A text agent can think for five seconds and nobody notices: the chat shows "typing…". A voice agent cannot. The person on the line hears silence and asks "hello?", and if the agent reads an order number as "fifty-one thousand one hundred thirty-six", there is nothing to reread. Grow2.ai, Auspex's AI division, builds voice agents for client calls: order confirmation, setting up delivery, writing the outcome to the CRM. Below are seven differences from a text agent, and each of them first appeared for us as a defect in a live call.
The examples come from two clients we do not name by agreement: an online clothing store (the agent confirms orders) and a service company whose callers are mostly older people (the agent takes orders for delivery via the Ukrainian parcel service Nova Poshta). Call lines below are translated from Ukrainian. How the agent hears is covered in a separate article: where Ukrainian speech recognition breaks. This one is about how it speaks and keeps the pace.
1. You have about one second to answer
In a chat, a reply in three seconds is normal. In a call, two seconds of silence already sound like "something broke". After the first tests, the store client wrote to us: "you can tell it's AI — the voice, the intonation, the pauses before answering". The pause before an answer on those recordings was 2.0 seconds. For the second client we measured 2.6–2.9 seconds.
That pause is made of three parts: speech recognition has to decide the person has finished, the model has to write a reply, and speech synthesis has to start playing. Each part has to be trimmed separately. When we tuned end-of-phrase detection, the median dropped from 0.56 to 0.31 seconds on the same recordings, without a single extra cutoff of the person mid-word.
A separate trap is the first request after idle time. On 6 September, after the caller's "Hello", the model's first reply took 4.5 seconds, while the following ones took 0.7–0.8. Now, while the greeting is playing, the agent "warms up" the model with one short request, and the first reply does not start cold.
2. No time to think, and you pay for it in quality
A text agent can run in reasoning mode: it thinks first, then answers. For a voice agent that costs seconds it does not have. Here is the price: older people often spell a surname with first names, so "Maria, Oleh, Roman, Oksana, Zoia" means "Moroz". In fast mode, with no time to reason, the model did not recognize this kind of dictation once in five tries. So we write difficult cases as rules in the script or move them into code, as in the next sections.
3. Greetings and figures as prepared text
While the model composed the greeting for an inbound call, 1.4–2 seconds passed before the first sound. The caller's early "hello" could cancel that greeting, and an audit on 14 September found calls with up to 59 seconds of silence on the line. Outbound calls had a similar story: on 30 October, in three calls the agent never introduced itself. Now the greeting is prepared text that goes straight to synthesis, without the model.
The same goes for figures in the script. Plans, prices and dial codes are copied word for word from the client's documents. A code like *141# that once became *140# can only be heard in the call recording: the CRM does not show it, and nobody will notice for a month.
4. The agent does not work out how to say numbers, it receives them pre-split
A text agent writes "51136", and the person sees the digits. A voice agent has to say them, and the model decides how. On test calls on 19 September it read order number 51136 as a quantity, "fifty-one thousand…", and in the next call it got a digit wrong. The rule "read it in groups" was in the prompt the whole time.
The conclusion is simple: a code the person checks against their own order should not be left to the model to say. Now software prepares the number: groups of three digits from the right, each spoken as a separate number, with zeros pronounced. The agent says "51036" as "fifty-one, zero thirty-six". The same applies to branch numbers: the model read "№18" as "eight", so the number now reaches the agent already written out in words.
Sums are the opposite: an ordinary number, said the way people say it. In Ukrainian that means "thousand four hundred ninety-nine hryvnias" (1,499), without the bookish "one" before "thousand".
5. Latin letters, sizes and abbreviations as words
On screen, "XL" is read instantly. In a call, the model said "large" or "L-X-L", and passed Latin strings like "XXS-XS" to synthesis as they were. Now the size arrives as ready-made Ukrainian words for "L — XL", and the agent is not allowed to say it in Latin letters. The same words sit in the speech-recognition vocabulary, so the agent also hears "XS" correctly.
One more small thing: the agent names the ordered size in full once, then says "your size". Otherwise "XXS — XS" comes up three times in one call, and that is more tiring than it looks on paper.
6. The conversation is built differently: shorter and in turns
In text, the person sees the whole order at once. By ear, three items with colors and sizes in one line are already hard to hold. In our store case, on the first calls the agent talked for 59–66% of the call time: every line was "Thank you", a block of information and "Is that right?", plus two menu-style questions in a row. Now items come two or three at a time with a short "all correct?" after each group, one question instead of two, and no bureaucratic phrasing.
Dictation is a genre that does not exist in chat at all. While the person spells a surname one to three letters at a time, the agent answers with one word, "Yes.", like an operator taking notes. When a branch number comes with an address ("number 12, Shevchenka street, 5"), the agent does not read the street back: the city and the number are enough to ship.
The reverse matters too: a short answer from the person has to be heard as finished. The agent first treated "Yes, one." as an unfinished phrase and waited for more, which meant 4 seconds of silence. Now it calibrates short answers to the audience.
7. What cannot be said must not reach synthesis
A text agent can add an emoji, and that is fine. A voice agent sends everything the model wrote to synthesis. Between 13 and 19 October the line carried "What's the difference? ☎️", and one reply consisted of a single ☎️. Now emoji are stripped before synthesis, while "№" and "°" stay, because synthesis reads them as words.
The same applies to service tags. We set speaking pace with a tag for synthesis, and on 4 September 242 words of that tag leaked into the transcript and the model's context, so the model started repeating "[speaking very quickly]" on its own. On another call the pace changed from phrase to phrase, because different parts of a reply received different tags. Now there is one pace for the whole call.
How to test a voice agent: by ear, not by transcript
Most of the defects above would look fine in a transcript. "51136" is correct as text, *140# looks plausible, and a 2-second pause in the logs is just a number. That is why we accept a voice agent by ear: through recordings of live calls and test calls on every new version of the script. A plan for that kind of pre-launch check is in the two-week pilot checklist, and what to do with recognition errors is covered in the article on transcript correction.
If you are choosing a voice agent for your calls and want to know where it will stumble on your own data, start with a Grow2.ai AI audit.
By Andrew Maryasov, founder of Auspex and Grow2.ai, Auspex's AI division.