How to Test a Voice AI Agent Before Go-Live: The QA Checklist I Run on Every Build (2026)

TL;DR: Testing a voice agent in the platform's chat simulator proves almost nothing. The failures that lose clients happen on a real phone line: a caller who mumbles a zip code, a name the agent spells wrong into the CRM, a calendar that times out, a transfer that rings an empty office at 7pm. My pre-launch pass is five rounds of real phone calls, not one: the happy path, the messy paths, data capture accuracy, escalation, and the integrations behind the call. Then a soft launch where I listen to the first twenty live calls before anyone celebrates. I build production voice agents for US clients on Retell, n8n, GoHighLevel, and Twilio, and I run VoiceDash, the white-label client portal agencies hand to their clients. Here is the checklist I go through before an agent answers a real customer.

Nearly every bad launch I have seen came from the same shortcut. The builder typed a few messages into the platform's test chat, the agent responded beautifully, and it went live on a business phone number that afternoon. Two days later the client forwards a recording where the agent booked "Jon" as "John Smith Junior" at a time the calendar had already given away.

The simulator is not the product

Text testing checks one thing: whether your prompt logic holds together. That is worth doing first, and it is worth doing fast, because it is free.

It cannot check anything that only exists on a phone call. It has no turn detection, so it never tells you the agent interrupts people. It has no telephony, so you never learn that the voice sounds fine in the preview and thin over a cell connection. It has no background noise, no dogs, no drive-through speakers, no toddlers.

So the rule I hold every build to: no agent goes live until I have personally called it from a real phone, at least twenty times, in the conditions its callers will actually use.

Round one: the happy path, out loud

Call the number. Be the ideal caller. Say exactly what a cooperative customer would say, in order, with no curveballs.

What I listen for, in this order:

• Does the greeting sound like this business? Right name, right pronunciation, right energy. A dental office and a plumbing company should not sound identical.
• Does it collect in the right order? The reason for the call before the phone number, always. People hang up on forms.
• Does it confirm before committing? Read back the appointment, the spelling, the callback number.
• Does it close cleanly? A specific next step, not "is there anything else I can help you with" on a loop.

If the happy path is not clean, stop. There is no point testing edge cases against a broken center.

Round two: the calls that actually break agents

This is the round most builders skip, and it is the one that decides whether the client is happy in month two. I keep a standing list and run through it on every build:

• The interrupter. Start talking over the greeting. Does the agent stop and listen, or plow on?
• The rambler. Give a sixty second story with the actual request buried in the middle. Does it extract the request, or ask you to repeat everything?
• The one word answerer. "Yeah." "Nope." "Sure." Does it keep momentum?
• The topic jumper. Start booking, then ask about pricing, then go back to booking. Does it hold the thread?
• The angry existing customer. Not a new lead. Somebody with a complaint about work that was already done.
• The wrong number. Somebody who clearly meant to call somewhere else. Does it end gracefully or interrogate them?
• The out of scope ask. Something this business genuinely does not do. Does it say so, or invent a service?
• The silence. Say nothing at all for fifteen seconds. Does it prompt, then end politely, or hang in dead air?
• The noise test. Call from a car with the windows down, or a busy room. This one surprises people.
• The accent and speed test. Have somebody who does not sound like you make the same call. If your whole test set is one voice, you have tested one voice.

I write the failures down as I go rather than fixing them mid-round. Fixing between calls tempts you into changing three things at once and losing track of what actually helped.

Round three: data capture accuracy

Everything the agent collects ends up somewhere: a calendar, a CRM, a text message to the owner. Wrong data is worse than no data, because somebody acts on it.

Test each field hostilely on purpose:

• Names that are not obvious. Give a name with a common misspelling and check what lands in the CRM, not what the agent said out loud. Say "Sara without an H."
• Phone numbers. Read one with pauses in odd places. Read one with a mumbled digit. Check that the number stored is dialable, with the right formatting and country code.
• Email addresses. The single most error-prone field on any voice call. Confirm whether the agent reads it back, and whether it survives the trip into the CRM.
• Addresses. Street names that sound like other words, apartment numbers, "Northwest" versus "NW."
• Dates and times. "Next Tuesday" on a Monday. "The first available." "After 5." And the one everybody misses: does the booking land in the client's time zone or your server's?

Then go look at the record, every time. The recording is where the agent sounds right; the CRM row is where the truth is. The spelling and confirmation patterns I bake into the prompt for this are in the voice AI receptionist prompt guide.

Round four: escalation and the fallback chain

An agent that cannot get a caller to a human is a wall, and this is the part that goes untested most often because it is annoying to test.

Call and ask for a human immediately. It should transfer on the first ask, without arguing.

Then run each of the client's emergency phrases for that vertical: gas, flood, no heat, locked out, chest pain, whatever applies. Confirm those route to the destination the client actually wants, which is often a different number than routine calls.

Then test the failure case, because this is the one that costs money. Have the transfer destination ignore the call. Does the agent come back on the line and take a message, or does the caller get dropped into silence? Does the owner get notified that a transfer failed? Run the same test after hours, when the front desk number rings an empty office. The full pattern for warm versus cold transfers and fallback chains is in how to transfer a voice AI call to a human.

Round five: the integrations behind the call

The call can be perfect and the job still unfinished. Test the plumbing separately from the conversation.

• Booking round trip. Book a real slot, then check the calendar. Then try to double book the same slot from a second call and confirm the agent refuses.
• Cancellation and reschedule. Half of the calls a receptionist gets are not new bookings.
• Post-call automations. Confirmation SMS sent, contact created in GoHighLevel, owner notification fired, transcript stored. Fire the webhook path deliberately and watch it in your workflow logs, not in your imagination. My split between live tools and post-call work is in how to connect Retell AI to n8n.
• Failure injection. Break the calendar credential on purpose and call again. The agent should apologize, take a message, and let somebody know, not stall silently or invent a confirmation.
• Concurrency. Two calls at once, then five. Agencies discover concurrency limits during a client's busiest hour otherwise.

The latency pass, and then regression

Once behaviour is right, listen for timing. Pull the recordings from your test calls and find the turns with real dead air. It is almost always the turn with a lookup in it. The fix order I use is in how to reduce voice AI latency.

Then the discipline nobody enjoys: after every prompt edit, re-run rounds one and two. Prompt changes are not local. Adding a rule about insurance questions can quietly change how the agent handles bookings, because you moved the model's attention. I keep a fixed set of ten test calls per client and re-run them before any change goes live, which turns a vague "seems fine" into a repeatable check.

Go-live is a soft launch, not a switch

I do not flip a business's main number to an agent on day one. The sequence I use:

1. Live on a forwarding condition first. After hours only, or overflow when the front desk does not pick up in four rings. Real callers, limited blast radius.
2. Listen to the first twenty real calls. Every one, start to finish. Real callers do things no test list predicts.
3. Fix what those twenty surface, then widen the hours.
4. Take the main line once a full week passes with no escalation you did not expect.

Then set a standing review. I check the flagged calls weekly for the first month, and monthly after that, alongside the outcome metrics I report on in voice AI client reporting.

Let the client see the same calls you do

Here is the part that turns QA from your private chore into something the client values: they should be able to hear the calls themselves.

That is what the portal is for. In VoiceDash your client logs into a portal with your logo on your domain and sees their own calls, recordings, transcripts, and usage, scoped to their data and nothing else. During a soft launch that is enormously useful, because the owner can spot-check the agent on their own terms instead of waiting for you to email a clip, and every fix you ship is audible in the next recording. It connects to your Retell account, is live in under 10 minutes with no code, and plans start at $19/mo with a 7-day free trial. VAPI and Bland support are coming soon; today it is purpose-built for Retell agents.

The bottom line

Testing a voice agent is not one activity, it is five: does the ideal call work, do the messy calls work, is the captured data correct, does escalation work when the destination fails, and do the integrations behind the call actually complete the job. Run all five on a real phone, from more than one voice, before a customer ever hears it. Then soft launch and listen to twenty real calls, because real callers will always be stranger than your list. It is a couple of hours of work, and it is the difference between a client who renews and a client who forwards you a recording. The rest of the go-live sequence is in how to onboard voice AI clients.

Want your clients hearing those calls in a portal with your logo on your domain? Start free on VoiceDash or book a demo and I will walk you through the setup on a call.