Voice Agent Go-Live Checklist

Pre-deployment checklist for a production voice AI agent (Vapi / ElevenLabs Conv AI / Retell / Synthflow). Used at PxlPeak for every SMB voice deployment. 40 items across 6 categories.

Skip any item at your own risk. Every one of these represents a real incident we’ve seen.

---

1. Intent & scope

• Written “what this agent does / does not do” one-pager approved by the business owner
• Top 10 expected user intents documented with example phrasings
• Escalation path defined: when does the agent transfer to a human? What numbers? What hours?
• “I don’t know” behavior specified — how does the agent handle out-of-scope questions?
• HIPAA / PCI / other compliance scope confirmed in writing with the client before going live

2. Prompt & behavior

• System prompt version-controlled in a git repo (not just pasted into the vendor console)
• Business hours logic handled explicitly — the agent should not schedule Sundays for a Monday-Friday business
• Timezone handling is explicit and tested (naive clients assume America/New_York)
• Agent identity is consistent: same name every time, same voice, same tone
• Profanity and hostile-user handling tested with at least 5 adversarial prompts
• Hallucination guardrails: if the agent doesn’t know a business detail, it says so and offers to connect a human

3. Integrations

• Booking system webhook end-to-end tested with real bookings (not mocks)
• CRM write verified — new leads actually land in the CRM, not just in a log file
• SMS follow-up tested on a real phone (twilio sandbox ≠ production)
• Payment collection (if applicable) runs in a sandbox environment for go-live week one
• Failure path for every integration: what happens if the CRM API is down mid-call?

4. Telephony & voice

• Phone number provisioned and verified to receive calls (test from 3 different carriers)
• Outbound caller ID shows the business name, not “Unknown” or the vendor default
• Voice latency tested below 1.5 seconds (anything above 2s feels broken)
• Voice and accent match the business’s customer base — a Southern plumber should not have a British accent
• Call recording disclosure enabled where legally required (CA, FL, IL, MD, MA, MT, NH, PA, WA at minimum)
• Barge-in (user interrupting the agent) works cleanly
• Silence timeout is tuned — 8s default is often too long for impatient callers

5. Observability

• Every call logged with transcript, duration, outcome, and cost
• Failure alerts wired to Slack or SMS — someone gets paged if the agent is down for 5 min
• Weekly review dashboard in place (call volume, resolution rate, escalation rate, cost/call)
• Client has read-only access to their own agent’s dashboard
• Cost ceiling alerts set (at PxlPeak, typically $X/day with auto-page)

6. Human handoff & trust

• A human answers when the agent can’t — voicemail is not an acceptable fallback
• The client can listen to 5 random call recordings per week without asking us
• A “kill switch” exists: the client can disable the agent and route all calls to a human in under 60 seconds
• First 7 days of production, a human reviews every transcript (not just sampled)
• Week-two review with the client: what worked, what didn’t, what’s getting edited

---

Anti-patterns we no longer tolerate

• “We’ll monitor it manually for now.” Not a real observability plan. Set up alerts before go-live.
• “Let’s launch and iterate.” Fine in principle. Not fine when the client’s main business line is the phone.
• “The vendor dashboard is enough.” The vendor dashboard doesn’t tell you if bookings are actually landing in the CRM.
• “We’ll train on call data over time.” You can train on data. You cannot train on customer trust that evaporated in week one.

---

Want this done for you? PxlPeak ships production voice agents for SMBs in 5 business days. Fully owned by the client, no lock-in, $997/mo starting price. Book a 15-min intro.

---

Related

• n8n Webhook Security — Production Patterns — production webhook hardening for the automation layer behind voice agents
• RAG Chatbot Quality Bar — the chatbot counterpart to this checklist
• AI Agent ROI Calculation — how to model whether a voice agent is worth deploying in the first place