The telephony layer decides whether a conversation happens at all, which makes it the wrong thing to treat as a commodity. We run our own: the SIP trunks to operators, the numbers, the rotation logic, the line checks, the machine detection and the capacity, because in outbound contact that layer sets the ceiling on every result above it.
Another operating model is built the other way round. The product is the model and the voice; the phone line is rented from a communications platform and treated as plumbing. That works for a demo and often for inbound, where the customer dials you. It works far less well when you are calling hundreds of thousands of people who did not expect you, many of whom will not pick up.
Without reaching the customer, there is nothing else. – Dawid Wójcicki, Founder & CEO of Vee
Reach is made before the first word#
We measure every process on a four-rung ladder: reach, continuation, obtaining the key information, and the real business effect. The first rung is the one most AI discussions skip. If the call does not connect, or connects to a voicemail the system mistakes for a person, or is labelled as spam before it rings, nothing about the dialogue matters.
Our cumulative operating data shows the scale of the gap. To 31 October 2026 we had made 298,684,764 single call attempts and held 38,325,799 connected AI conversations. Those are different units and should not be divided into a neat rate, because attempts include retries, unreachable numbers and calls scheduled deliberately at different times. But the order of magnitude is the point: most of the work in outbound contact happens before anyone says hello.
Clients often assume outbound is simpler than inbound. Someone who calls you will cope even with an IVR. Someone you call may not even want to pick up, let alone keep talking. – Dawid Wójcicki, Founder & CEO of Vee
What running our own telephony means#
Our telephony and infrastructure platform, Esense, has been part of the service from the early years. In practice that covers several distinct jobs, each of which affects reach or reliability.
- SIP trunks with operators. We hold the operator relationships and integrations ourselves, in Poland and in the other European markets we work in, and we deal with the configuration changes operators make on their side.
- Numbering. We supply the outbound numbers, choose how they are presented and manage their reputation over time.
- Routing and transfers. When a customer needs a person, we transfer the call to the client's number, queue or skill over SIP or VPN, with the context the client's staff need.
- Dialler and scheduling. Calling windows, outcome-based retry intervals and number strategies run in our own stack, enforced at the system level.
- Machine and voicemail detection. Our own logic decides whether a human, a voicemail or a call-screening service has answered, because that decision shapes both the first second of the call and the next retry.
- Capacity. Lines and channels are sized and expanded by us, not requested from a platform's quota.
The client does not need to supply servers, numbers or calling infrastructure. That is not a convenience add-on. It means one party is accountable for the whole path from the number dialled to the result reported, and that party can change any part of it when the data says it should.
Numbers have reputations#
The outbound number is one of the six parameters of a contact strategy, alongside the destination number, the time window, channel and content, attempt intensity and the persona. It is also the one most often ignored. A number that has made too many short calls can be labelled as suspected spam by handsets and network services. A presentation customers do not recognise reduces pick-up. A number that looks foreign or unfamiliar in the customer's market behaves differently from a local one.
Because we own the numbering, we can treat the outbound number as a variable: monitor how each one performs, rotate where appropriate, and match presentation to the client and the market. On the destination side, when a customer has several numbers on file, we work through them in rotation rather than hammering the first one. Neither is possible if the number is whatever a rented platform assigns.
Checking the line before you call it#
Our first outbound campaign, designed by us and tested with a call centre during 2016–2018, taught us a lesson we still apply. One of the changes that made it work was checking whether numbers were active before calling them. A base full of dead numbers wastes attempts, distorts reach statistics and makes a campaign look weaker than it is. The same campaign also showed how much the choice of voice matters in cold contact, and that is when conversion appeared.
Line checks are not glamorous, and nobody puts them on a product slide. But the quality of a base is a business variable, and "we can't jump over a bad base" is a phrase our team uses for a reason. Knowing which numbers are reachable is the start of knowing what the base can deliver.
Voicemail, call screening and the first second#
The first second of a connected call is where telephony and dialogue meet. The system has to decide almost instantly what has answered. If it greets a voicemail as a person, it wastes the attempt and may leave a confusing message. If it treats a person as a machine and hesitates, the person hears silence and hangs up. Call-screening features on modern smartphones add a third case: an automated assistant asking the caller to state their business before the phone even rings for the owner.
Each of these outcomes also feeds the retry logic. Different outcomes lead to different next steps. If machine detection is wrong, the retry clock is wrong too, and a campaign either over-contacts some people or gives up on others too early. That is why we run our own detection rather than accepting a default, and why raw telephony statuses are something we verify against what actually happened on the line.
Capacity is a business variable#
In early collections, each day without contact moves some debtors into a later bucket, and for a bank a later bucket means higher reserves. In lead qualification, a fresh lead cools by the hour. In appointment management, a campaign for the coming days has to be completed in hours so freed slots can be refilled. In all of these, how fast a whole assignment can be worked is part of the result.
We learned how much this matters in March 2020. When the pandemic began, demand from our large clients multiplied roughly fivefold within weeks. Because capacity depended on infrastructure rather than on hiring and training people, we could expand channel capacity at a mobile operator and absorb the surge. A contact centre cannot add staff that fast. A vendor whose telephony sits on someone else's quota is limited by that quota.
| Decision | What it changes in the result |
|---|---|
| Which outbound number, and how it is presented | Whether the customer picks up at all, and whether the call is flagged as spam |
| Which of the customer's numbers to try, in what order | Reach on accounts with several numbers on file |
| Whether a number is checked before dialling | Wasted attempts and the accuracy of reach statistics |
| How the answering party is classified | The opening of the call and the timing of the next attempt |
| When the line is opened and how often it is retried | Contact rate, over-contact risk and compliance with market rules |
| Where and how a transfer lands | Whether the customer reaches a person with context, or starts again |
Each decision is configured per process and per market, and reviewed against results.
Markets add their own layer. Calling hours, permitted attempt frequency and number presentation rules differ between Poland, Austria and Germany, and a process that works in one has to be re-checked before it runs in another. When the telephony is ours, those differences are configuration we control and can evidence to the client's compliance team. When it is rented, they become questions for a third party who is not in the room when results are reviewed.
Why AI vendors outsource it, and why we don't#
There are good reasons a voice-agent platform may choose not to run its own telephony. It is slow, unglamorous work: operator contracts, number regulation that differs by country, configuration changes on the operator side, codec and latency issues, monitoring of line quality. None of it shows in a demo. A platform can be launched far faster on rented telephony, and for many use cases that is a reasonable choice.
It is not a reasonable choice for a service whose value is measured in reached customers, kept promises and confirmed visits. If reach is the first rung of the ladder, the party responsible for the result has to control the layer that determines reach. That is why, from our earliest years, we built telephony into the service, and why we still see it as one of the most underrated parts of what we do.
What to ask a supplier#
- Who owns the SIP trunks and operator relationships for this service, and who fixes it when an operator changes its configuration?
- Whose numbers will customers see, how is their reputation monitored, and can they be rotated or changed per market?
- How does the system tell a person from a voicemail or a call-screening assistant, and how does that decision affect retries?
- Are calling windows and attempt limits enforced by the system, or just scheduled?
- How quickly can capacity be expanded if our volume doubles next month, and who decides?
- Will you report attempts, connections and conversations as separate units, so we can see where reach is lost?
If the answers point to a third party for most of these, the supplier is selling a conversation layer, and someone else is deciding whether the conversation happens. For the processes we run, we would rather answer for both.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS