← All work
AI & automation / Concept
AI customer support, from question to resolution
A solution concept for voice and chat agents that check orders, follow support policies and hand complex cases to people.
- Designed for
- Concept for commerce and subscription teams
- Field
- AI & automation
- Scope
- Reference architecture: conversation, tools, evaluation and handoff
- Stack
- TypeScriptPythonVoice APIsPostgreSQLpgvectorTemporalOpenTelemetry
The problem
Customers want an answer about their own order, not a link to an FAQ. Support teams need a system that can use live account data, take permitted actions and bring in a person with the full context when the request needs judgment.
Proposed architecture
- Voice and chat entry points sharing conversation context, with clear AI disclosure and customer verification before account access
- Retrieval over approved support content, with source references and an escalation path when the answer is uncertain
- Scoped tools for order lookup, delivery updates and subscription changes, with policy checks before every action
- Human approval for refunds and other sensitive changes, plus handoff with the transcript and actions already attempted
- Replayable evaluations, redacted traces and dashboards for resolution quality, latency and cost per conversation
Goals and validation
- Target: resolve routine requests using live business data, with fewer repeated questions
- Target: give human agents the conversation, relevant records and a clear reason for escalation
- Validation plan: measure resolution quality, escalation accuracy and cost in a supervised pilot before expanding access