with replies drafted and held for approval.
This is a reference design we build from — one of a set that are not delivered client projects, and the figures are estimates rather than measured results.
Trigger
Inbound email, web form, or chat handoff
Timeline
1 wk evaluation set → 4 wks build → 2 wks supervised
Stack
LangGraph or Google ADK agent on Cloud Run · pgvector over the help centre and past tickets · Zendesk / Freshdesk / Outlook Graph API · Evaluation harness on 200 labelled historical tickets · OpenTelemetry tracing
Outcome
The queue is sorted before anyone opens it. Agents edit instead of compose. Escalation is a rule rather than a hope.
The process today
A shared inbox where everyone reads everything, urgent things get buried, and the same twenty answers get retyped weekly.
What the automation does
Classifies intent and urgency, deduplicates threads, and routes to the right queue or owner. Retrieves the relevant help-centre article or prior resolution. Drafts a reply in the team’s voice. Holds every draft for human approval until measured accuracy clears an agreed threshold. Escalates anything below confidence, or containing a complaint or legal keyword, straight to a human.
Where a human stays in the loop
Every drafted reply is held for approval until the evaluation harness shows accuracy has cleared the agreed threshold. Low-confidence and sensitive tickets escalate straight to a human, always.
What you end up owning
The agent code, the evaluation harness and its 200 labelled tickets, the retrieval index, and the tracing setup — all runnable and inspectable in your own environment.