The most expensive mistake in voice agent projects is going live before you are ready. A software bug gets fixed, but a customer who had a bad conversation may not come back — and will tell others about it.
The fifteen items below are what to complete before going live. Do not start until every box is ticked.
Scope and purpose
1. The agent's purpose is written in one sentence. If answering "what is this agent for" takes more than a sentence, the scope is too broad.
2. Out-of-scope topics are listed. What the agent will not discuss, decide or commit to is written explicitly. For health, legal and financial content that list must be especially clear.
3. A success criterion is defined. Which number will tell you this agent is working? Define it before going live, not after.
Content
4. The knowledge base is current and verified. Prices, hours, campaign terms and delivery times checked against their source. Stale information is more insidious than wrong information, because it was once true.
5. The top five questions are answered flawlessly. The majority of calls cluster there. Answer five perfectly rather than a hundred adequately.
6. Behavior on unknown topics is defined. What does the agent do when it does not know? The rule "do not guess, hand off" must be explicit in the prompt.
Conversation design
7. The greeting is short and complete. Identity statement, recording notice, open question. Under forty words.
8. Answer length is capped. Long answers are torture on a voice channel. The word limit belongs in the prompt.
9. Call closing is defined. When and how does the agent end a call?
Handoff
10. Handoff thresholds are written as rules. "Hand off if necessary" is not enough; concrete thresholds must be defined.
11. Handed-off calls land somewhere. Is there a counterpart on the team? Who owns the queue, and within what time is it answered? An unanswered handoff is worse than none.
Compliance
12. Recording, retention and access are set up. The notice is given, the retention period is defined, automatic deletion works, access is role-limited.
13. The privacy notice is updated. Voice calls and recording are covered. If outbound calls will be made, the permission framework is defined.
Testing and rollback
14. The eight test scenarios have passed. Normal path, interruption, silence, out-of-scope question, handoff insistence, noisy environment, language switch, and post-call output check. We covered these in detail in the 6 August post.
15. A rollback plan is ready. What happens if something goes wrong? The path to switch the agent off and route calls to the team must be defined and tested. Teams that skip this item make decisions in a panic when trouble arrives.
Rollout order
Even after completing the list, do not switch everything on at once. An order that works:
Week 1 — shadow mode. The agent runs but does not reach customers; alongside calls the team handles, it records what it would have answered. That week of data is the fastest way to expose knowledge base gaps.
Week 2 — limited live. After-hours calls only, or a single topic only. Low risk, measurable volume.
Week 3 — expansion. If results are good, scope opens up.
Week 4 — full scope. With measurement in place.
This four-week transition builds the team's confidence and caps the cost of a bad first impression.
What to do in week one
The first week after going live is the project's most critical period:
- Listen to at least ten call recordings every day.
- Assign every problematic moment to one of three categories: knowledge base, prompt, rule.
- Fix accuracy violations the same day; batch the rest weekly.
- Track the handoff rate daily; high in the first days is normal, but if it does not fall, revisit the thresholds.
- If a customer complains, find and listen to the call. The source of a complaint is usually different from what was assumed.
That loop raises quality more than anything else in the first month, and it requires no technical investment.
If you would like to test the agent with your own ears, you can start from the demo page.