On a live call for a home-services client, a caller gave his last name. The transcription added a letter. The dispatch text went out with the wrong spelling, and by the time anyone noticed, it was in the system.
This isn't an edge case, and it isn't a sign of a bad transcriber. Speech-to-text is excellent on ordinary words and genuinely bad on proper nouns — surnames, street names, the part of an email before the @. And critically, it doesn't tell you it's unsure. It picks the most statistically likely spelling and moves on with total confidence.
Let the caller spell it.
The obvious fix made it worse
Our first instinct was to have the agent spell everything back. That was worse.
Text-to-speech reads run-together letter strings badly — it came out blurred, letters smearing into each other. Phone numbers were the worst offender. And having the agent announce "I'm going to spell your email back to you" and then not spell it cleanly sounds more broken than saying nothing.
What actually works
Let the caller spell it. "Can you spell your last name for me?" Their spelling, captured by the transcriber, is the source of truth. You've moved the hard part to the human who actually knows the answer.
Read it back slowly, one character at a time, with a real beat between them — and never announce that you're about to. Just do it. Announcing sets an expectation the voice usually can't meet.
Say it explicitly in the instructions: don't rattle letters or digits together so they blur. This has to be a written rule; agents don't do it by default.
Don't read phone numbers back at all. Default to the number they're already calling from. Only read back a different number they give you, and then in small groups.
When unsure, ask them to repeat. Guessing is what created the problem.
One more that cost us a call to find: the "@" in an email address. One agent read it aloud as "En En." The intake logic was perfectly correct — the pronunciation was the only thing wrong. Fixed with a pronunciation rule mapping "@" to "at." Worth checking on any symbol or proper noun your agent trips over.
What this means for you
Every wrong character downstream becomes something real: an email that bounces, a text that never lands, a duplicate customer record, an invoice with someone's name misspelled on it.
So when you're evaluating anyone's AI phone agent, ask a specific question: how does it confirm spelling before it writes anything down?
If the answer is some version of "the transcription is very accurate," they haven't run enough live calls yet. Accuracy isn't the point. Knowing when to stop trusting yourself is the point.
Want us to look at where a phone agent could take work off your plate? Grab a free workflow audit.