The robot asks: “Will you come to your appointment tomorrow?” The scenario has three branches: “yes,” “no,” and “call back later.” And the person on the other end says: “well, I’ll try, but only if it’s after lunch.”
This gap is where all the complexity of an interactive call campaign lives. Speech recognition worked perfectly, and you have the text word for word. All that’s left is a small thing: understanding what the person meant. That’s a separate task, and it’s solved in fundamentally different ways.
Speech ≠ meaning
The first thing to separate: speech recognition (STT) and intent classification are two different stages, and you don’t get the second for free with the first.
STT turns sound into a string. That’s it. Next you have the text “well, I’ll try, but only if it’s after lunch” and a list of branches to choose from. Connecting one to the other is work someone has to do: either your code or a separate model.
The last line is a classic trap. A naive search for the word “yeah” in the text will confidently send the person down the agreement branch, and tomorrow the clinic will be waiting for him.
Four approaches
You list the expected words for each branch, and the system looks for them in the recognized text. No machine learning involved, just a plain substring search.
It works instantly, costs nothing, is fully predictable, and can be debugged by eye: you can see which word matched and why. On short closed questions it covers most answers, since people talking to a robot really do tend to answer in one word.
Breaks on: synonyms that aren’t on the list, negations (“yeah no”), long answers, and any phrase the scenario author didn’t think of.
The same search, but smarter: words are reduced to their base form (“will come,” “coming,” “came” → one root), and the distance between strings is calculated to survive slips of the tongue and recognition errors.
Still fast and free, still explainable. It noticeably widens coverage compared with literal search, especially in Russian with all its inflections.
Breaks on: the same things as the first approach. Morphology is about the form of a word, not its meaning. “I’m not against it” and “I’m against it” differ by one word and mean the opposite.
This is where meaning comes in. The model turns a phrase into a vector, a set of numbers whose position reflects meaning rather than spelling. You compute vectors in advance for the reference phrases of each branch, and during the call you look for the one closest to what the contact said.
The main advantage is paraphrases. “Yeah, sure,” “of course,” and “absolutely” end up next to “yes” even though they look nothing like it on paper. Models of this class are small, you can realistically host them yourself, and then the data never leaves your perimeter. That’s a strong argument if you work in healthcare or finance.
The cost: infrastructure. The model has to be hosted, updated, and monitored somewhere, and you have to check its quality on your domain yourself. And a fundamental limitation remains: vector similarity doesn’t understand the context of the question. “After lunch” is equally close to “I’ll come” and to “call me back later,” and the right answer depends on what exactly was asked.
The model is given the question the robot asked, the list of branches, and the contact’s phrase, and asked to pick one branch. Not to retell or explain, but to choose from the options listed.
It’s the only one of the four approaches that sees context. “After lunch” in reply to “will you come?” and in reply to “when is a good time to call you back?” are different intents, and the model can tell them apart. It will also get the meaning out of “yeah no, I won’t come” and “I’m not against it.”
The cost: latency and money on every request. Plus a separate engineering task: making the model answer with strictly one option from the list instead of three paragraphs of reasoning.
Comparison
| What matters | Keywords | Fuzzy | Embeddings | LLM |
|---|---|---|---|---|
| Understands paraphrasing | No | Barely | Yes | Yes |
| Understands negation | No | No | Partly | Yes |
| Takes the question into account | No | No | No | Yes |
| Latency in the conversation | Unnoticeable | Unnoticeable | Small | Noticeable |
| Cost per request | No | No | Your own hardware | Per call |
| Clear why it decided that | Fully | Fully | Via distances | Not always |
| What to fix on a miss | Add a word | Add a word | Reference phrases, threshold | Branch wording |
Three things that matter on a phone call
Articles on text classification usually compare accuracy and stop there. In telephony, accuracy isn’t the main criterion, and here’s why.
A pause costs more than a mistake
In a chat, taking a second to think is normal. In a conversation, silence after your line sounds like “the call dropped,” and the person starts talking over you, asks again, or simply hangs up. A classifier that is more accurate but noticeably slower loses in practice to one that is fast and a bit dumber.
This leads to a simple conclusion: the time budget for a decision is fixed, and it’s small. Anything that doesn’t fit into it doesn’t matter, however smart it is.
The smart stage should be the exception, not the rule
If you run a heavy classifier on every utterance, you pay for every answer, including all those “yes” replies that keyword search would have handled. And those are the majority.
A sensible setup has two stages: a cheap check first, and the smart stage only when the check didn’t produce a confident match. Then cost and latency appear only on genuinely difficult answers, and on simple ones the scenario works as before.
There’s a nice side effect: the better the scenario author chose the expected words, the less often the expensive stage kicks in. A carefully built scenario is literally cheaper to run.
A failure must not drop the call
A person is on the line, and the next line has to be said now. You can’t retry the request the way you would on the web: the conversation happens in real time and won’t wait.
That’s why the classifier must have a hard timeout and clear behavior on failure. The right failure is falling through to the “none of the above” branch, the same one the scenario already goes to when it can’t make out an answer. Then even a completely unavailable classifier just means “we work as if it weren’t there,” not a dropped call.
Build this into the architecture from the very start. A classifier that throws an exception instead of saying “I don’t know” when it’s unavailable turns an external service into a single point of failure for all your conversations.
How it works in AutoCall.kz
In interactive scenarios, every recognition element has an “Else” branch, where a contact goes when their answer didn’t match any of the expected words. That’s exactly where the smart stage plugs in: first the matching against the words you set runs, and only answers that didn’t match are analyzed further, taking the question itself into account and not just the phrase.
The result is always one of your branches or “Else”: by design, the classifier can’t invent a new branch. And on any failure or timeout the call goes to “Else,” so it behaves exactly as it would without any classification.
A practical tip for those who build scenarios: phrase the question so that people want to answer briefly. “Are you coming tomorrow, yes or no?” will get far more one-word answers than “We’d like to confirm your plans for tomorrow.” The clearer the question, the less often you need the smart stage.
What’s next
Everything above is about scenarios where the branches are set in advance. A related topic is voice AI agents, where there are no branches at all: the agent runs the entire conversation and decides what to do with each answer on its own. We wrote separately about the difference between these approaches.
A separate topic we’ll come back to is choosing an omni model: a single model that takes in audio and returns both the text and the meaning at once, without splitting recognition and classification. That changes the picture entirely, because it removes one stage from the chain and its latency along with it. How we choose such a model for AutoCall.kz will be the subject of a separate article.