In August 2013 our first working dialogue prototype could already understand a driver who said “I put diesel in it” instead of choosing from a menu. It also did simple arithmetic on the answers and noticed when a car make and model did not belong together. Large language models have since made language understanding far easier, but the hardest part of that prototype was never the vocabulary. It was knowing what the conversation was for.
We tell this story not out of nostalgia but because the current debate about conversational AI often assumes that understanding people began with LLMs. It did not. What began with LLMs is a new and powerful way of doing one part of the job. Understanding which parts have changed, and which have not, is the difference between buying a model and running a process.
What the prototype did#
The prototype simulated a motor-insurance premium process. The caller talked to an automated voice about the car, the driver and the factors that push a premium up or down. The voice was a rough synthetic one; this was 2013. The telephony was supplied by a partner, because we had not yet built our own integration with operators’ SIP lines. The dialogue was stiffer than anything we would put in production today. And yet, for its time, it was remarkably fluent, because it did three things the helplines of the day did not.
It understood people’s own words#
The standard approach in 2013 was the menu disguised as a conversation: “If your car runs on petrol, say petrol. If it runs on diesel, say diesel.” The caller had to learn the machine’s vocabulary. We never built it that way. In our prototype you could say “diesel”, or “I put oil in the tank”, or any other everyday way of describing the fuel, and the system understood that all of them meant the same thing. Synonyms and colloquial speech were part of the design from the first day.
It did simple arithmetic#
The process asked for the car’s mileage and its year of manufacture, or simply how old it was. The prototype then divided one by the other and proposed an assumption: if the car has done this many kilometres over this many years, we can assume it will do roughly the same next year. Trivial maths, but nobody did it on telephone lines then. It meant the caller was not asked a question the system could already answer.
It checked consistency#
Asked for the make, a caller might say “Ford”; asked for the model, “Fabia”. The prototype knew that Ford does not make a Fabia and asked whether the caller meant a Škoda Fabia. Again, nothing clever by today’s standards. But it showed that the system was not transcribing answers into fields; it was holding a model of the situation and noticing when the pieces did not fit.
Nobody believed it#
We took the prototype to the market in 2014 and 2015. The most common reaction was disbelief. People assumed there had to be a person somewhere behind the curtain, clicking tiles or typing responses. With hindsight, we went to market much too early. The language understanding worked, but the speech recognition available to us did not: the engines of the time relied on restricted, dictionary-bound models, were slow, and did not work well in real time on telephone audio.
That is why we wrote our own algorithms for understanding intent rather than relying on the language components that speech vendors offered. The vendors’ “literary” language processing failed on how people actually speak on the phone. The real unblocking came in 2016, when cloud speech recognition reached a usable beta: suddenly the words reaching our understanding layer were good enough for the understanding to matter. Our technology for intent understanding, Brilliance, dates from 2015; production at scale in regulated sectors followed from 2017.
In June 2016, Rzeczpospolita described our approach in an interview: the system did not just recognise speech, it interpreted the utterance and led the dialogue. That distinction between recognition and understanding has been our most consistent public position for a decade.
What LLMs changed#
A great deal, and we say so plainly. Synonyms and paraphrases that once had to be anticipated, listed and trained can now be handled by a model that has seen most of the ways people express things. Building a new process takes less training data. Coverage of unusual phrasings is better. Classification of intent across languages is easier. A prototype that took us months in 2013 could be sketched far faster today.
We use these capabilities where they make sense. Our technology runs in variants matched to how sensitive a process is. Brilliance H is fully deterministic with no LLM. Variants 1 and 2 use a small proprietary model for intent classification only. Variant 3 uses an LLM for classification, so we need less training data and can launch faster, but the LLM does not generate the replies. Variant 4, generative, is intended for low-sensitivity processes and is technically still being built. Deterministic where it matters, generative where it helps.
Even the vocabulary problem has not disappeared completely; it has moved. A general model knows that “diesel” and “oil in the tank” mean the same thing. It does not automatically know the product names a particular insurer uses, the way customers of a particular bank refer to their instalments, or what a patient means by the name a clinic’s own staff give to a department. Each process still has its own words, and somebody still has to listen to real calls, find the phrases the model misreads and decide how they should be handled. The work is lighter than it was in 2013. It is not optional, and it is never finished, because customers keep finding new ways to say the same thing.
| Capability | 2013 prototype | Today |
|---|---|---|
| Synonyms and everyday speech | Designed and listed by hand for each process | Handled largely by trained models, including LLMs for classification |
| Speech recognition | Weak, dictionary-bound engines; a major constraint | Strong engines, chosen per language and process |
| Arithmetic and derived answers | Rules written by dialogue designers | Still rules written by dialogue designers |
| Consistency checks | Rules based on process knowledge | Still rules based on process knowledge, now with richer data |
| What the system says | A fixed set of prompts | In regulated processes, a closed and versioned library; never improvised |
Describes Vee’s own technology path; not a general market history.
What LLMs did not change#
Look again at the prototype. Recognising “diesel” was the easy part even then. The arithmetic existed because a dialogue designer understood that an insurer needs an estimate of annual mileage and that a caller should not be made to calculate it. The make-and-model check existed because someone knew which mistakes callers make and what a wrong vehicle record would cost later. None of that came from language technology. It came from understanding the process and the people in it.
That is still where results come from. A model can recognise that a customer in early delinquency has said they will pay on Friday. It does not know, unless someone designs it, that the bank needs the date as a structured field, how the customer's history should shape the conversation, that identity must be verified before any detail of the debt is discussed, or what should happen when a customer mentions financial hardship. Those are process decisions, and they are made by people who know banking, insurance or healthcare.
Building a voicebot is not the same as serving a customer. Between the two lies a very long road. – Dawid Wójcicki, Founder & CEO of Vee
There is a second thing LLMs did not change: accountability. In a regulated process, someone has to be able to say exactly what the system said to a customer and why it took each step. That was easy with the 2013 prototype because everything was explicit. It is easy today in our regulated configurations because replies come from a closed, versioned library and every call can be reconstructed from the recording, logs, statuses and script version. It is much harder if a model is allowed to improvise. Better understanding does not remove the need for control; it raises the stakes of not having it.
Machines should learn people#
The idea behind the prototype was simple: the caller should not have to adapt to the machine. In a 2022 podcast our founder put it as machines learning people, not people learning machines. In 2013 that meant synonyms instead of menus. Today it means more: speech engines tuned for each language, voices that suit the people being called, dialogue that remembers what was agreed last time, and timing that respects when someone can actually talk.
LLMs have made the first part much easier, and that is good for everyone, including our clients’ customers. But in our view a system can understand words well and still fail to serve anyone, if nobody designed the arithmetic, the checks, the exceptions and the route to a person. We built those things by hand in 2013 because we had no other option. We still build them deliberately, because they are where the value is.
What to ask before you buy “understanding”#
- Which answers will the system derive rather than ask for, and who decided that?
- Which consistency checks run during the call, and what happens when one fails?
- Where is a language model used: to understand the customer, or to write what is said back?
- If it writes replies, how does your compliance team approve what can be said?
- Can you reconstruct any single call: what was heard, what was understood, what was decided and which script version was live?
A supplier who can answer these has thought about the process, not only the model. That is the lesson our first prototype taught us in 2013, and it has outlasted every change in technology since.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS