Context
Most conversational AI deflects. The valuable version completes: it can look up the caller, answer from their real record, and execute the follow-up in the operational system.
The problem
Conversations span channels — a request starts on a call and finishes over SMS — while state has to stay coherent.
Answering from account data means live integration with systems of record, with the access-control questions that follow.
Voice adds hard latency budgets: retrieval, reasoning, and synthesis must fit inside a natural pause.
The system
Critical decisions
Agent actions are typed operations, not free-form tool calls
The agent chooses from an explicit, auditable action catalog with per-action authorization — capability boundaries are architecture, not prompt text.
Escalation is a feature, not a failure
Confidence thresholds route to humans with full conversation context. The metric is resolved work, not deflection rate.
Latency budget drives the retrieval design
Pre-fetched caller context and cached account summaries keep voice interactions inside conversational timing.
Execution
Channel by channel: SMS first (forgiving latency), then voice. Every production conversation reviewed in early weeks to build the evaluation and escalation tuning set.
Outcome
- Customer conversations that end in completed work inside operational systems.
- Human agents concentrated on the interactions that genuinely need them.
- [Quantified outcomes to confirm before publication]
Lessons
- The integration layer, not the conversation layer, is where the value lives.
- Typed actions with authorization beat open-ended agency in production.
- Latency is a product constraint that shapes architecture.
Have a similar problem?
Discuss an AI initiative