Voice removes your safety nets
In text, a person can re-read, a draft can be reviewed before sending, and a mistake sits still while someone fixes it. On a live call none of that exists — the answer is delivered the instant it is produced, to someone who cannot scroll back.
The failure modes are different, not just worse
Mishearing, accents, background noise, and people talking over each other are voice-specific problems with no text equivalent. They arrive before the model has even been asked to be right about anything.
Text lets a human check the answer before the customer sees it. Voice never does.
Where voice genuinely wins
When your customers phone rather than write, when the alternative is an unanswered line, and when the task is narrow — taking a booking, confirming details, routing a call. Narrow and transactional is the profile that works.
Escalation has to be instant
A caller cannot wait for an email. Voice needs a live transfer path that works during your hours and a plainly stated alternative outside them, which raises the bar on what after-hours coverage really costs to deliver.
Disclosure is more important, not less
A synthetic voice that sounds human without saying so is where the trust penalty is steepest. The disclosure rule applies with more force on a call, not less.
Text first, then voice on the same knowledge
Running text for a few months produces the corrections, the documented rules, and the escalation habits a voice deployment needs. Doing it in the other order means learning all of that live, on the phone, in front of customers.