3 min read
What to verify before an AI voice agent takes a real call
A voice agent that fails badly is worse than no agent at all, because the caller experiences being dismissed rather than merely being missed. These are the checks worth making first.
Quick answer
What should you check before launching an AI voice agent?
Before go-live, confirm the agent's scope is narrow and explicit, disclosure meets your jurisdiction's rules, there is a working escalation path to a human, calendar availability is genuinely live, every outcome writes to the CRM, and the agent has been tested against real recordings rather than scripted ones.
The gap between an impressive demo and a deployment you can leave running is almost entirely made of the items below. None of them are about the model.
The asymmetry that should shape every decision
A missed call is neutral. The caller knows they were not answered and tries again or goes elsewhere.
An agent that answers and handles the call badly is negative. The caller experiences being dismissed by a company that could not be bothered to staff its phone, which is a worse outcome than the voicemail it replaced.
That asymmetry is the argument for narrow scope, aggressive escalation and a working kill switch. It is cheaper to hand over too often than to answer wrongly once.
What a good agent sounds like
Short turns. One question at a time. No narration of what it is about to do. A natural holding phrase when it genuinely needs a moment, rather than silence.
Most of that is prompt and configuration work rather than model selection, and most of it is audible in the first minute of a pilot call. Related reading: why latency decides the call.
The steps, in order
- Step
Define what the agent will refuse to do
Scope is defined by exclusions, not inclusions. Write down the situations where the agent must hand over rather than attempt an answer — complaints, medical or legal specifics, pricing outside a defined range, anything emotionally charged. An agent with no refusal list will confidently answer things it should not.
- Step
Settle disclosure before launch
Several jurisdictions require callers to be told they are speaking with an automated system, and the FCC has confirmed AI-generated voices fall under existing artificial-voice restrictions. Decide the exact wording and where in the call it appears. Confirm your own obligations with your advisor rather than copying someone else's script.
- Step
Build and test the escalation path
Every agent needs a route to a human, and it must work at the times the agent is actually running. An escalation that rings an unstaffed desk at 2am is not an escalation. Decide what happens when nobody answers the transfer — voicemail with context is acceptable, silence is not.
- Step
Verify calendar availability is genuinely live
Book a real appointment through the agent, then check it appears on the right calendar with the right owner and duration. Then block that slot manually and confirm the agent stops offering it. Stale availability produces double bookings, which cost more trust than a missed call.
- Step
Confirm every outcome writes to the CRM
Booked, not interested, wrong number, callback requested, escalated. Each needs to land as structured data on the contact record, not only in a transcript. Without this you cannot calculate cost per booked appointment, which is the only number that tells you whether the agent is working.
- Step
Test against real call recordings
Scripted tests pass. Real callers interrupt, mumble, give addresses in fragments and change their mind mid-sentence. Replay actual recordings from your own phone system through the agent and watch where it breaks. This is the step that separates a demo from a deployment.
- Step
Tune endpointing against your own callers
Endpointing decides when the agent believes the caller has stopped speaking, and defaults are rarely right for a specific market. Too eager and it interrupts; too patient and every turn feels slow. Adjust it against the recordings from the previous step.
- Step
Set volume and retry limits before going live
For outbound, define maximum attempts per contact, the interval between them, and the calling-hours window. For inbound, define concurrency limits. These are the settings that turn a configuration error into a small problem rather than a large one.
- Step
Run a limited pilot
Route a fraction of calls, or run it for a defined window, before handing it the whole line. Listen to every call in the pilot. Whatever the transcripts say, listening reveals things reading does not.
- Step
Define the kill switch and who can pull it
One person, one action, no deployment required. Decide in advance what the line falls back to when the agent is disabled. Every automated system needs an off switch that someone non-technical can reach.
Questions
Related questions
How long should a voice agent pilot run?
Long enough to cover the range of calls you actually receive, which for most businesses means at least two weeks including a weekend. Volume matters less than variety — a pilot that only sees routine weekday enquiries has not tested the cases most likely to fail.
Should the agent say it is an AI?
In several jurisdictions you are required to disclose it, and the FCC has confirmed that AI-generated voices in calls fall within existing artificial-voice rules. Beyond compliance, disclosure tends to reduce complaints rather than increase them, because callers who suspect it and are not told react considerably worse. Confirm your obligations with your advisor.
What is the most common launch failure?
Scope that is too broad. An agent asked to handle every possible call will answer questions it should have escalated, confidently and incorrectly. Narrow scope with a clean handover outperforms broad scope with a weak one on nearly every measure that matters.
How do you know if the agent is working?
Track cost per booked appointment rather than cost per minute, along with the transfer rate and how far into calls people hang up. A rising hang-up point usually indicates a latency or scope problem, and both show up in that number well before anyone complains.
Services mentioned here
AI voice agents that hold a real conversation
An AI voice agent is software that conducts a spoken phone conversation in natural language. It understands what a caller says, responds in real time, follows the rules it was given, and takes actions such as booking an appointment, creating a CRM record or transferring to a human when the conversation requires one.
An AI receptionist that answers every call, at any hour
An AI receptionist answers inbound calls around the clock, identifies why the caller is ringing, answers routine questions such as opening hours, pricing ranges and location, captures caller details into the CRM, books appointments against live calendar availability, and transfers urgent or complex calls to a human.
Vapi voice agents wired into your CRM
Vapi is a platform for building programmable AI voice agents. It handles speech recognition, language model orchestration, speech synthesis and telephony, and exposes function calling so an agent can read and write external systems — checking calendars, creating CRM records or booking appointments during the call.