Don’t judge a voice agent by its transcript
An AI voice agent’s transcripts can make good calls look broken and hide what callers really hear. Why that happens, and how to review calls instead.
The Utter team · October 31, 2026 · 4 min read
You open a call log and read this line from your AI agent:
“The rent at $322, Maple Terrace is $21.95 a month.”
That looks like a serious bug. Then you play the recording. The agent said, clearly, “The rent at three twenty-two Maple Terrace is twenty-one ninety-five a month.” The caller heard it correctly. Nothing was wrong except the transcript.
We build AI voice agents for property managers, and we spent a long time comparing what our agents wrote, what they said and what the logs showed. The main lesson: a transcript is a poor way to judge a voice agent. It makes good calls look bad and bad calls look good.
A transcript is a copy of a copy
Most people assume the agent’s side of a transcript is the text the AI wrote. On some voice platforms it isn’t. It goes more like this:
- The AI writes a sentence.
- A voice reads it out loud.
- A speech recognizer listens to that voice and writes down what it heard.
Step 3 is what you read. And speech recognizers are built to tidy up numbers, so they often turn a spoken number into a price, a time or a date.
| What the agent said | What the transcript shows |
|---|---|
| three twenty-two Maple Terrace | $322, Maple Terrace |
| sixteen twenty-four Birch Lane | 16:24 Birch Lane |
| twenty-one ninety-five a month | $21.95 a month |
The streets here are made up, but this is the pattern we saw on real calls.
It makes good calls look bad
Say you read “$21.95 a month” and believe it. You file a complaint, or you add a rule telling the agent to be careful with rent. The rule can’t fix anything, because the agent never made the mistake. And every rule you add can change how the agent behaves somewhere else, so now you may have a new problem. (We wrote about that in “Never” is the riskiest word you can give an AI agent.)
It makes bad calls look good
This is the more dangerous direction.
We once had a voice that read “11:45 AM” out loud as “one one four five AM”. Callers were confused. The transcript said “11:45 AM”, because the speech recognizer heard those digits and neatly wrote them back as a time. On paper the call was flawless. We only found it by listening.
A transcript also can’t show you most of what makes a call feel bad:
- Silence. A seven-second wait looks the same as no wait at all.
- Talking over people. The words are all there, in tidy turns.
- Tone. A cheerful “Sorry to hear your ceiling is leaking” reads fine.
- Mispronunciation. Your street name is spelled right on the page and sounds wrong on the phone.
It can confuse the agent too
On some platforms, the same copy also becomes the agent’s memory of the call. So when the recognizer garbles an address, the agent’s own record of what it said is garbled too. It may then ask again about something the caller already confirmed.
To the caller, the agent isn’t listening. In the transcript, it looks like the AI is repeating itself. The real cause is neither of those, and rewriting the agent’s instructions won’t reach it.
How to review calls instead
Start with what happened. Was the tour booked at the right time and at the right building? Did the text go out? Did your team get the message? Check your calendar and your messages, because those are facts.
Listen to anything with a number in it. Prices, times, addresses, phone numbers. Don’t trust the page on these, in either direction.
Listen for the feel. Pick five calls a week and play them at normal speed. Count the moments you wince: cut-offs, long waits, the same line twice. These are what callers remember, and they’re invisible on paper.
Use the transcript to find your place. It’s a good index. Search it to jump to the right moment, then press play.
Ask your vendor two questions. Is the agent’s side of the transcript what it wrote, or what a recognizer heard? And can I see what it actually wrote? If they aren’t sure, that tells you how closely they’ve looked.
Our own call reviews now work this way. We check the audio and the text the agent really wrote, and use the transcript only to find our place.
If you’re choosing an AI voice agent for your properties, don’t judge it by a transcript. Call it and listen: (984) 363-1437 in the US, (604) 706-8349 in Canada.