Anyone can build a voicebot today. That is the most important change in our market in 2026, and it is not the change most people think. Lower barriers to building conversational software have moved the competition away from who has the technology and towards who can make it work in a real operation: integrated with existing systems, secure, at scale, and judged on business results.

We have watched this shift from an unusual vantage point. Vee has run telephone customer contact for large organisations since 2012, with production at scale since 2017. We employ no call agents; the conversations are executed by our AI, and our people design, listen, analyse and improve. When building the conversation was the hard part, it was hard for us too. Now that it is easy, the rest of the job is more visible than ever.

Anyone can build a voicebot today. Our advantage is not that we can build one, but what we can achieve with it for the client. – Dawid Wójcicki, Founder & CEO of Vee

What actually changed in 2026#

In his letter to shareholders for the third quarter of 2026, our CEO described the year as a fundamental change in the conditions under which technology companies grow. Large language models, AI-assisted programming tools and increasingly capable automation have made it possible to develop, in much shorter cycles, software that recently required large teams and many months of work. That applies to conversational systems, voice technologies, analytical tools and business-process automation alike.

The consequences are easy to see from the buyer's side. A convincing demo of a voice agent can now be produced in days. Synthetic voices are good enough that many listeners cannot tell them from a person on a short sample. Platforms let a business configure a conversational flow without writing code. Procurement teams receive more proposals, from more suppliers, at lower entry prices, than at any point in the last decade.

None of that is bad. It is simply a change in what is scarce. When everyone can build the conversation, the conversation stops being the differentiator. The letter put it plainly: lower technological barriers accelerate the market while changing the nature of competition, and what increasingly matters is the ability to use technology effectively in a real operating environment.

The market is discovering the gap#

If building were the hard part, the boom in available tools would have produced a boom in results. The research published over the past year suggests otherwise. These are different studies with different samples and should be read separately, but they rhyme.

Deployment is not the same as value
SourceSampleFinding
BCG, Agentic AI Is the New Frontier in Customer Service Transformation, Dec 2025180 customer-service leadersOnly 28% report measurable business value from GenAI in customer service
Gartner, survey published 4 Nov 2026275 customer-service leaders, Jul–Aug 2026Just 4% have achieved the expected ROI of their AI initiatives
Deloitte, AI-assisted customer service in banks, 30 Jun 202630 US banking executives77% cite integrating new technology with other systems as a modernisation challenge
EY-Parthenon with PSMB, Contact center w Polsce, Sep 2026About 50 organisations in Poland98% record calls, but 69% of them analyse less than 5% of all contacts

Surveys of self-reported outcomes, not forecasts. BCG's figure is a self-assessment; the Deloitte study covers US banks only; the EY-Parthenon sample is small.

Read together, they describe one pattern. Organisations have bought or built the technology. Far fewer can show that it changed a business result. The obstacles they name are not the quality of speech recognition or the fluency of a model. They are integration, measurement and the discipline to analyse what the calls are telling them.

We wrote about the value gap in more detail elsewhere in this series. Here the point is narrower: the gap is not closing as tools get better, because tools were never the bottleneck.

Building a voicebot is not serving a customer#

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

It helps to be concrete about what that road contains. Take an outbound process, such as early collections for a consumer-finance lender, and follow a single case from assignment to outcome. The conversation design is one step among many.

  1. The assignment. The bank's analysts decide which accounts to hand over today, by portfolio, bucket and customer category. A file arrives, is validated, and exclusions are applied.
  2. Reach. Someone has to answer. That depends on how contact is set up: numbers, timing, retries, supporting text messages, and the voice and persona. Without reaching the customer, there is nothing else.
  3. Continuation. The first sentence decides whether the person stays on the line. That includes the AI disclosure required under the AI Act since 2 August 2026, delivered so that it informs without ending the call.
  4. Identity before detail. No debt information until identity is verified, enforced by the platform rather than by a line in a script.
  5. The conversation. Short, specific and adapted to the customer's history, so that a broken promise is named rather than ignored.
  6. Exceptions. A request for a person, failed verification, withdrawn consent to recording, a bereavement, a legal dispute, an erasure request, financial hardship. Each needs a defined route: transfer, stop or referral.
  7. The result. A structured status per case, returned to the bank's systems in a form it can act on the same day.
  8. The outcome. The payment, which arrives in the bank's ledger days later and has to come back to be matched, so the process can be judged on what it achieved and not only on what was said.

A voicebot platform gives you a good tool for step five. The other seven steps are where most of the effort, and most of the value, sit. They are also where a cheap demo and a production service look most different.

What is still rare#

If building is common, what is scarce? From where we stand, five capabilities, none of which arrives with a software licence.

Operating experience at volume#

To 31 October 2026, Vee had handled 48,626,959 cases, cumulatively, with 38,325,799 connected AI conversations and 28,569,847 AI conversation minutes. A case is a record assigned for handling, not a person or a call. Every one was executed by Vee's AI. That volume is not a marketing number for us; it is the source of most of what we know about reach, retries, wording and exceptions. Many patterns only become visible after the hundred-thousandth call in a given bucket, market and season.

48,626,959Cases handled by Vee's AIA case is a record assigned for handling, not a person or a call.Cumulative to 31 Oct 2026, Vee operating data.

Integration with the client's real systems#

Banks, insurers and utilities do not rebuild their architecture for a supplier. Data comes in through SFTP files, REST APIs, VPN connections and SIP transfers, often in combinations that took years to settle. The integration has to adapt to the client, and that adaptation is engineering, security review and patience rather than configuration.

Security and regulation designed in#

In regulated sectors, a supplier is audited before the first call: information security (for us, ISO/IEC 27001:2022 for Vee Center Sp. z o.o.), sub-processing under GDPR, DORA on the client's side, and the AI Act's transparency obligations. In those processes, what the customer hears comes from a closed, versioned library, and every call can be reconstructed from the recording, logs, statuses and script version. These are not features you add after a successful demo.

Measurement against a comparator#

Counting minutes, records and call durations is not analytics. A process has to be compared with an alternative on the same buckets and portfolios, using outcome data from the client's systems, and a promise to pay has to be distinguished from a payment. That is uncomfortable, because it shows the incremental effect rather than a flattering total. It is also the only kind of number a finance director should trust.

People who know customer service#

The scarcest ingredient is not technical. It is a team that combines customer-service judgement with telecoms, heavy back-end systems, compliance and data. In every project, our people listen to selected calls every day, especially unusual ones, to hear what technology still cannot. That kind of team cannot be hired off the shelf; it is built from shared experience over years.

Scale is not a slider#

Platform pricing pages often make scale look like a setting: more channels, more concurrent calls, more minutes. In practice scale is a set of operating problems that only appear when volume arrives, and they arrive in a particular order.

The first is telephony. Number reputation, rotation, carrier behaviour, voicemail detection and channel capacity decide whether a large campaign reaches anyone at all. The second is timing. A collections portfolio or a fresh lead file loses value by the day, sometimes by the hour, so a process that can technically handle a million records but needs a week to do it has not scaled. The third is variety. At volume, rare situations stop being rare: every day brings bereavements, disputes, people who are driving, people who want a human, people who answer in another language.

We learned this most sharply in March 2020, when demand from large clients rose roughly fivefold within weeks and channel capacity for one mobile operator had to be raised from 250 to 900. Scaling then depended on infrastructure and process design, not on hiring. That has remained true. The same conversation run for ten thousand people and for a whole national portfolio is not the same service.

Outbound shows the difference fastest#

Clients often assume outbound is simpler than inbound. It is usually the other way round. Someone who calls you has a reason to stay on the line and will cope even with a clumsy menu. Someone you call may not want to pick up, and has every reason to hang up after hearing your company name. That is why outbound processes expose the gap between a voicebot and a service so quickly: the customer owes you nothing, and every weakness in timing, wording or voice shows up in the numbers within days.

Every market is a new process#

Taking a working process to another country is another place where cheap building tools help little. Identity verification, politeness conventions, the language of consequences, permitted call windows and consent handling differ between Poland, Austria and Germany, and they are configured per market and enforced at system level. When the content belongs to a bank, localisation is a compliance exercise, not a translation. In October 2026 foreign markets accounted for 25.27% of our revenue from running contact processes, excluding implementation fees, extra services and one-offs. Each of those markets required this work before the first call.

We have seen this cycle before#

Cheap tools flooding a market are not new to us. In commentary for an industry white paper on voicebots at the end of 2020, our CEO argued that growth would depend on the quality of the technology and on deployment experience rather than on availability alone. At the end of 2021, looking back on the year in BRIEF, we described firms that had rushed into cheap, ineffective bots and been disappointed. Customers remember those experiences: Gartner's 2026 survey of 3,566 customers found that only 27% would try a chatbot again after a negative experience.

The 2026 wave is larger and the tools are far better, so the risk is not identical. A modern voice agent sounds much more convincing than a 2020 command bot. That can make the problem harder to spot: a fluent conversation that reaches the wrong people at the wrong time, mishandles an exception, or cannot be measured is still a failed process, only a more articulate one.

What does not change: the person on the line#

Through every phase we have worked in, from early speech recognition and studio-recorded voices to synthesis and language models, one thing has stayed constant. The customer judges the conversation, not the technology behind it. A person will not forgive a weak conversation just because it is automated, and they should not have to.

That is why we treat customer experience, technology and business effectiveness as three layers of the same service rather than three departments. Voice, pace, timing and memory of what was agreed sit next to telephony and data, and both sit next to segmentation, contact strategy and tests. It also means designing the way out. Gartner's survey of 3,566 customers, published in August 2026, found that 87% say companies using GenAI for service must provide access to a human agent. In our processes, transfers to the client's own staff are part of the design from the first specification, paced to that team's capacity and working hours.

Cheaper tools do not change any of this. If anything, they raise the bar, because customers now meet more automated conversations and are quicker to tell a good one from a bad one.

What this means for buyers#

Lower barriers are good news for buyers, provided they change how they buy. Three adjustments follow.

Stop evaluating the conversation in isolation. A demo call tells you whether a supplier can build a voicebot. Almost everyone can. Ask instead to see how a live process is run: the daily data cycle, the status model, the exception routes, the quality-listening routine, the reporting your analysts would actually use.

Buy outcomes you can verify, not activity. When building is cheap, activity is cheap too. Minutes and attempts are easy to produce and say little about value. We no longer bill new agreements by the minute, because a minute rewards long calls while results often come from shorter ones. A pricing model with a setup fee, a base fee per case or account, and an optional performance layer linked to a result measured in your systems keeps both sides pointed at the same goal.

Start on a sample, then scale. The fastest way to separate a service from a demo is a controlled pilot on part of a real portfolio, measured against a comparator, with a joint decision on what comes next. If a supplier wants to start on your whole database, ask why.

What it means for us#

The same change applies to Vee. The new tools make our own development faster, and we use them. They also mean that our technology alone is not a moat, and we do not pretend otherwise. What we offer is the combination the CEO's letter describes: production experience accumulated since 2017, applied with new AI capabilities, inside a managed service that covers organising contact, running the process, quality control, data analysis, measuring effectiveness and continuous improvement.

That is also why we describe ourselves as a managed AI Call Center rather than a voicebot company. A voicebot is a tool. We answer for the whole process: telephony, dialogue, schedules, data, reports and optimisation. The test is not whether the conversation sounds good. It is whether the bank's customers pay on time more often, whether the insurer's policyholders renew, whether the clinic's empty slots are refilled. For example, in early-stage collections we have seen on-time payment frequency rise by 7–13% versus a human-agent comparator (historical results).

The market is reaching this conclusion from many directions. Lower barriers accelerate it. The more voicebots there are, the clearer it becomes that the voicebot was never the job.

Questions to ask any supplier in 2027#

  • What happens to a case from the moment our file arrives to the moment the outcome returns from our systems?
  • How will you reach customers who do not answer the first call, and who decides the retry strategy?
  • Which exceptions are defined, and where does each one go?
  • How will results be measured against a comparator, and on which data?
  • Who listens to calls, how often, and what happens with what they hear?
  • Which parts of your service would still exist if the model underneath were replaced tomorrow?

The last question is the most revealing. If the honest answer is "not much", you are buying a tool, and you will need to build the service yourself. If the answer is a long list, you are talking to someone who does the job.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS