The question.
Support automation often answers a question without resolving the underlying case, or modifies records without enough context.
The design gives each support action a typed contract and preserves evidence, entitlement, and escalation state across channels.
The proposed approach.
- Resolve policy and product questions against maintained sources
- Use scoped tools for case lookup and eligible account actions
- Confirm consequential changes and make retries idempotent
- Escalate with a concise, source-backed case summary
Target architecture.
Voice / text requestIdentity and case contextKnowledge retrievalService toolsResolution or escalation
The difficult parts.
- Ambiguous requests and fragmented customer context
- Knowledge freshness and conflicting policies
- Repeated calls, duplicate tickets, and failed handoffs
How to evaluate it.
- Case resolution and correct escalation rates
- Policy-grounded answers on representative scenarios
- Tool-action correctness, latency, and caller interruption handling
Planned deliverables.
- A cross-channel case-state model
- A source-grounded service workflow
- A conversation and action-quality evaluation suite
Design principles.
Measure a resolved case, not a contained conversation.
A human should inherit context rather than start over.