AI Voice Agent — L1 informational calls (revenue-aware design)

An AI voice agent that handles L1 informational calls — account balance, product status, "how do I…" — while keeping all revenue-generating voice interactions (upsell, retention, complex troubleshooting) on human agents. The split is the design — not an afterthought. Targets the 14K-voice-calls/month range where 70% of contacts are voice but only a subset is monetized.

Naïve "automate all voice" deployments cannibalize revenue. A payments group had to flag this explicitly: "voice itself is a revenue stream — 85% of merchants prefer voice, especially for upsell". Correct design splits informational L1 (AI) from revenue interactions (human) and routes accordingly. At 14K voice/month with 40% informational, this is roughly mid-six-figures USD-equivalent annually in deflection while protecting upsell.

1. Audit the last 1,000 voice calls — classify each as informational vs revenue-generating vs complex troubleshooting.
2. Pick the 5–8 highest-volume informational categories — these are AI candidates.
3. Stand up the voice agent (Whisper or Deepgram ASR + LLM + TTS) on a separate phone tree branch.
4. Route by intent classification at call open — informational → AI; anything revenue-adjacent → human.
5. Hard rule: AI must escalate on first mention of "cancel", "downgrade", "buy more", "complaint" — these are revenue interactions even if they start as a question.
6. Measure CSAT + revenue retention separately. CSAT shouldn't drop; revenue must not drop.
7. Expand intent coverage only after both metrics are stable for 4+ weeks.

• The "automate voice" trap is real — measure revenue impact, not just deflection rate. Most large players lose 3–8% upsell when they fail to carve out revenue calls. Run a 4-week A/B before rolling out wide.