AI customer support workflow

Support teams spend most of their time answering the same routine questions, while genuinely difficult cases wait in the same queue.

A support workflow design where AI answers routine requests from a knowledge base, checks its own confidence before responding, escalates uncertain or sensitive cases to a person, confirms satisfaction, and raises priority when a customer isn't satisfied.

A worked design for human-in-the-loop support: routing rules, confidence thresholds, escalation of uncertain or sensitive cases, and satisfaction follow-up. This is a designed workflow, not a deployed product.

• Problem definition
• Workflow design
• Architecture

1. 01An inbound request is classified by intent and sensitivity before any answer is drafted.
2. 02Routine requests are answered from the knowledge base. The response is only sent when the confidence check passes.
3. 03Uncertain or sensitive cases are escalated to a person with the request, the retrieved context, and the reason for escalation attached.
4. 04After a resolution, the workflow confirms satisfaction. A dissatisfied response raises priority and re-queues the case to a person.
5. 05Outcomes are recorded so routing accuracy can be reviewed over time rather than assumed.

Orchestration across Zapier and n8n, with knowledge base lookup through Glean and voice intake through VAPI.

Confidence gating is a distinct step with its own threshold, separate from the answer generation step.

Escalation carries context forward so the human agent does not restart the conversation.

Priority is a computed field driven by satisfaction and wait time, not a manual flag.

• This is a design, not a running system. No production traffic has passed through it.
• Thresholds and routing rules are reasoned choices, not values calibrated against real ticket data.
• Knowledge base quality determines the ceiling. A thin or stale knowledge base makes the automated path useless.
• The design assumes a person is available to receive escalations, which is a staffing commitment, not a technical one.