Vee has run automated telephone customer contact since 2012, with production at scale since 2017, and in that time the technology has changed almost completely. The lessons about what makes customer contact work have changed very little. This is our history of the period, told stage by stage, with what worked and what did not, and with the press coverage and public appearances that marked the way.
It is not a neutral history of the market. It is the history as we lived it: a company that started in research, went to market too early, found its footing in outbound processes for regulated sectors, and today runs the whole process for large organisations in several European countries. Where we were right before the market, we say so with dates. Where we were wrong, we say that too.
Up to 2015: from menus to a prototype nobody believed#
Before 2012: the age of “press one”#
Automated telephone contact in the late 2000s meant the interactive voice response menu. Callers pressed keys or said single words from a list the machine could recognise. It was efficient for the company and frustrating for everyone else, because the caller had to learn the machine’s vocabulary rather than the other way round.
Our founder, Dawid Wójcicki, had been working around conversational software for years through earlier ventures, and by 2009 he had settled on the idea of a talking assistant: instead of building yet another web interface, the system should simply talk. Work on a voice system began in 2010. It ran straight into a wall. Off-the-shelf speech recognition and synthesis, and the “literary” language processing of the time, failed on real telephone speech. So the team wrote its own algorithms for understanding intent. That decision, made out of necessity, shaped everything after it.
2012–2015: a prototype nobody believed#
The joint-stock company was formed on 25 September 2012, and Voicetel Communications S.A. was registered on 11 December 2012, founded by Dawid Wójcicki and Krzysztof Płachta. In August 2013 we had our first working dialogue prototype. It simulated a motor-insurance premium process and already understood synonyms and colloquial speech: a driver could say “diesel” or describe putting oil in the tank, and the system knew what was meant. It did simple arithmetic on the answers and caught inconsistencies such as a car model that did not belong to the stated make. In October 2013 Gazeta Ubezpieczeniowa described how the insurance multi-agency Asist was introducing Asystent VT, an assistant built on our technology, to serve the agents it works with.
Then came two difficult years. From 2014 we worked with established speech-recognition and speech-synthesis vendors, but their promises did not match performance on telephone audio. We demonstrated the prototype widely, and the most common reaction was that it could not possibly work without a human behind it. In 2015 we created Brilliance, our proprietary technology for understanding intent, and in March 2015 signed our first contracts; the first client was a national network of clinics and hospitals.
2016: recognition catches up, and we go public#
The turning point was technical. In 2016 cloud speech recognition reached a usable beta, and for the first time the words reaching our understanding layer were reliable enough on the phone. That unblocked commercialisation. On 10 May 2016 Voicetel debuted on NewConnect, the Warsaw Stock Exchange’s alternative market, under the ticker VOI (today VEE). Puls Biznesu and Parkiet covered the debut; Parkiet called it the debut of a virtual call centre.
The interviews from that spring read well a decade later. On 15 May 2016 Puls Biznesu reported that we would not hide from customers that they were talking to a robot. In June, Rzeczpospolita explained that our system did not merely recognise speech but interpreted what was said and led the dialogue. In December, in an interview with CCNEWS, we argued for automating repetitive work and moving people towards advice and empathy. All three positions are still ours.
Not everything about 2016 was easy. Cloud speech recognition was the breakthrough, but in those years some institutions were wary of sending any audio to the cloud, and conversations with some domestically owned firms moved far more slowly than with internationally owned groups that had already thought through cloud risk. We learned early that in regulated sectors the technical answer is never the whole answer: the security, data-protection and supervisory questions have to be answered just as carefully, and in the client’s language rather than ours.
2016 was also the year we designed our first outbound campaign ourselves and tested it with one call centre. Numbers were checked before dialling so we only called live lines, and when we replaced the synthetic voice with a studio-recorded one, conversion appeared. That campaign led to our first large one, confirming customers’ interest in a loan.
2017–2019: production at scale#
From 2017 we were running production at scale in regulated sectors. In February 2017 we took part in the Polish mission at CCW Berlin. In June we spoke at the Trendy Contact Center conference on contact-centre transformation through AI. On 17 July 2017 a passenger phone assistant for Gdańsk Lech Wałęsa Airport went live in Polish and English, built on Brilliance. It was one of the turning points in our history: a live service answering open questions from travellers, not a demonstration. In the years that followed, when the boards of insurers, banks and medical companies asked whether automated telephone conversation could work with real customers, it was the deployment we could point to. CCNEWS reported in August that the project used two separate speech-recognition engines to improve results, an early sign of a habit we still have: choose the best engine for the job rather than marry one vendor. The airport assistant was also designed with a route to a human consultant from the start.
On 22 March 2018 we held our own event, #GdzieJestDialla, at Teatr Kamienica in Warsaw. We presented outbound processes then packaged as named robots: Maya for outbound campaigns, Pola for collecting GDPR consents, Lydia for lead pre-validation, Zoya for collections, and Tessa and Vera for appointments. Shortly after, a proof of concept for four processes at a large bank followed; all four succeeded, and that cooperation continues today. Pola is worth a note: built in May 2018 as GDPR took effect, it treated compliance as part of the process rather than an afterthought.
In 2018 our founder was a finalist in EY Entrepreneur of the Year. In January 2019 Rzeczpospolita wrote that a Polish company had built human–machine telephone dialogue before Google Duplex. On 11 September 2019 Vee Center Sp. z o.o. was registered; today it carries all corporate sales, infrastructure and R&D.
Those years taught us most of what we still teach clients. Outbound is harder than inbound, because the person you call did not ask to hear from you. Reach comes first: without reaching the customer, there is nothing else. The share of people who hang up after hearing the company name is a measure in its own right. Calling schedules, retry spacing and the choice of outbound number move results as much as the dialogue does. We also settled on a principle that still shapes how we work: we do not call on behalf of a bank or an insurer, we call from it, as part of its own operation and under its data-protection arrangements. And from the first projects, people on our team listened to real calls every day, because the metrics never tell the whole story.
2020–2022: the pandemic, a new name and a crowded market#
In March 2020 the pandemic multiplied demand from large clients, and we raised channel capacity for one mobile operator several-fold within a short time. Puls Biznesu headlined it “Bot army versus the virus”. It proved something we had said in a CCNEWS interview that same month: scaling conversational AI depends on infrastructure, not on how many staff you can hire. In July 2020 Puls Biznesu reported a proof of concept with Blue Media: a voice assistant in its Autopay app, used to pay motorway tolls. In that March interview we also warned that systems built on a narrow set of commands were not ready for real conversations and could put customers off automation altogether.
On 26 June 2020 the general meeting renamed Voicetel to Vee S.A. On 4 November 2021 Vee Center received its first ISO 27001 certification; the current ISO/IEC 27001:2022 certificate for Vee Center Sp. z o.o. is valid until 3 November 2027. In August 2021 we began working with Grupa Neuca on selected patient-service processes in one of the group’s companies. In 2022 we released the third generation of Brilliance; the clients named in the press that year included Rankomat and Blue Media.
The market around us filled up. In a December 2021 year review in BRIEF, we observed that many firms had rushed into cheap, ineffective bots and been disappointed. In a December 2022 article in NowyMarketing we argued that what matters is the cost of a single success, not the price of a call or minute: a cheap solution with a low success rate can cost far more per success than a dearer one that works. Those two observations, about implementation quality and about paying for outcomes, became the backbone of how we sell today.
We also made mistakes in this period. An attempt to package our technology as a product for small businesses produced only modest results. It showed us that a self-service product needs a large sales organisation or very strong online marketing, and that our strength lay elsewhere: in running demanding processes for large organisations, end to end. We still think smaller firms will eventually use voice automation widely, but we learned not to confuse what technology makes possible with what a company is built to deliver.
2023–2025: effectiveness, scale and the first foreign projects#
On 28 March 2023 we adopted our 2023–2025 strategy. Brilliance 3.0, Vee DJ and Vee Analytics followed, and we started our first foreign project, for an Austrian financial institution. In April 2024 we told StockWatch that our advantage lay in the effectiveness of the processes we run, measured in KPIs, not in the AI model. In 2024 we handled 8.2 million cases; in 2025, 11.61 million, an increase of 45.7%, with February 2025 the first month above one million cases.
The 2023 strategy was a deliberate narrowing. Instead of chasing every possible use of conversational AI, we concentrated on the processes where we had proof: early collections, sales and lead qualification, appointments, surveys and customer service, for banks, insurers, healthcare providers, telecoms and energy companies. In Q3 2023 the group reported its first positive EBITDA since 2019. Focus, not breadth, turned the technology into a business.
In May 2025, CCNEWS reported our April figures: more cases handled, fewer conversation minutes. We had started counting cases regardless of call length, because results, not talk time, are what clients pay for. In the same period we spoke publicly about the same themes: at the University of Gdańsk in 2023, in CCNEWS podcasts arguing that technology should be judged by customers rather than technologists, and on Radio Gdańsk in 2025 about how voicebots and customer service are evolving.
These were also the years in which the group became profitable. Since 2025 Vee has been profitable at operating and net level, according to its periodic reports, and in 2025 contact-centre processes accounted for 96.89% of revenue. Profitability did not come from selling more minutes. It came from running more cases more effectively, with a team of specialists rather than a floor of agents: Vee employs no call agents, so there is nobody to schedule, train or replace when volumes change. One change to a dialogue applies to every live conversation at once.
2026: language models, the AI Act and new markets#
Large language models changed what anyone can build. Speech that once required months of training to understand can now be classified quickly by a capable model. We use LLMs where they help, for classifying intent in some variants of our technology, and keep replies in regulated processes in a closed, versioned library. Deterministic where it matters, generative where it helps.
Article 50 of the EU AI Act, which requires people to be told when they are interacting with an AI system, has applied since 2 August 2026; the Commission published guidelines on these transparency obligations on 20 July 2026. For us that codified a rule we had stated in 2016. Commercially, 2026 brought a multi-market agreement with an international banking group and our first portfolio in Germany in August. On 15 September 2026 we ended regular publication of case volumes and minutes, in favour of reporting economic value and measurable results. In May 2026 we presented a collections case study at the Golden Solutions conference on effectiveness in collections.
Foreign markets moved from experiment to substance. The foreign share of our contact-process revenue rose from 6.78% in Q4 2025 to 21.95% in Q3 2026 and 25.27% in October 2026, excluding implementation fees, extra services and one-offs. We work in three foreign markets, with a fourth in preparation. The pattern is always the same: start with one process or portfolio, prove the result on the metric the client cares about, then extend to new portfolios, business units and countries.
The market in 2026 looks very different from the one we entered. Anyone can now build a voicebot in an afternoon, and many do. What we see failing is not the technology but the habit of judging it by the demo: by how natural the voice sounds, how quickly it answers and how little a minute costs. We call that AI theatre. The processes that work are the ones someone designed around a business outcome, integrated with the client’s systems, launched on a sample, measured against a comparator and kept improving for years. That was true of our first campaigns, and it is true of the newest ones.
The role of an external partner is to deliver success, not capacity. – Dawid Wójcicki, Founder & CEO of Vee
What stayed true through every phase#
| Phase | What changed | What stayed true |
|---|---|---|
| IVR menus | Callers adapted to the machine | People should not have to learn the machine’s vocabulary |
| Early recognition, 2013–2015 | Understanding worked, recognition did not | Understanding meaning matters more than recognising words |
| Cloud recognition, 2016 | Reliable speech on the phone | Choose the best engine for each job, not one vendor for all |
| Production at scale, from 2017 | Outbound in regulated sectors | Design the route to a human and build compliance into the process |
| Pandemic, 2020 | Sudden spikes in demand | Scale is infrastructure, not headcount |
| Crowded market, 2021–2022 | Cheap bots everywhere | Implementation quality decides results, and cost per success is the right measure |
| LLMs and the AI Act, 2026 | Anyone can build a voicebot; transparency is law | Disclose the AI, control what it says, and run the process for the outcome |
Vee’s own reading of its history; dates and press items as listed above.
Two things are striking in hindsight. The technology that mattered most in each phase was rarely the one people talked about most. And the companies that struggled were seldom short of technology; they were short of people who knew how to design, run and improve a customer process with it. That has been true since the age of “press one”.
That is also why our own team took longer to build than our technology. People who combine customer-service judgement with telecoms, AI, heavy data back-ends and the compliance demands of banks cannot be hired off the shelf. We grew them project by project, through the same phases described above. Each phase left the team with a skill the next one needed: understanding from the prototype years, operational discipline from the outbound years, resilience from the pandemic, measurement from the years of scale.
What to ask anyone selling you the next phase#
- Which phases of automation has your team actually operated through, and what did you get wrong?
- What results can you show from production, with scope and a comparator, rather than from a demo?
- How do you tell customers they are speaking with AI, and since when have you done it?
- Who designs, listens to and improves the process after launch, and how often?
- What do you measure and bill: minutes, or outcomes your client can verify?
The next phase will bring better models and new rules. The questions above will still separate those who run customer contact from those who demonstrate it. We have tried to stay on the right side of that line since 2012.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS