Almost every contact centre records its calls, and most of them barely listen to what they recorded. That is the most useful finding in this autumn's Polish industry survey, because it describes a gap that no new model will close on its own: the gap between keeping data and learning from it.

We read the survey as practitioners who run telephone contact for banks, insurers, healthcare providers and energy companies. Our view is simple. Recording is an archive, listening is a sample, and analytics is something else again: a way of asking which customers, which buckets and which weeks produce the result, and why.

What the survey found#

In September 2026 EY-Parthenon and the Polish Direct Marketing Association (PSMB) published “Contact center w Polsce”, a report on the state of the Polish contact-centre industry. It combines a survey of about 50 organisations with a consumer survey of roughly 1,000 people. A press release two days later repeated the headline findings; it is the same research, not a second confirmation, so we count it once.

Headline findings, EY-Parthenon and PSMB, Contact center w Polsce (16 Sep 2026)
FindingShare of organisations
Record their customer calls98%
Analyse less than 5% of their recorded calls69%
Report lower service costs (organisations using AI)83%
Report higher customer satisfaction (organisations using AI)23%

Declarative answers from about 50 organisations in Poland; small sample. The 83% and 23% are shares of the organisations that use AI. The press release of 18 Sep 2026 reports the same data.

Two caveats matter. The answers are declarations, not audited measurements, and the business sample is small. Neither caveat weakens the direction of the finding. If anything, organisations tend to describe their own analytics generously, so the real share of unexamined calls is unlikely to be lower than reported.

The second pair of figures tells the same story from a different angle. Among organisations that use AI, far more say it has cut their costs than say it has made customers happier. Cost is easy to see on an invoice. Satisfaction, payment behaviour and retention only become visible when someone connects contact data to outcomes, which is exactly the work that 69% are not doing at scale.

Why the archive stays unread#

Recordings were never designed for analysis. They exist so that a dispute can be settled, a complaint checked or a regulator answered. Retrieval is the use case, and retrieval only needs the file to exist. Nobody budgets for listening to the other 95%.

Manual listening also does not scale. A quality team can listen to a handful of calls per agent per month, score them against a form and give feedback. That is useful coaching, but it is a sample chosen for people management, not for understanding a portfolio. It answers “did this agent follow the script?” rather than “why did this bucket stop paying in March?”

Finally, most reporting is built around what the telephony platform counts by default: minutes, calls, handling time, queue time. These numbers are precise, cheap and comforting. They are also the wrong frame for the questions that matter to a head of collections or a head of customer service.

Counting minutes is not analytics#

We stopped treating minutes as a measure of value some time ago, and in September 2026 we stopped publishing them as a regular operating indicator. The reason is practical. In early collections a short, specific conversation that ends with a clear commitment is often the better one. A process that runs more minutes this month may simply be running longer, less effective calls.

Volume metrics answer “how much did we do?” Analytics answers “what worked, for whom, and what should we change tomorrow?” The first question matters for capacity and billing. Only the second improves results.

What analytics means in contact operations#

When we say analytics we mean a set of questions asked of every process, with the data to answer them. In our experience five dimensions explain most of the variation in outcomes.

  • The base. Where the records came from, how fresh they are, how complete the phone numbers are and how often the base has already been worked. We cannot jump over a bad base, and analytics is how we notice one in the first days rather than the last.
  • Reach. Which numbers, time windows, outbound identities and retry intervals actually connect. Without reaching the customer there is nothing else, so reach is the first rung we analyse.
  • The target group. Which segments respond, continue the conversation and give the key information. A segment that hangs up after the company name needs a different opening, not more attempts.
  • Buckets. In collections, accounts move from one delinquency stage to the next every day they are not handled. Results only mean something when compared within the same bucket and portfolio.
  • Seasonality. Payday cycles, holidays, school terms, heating seasons and renewal peaks change reachability and willingness to talk. A comparison that ignores the calendar compares two different worlds.

None of these requires exotic technology. They require data that is structured at the level of each case, a definition of success agreed with the client, and people whose job is to look.

How we work with contact data#

Because every conversation in our processes is executed by Vee's AI, every one of them leaves a structured trace: recording, logs, the version of the dialogue that ran, and statuses. We keep four status axes apart: technical (did the call connect), dialogue (how far the conversation went), processing (what happened to the case) and business (the client's version of the truth, in its own language). Mixing them is how most arguments about KPIs start.

That trace is what makes analysis possible across the whole volume rather than a sample. Our analysts look for correlations between the way contact is set up and the result, and they test changes on comparable groups with the client's agreement. When something works, one change applies to every live conversation at once.

Every day, in every project, our people listen to selected calls, especially unusual ones, to hear with their own ears what technology still cannot. – Dawid Wójcicki, Founder & CEO of Vee

Human listening has not gone away; it has changed its purpose. Instead of scoring randomly chosen calls, our quality team listens to the calls that the data points to: the unusual endings, the segments that suddenly behave differently, the first days of a new base. Analysis tells them where to listen, and listening tells analysts what the numbers cannot.

Measuring against something#

Analytics without a comparator is decoration. We report results against a comparator on the same buckets and portfolios, measured in the client's systems, and we keep declarations apart from outcomes: a promise to pay is not a payment. The honest claim is incremental. We do not deliver the whole recovery rate; we deliver the difference.

What the survey implies for buyers#

If your organisation is among the 69%, adding AI to the front line will not fix that by itself. Automation produces more data, faster. Unless someone is responsible for turning it into decisions, it joins the archive.

The split between 83% reporting lower cost and 23% reporting higher satisfaction is also a warning about how success gets defined. If the business case is written only in cost, it will be measured only in cost. Write the outcome into the business case from the start: payment within a window, attended visits, renewed policies, completed surveys.

Where to start if you are in the 69%#

You do not need a data-science programme to begin. You need one process, one agreed outcome and a habit of looking at the same few cuts every week. Start where the money is most visible: early collections, appointment no-shows, renewals or lead follow-up.

  1. Define the outcome in the client system, not in the dialler: paid within a set window, visit attended, policy renewed.
  2. Tag every case with its origin: base, bucket or segment, and the date it was assigned.
  3. Report reach, continuation and key information separately before you look at the outcome, so you know which rung broke.
  4. Compare the same bucket this month with the same bucket last month and the same month last year.
  5. Pick the ten strangest calls of the week and listen to them as a team.

The fifth step is the one most organisations skip, and it is often the one that produces the first useful change. A cluster of customers who misunderstand a single sentence, or a time window in which nobody answers, rarely shows up in a dashboard until someone has heard it.

Questions to ask a supplier#

  1. What share of conversations do you analyse, and how: transcripts and statuses for all, or listening to a sample?
  2. How do you separate technical, dialogue, processing and business statuses, and which of them do you report to us daily?
  3. How will you analyse our base before launch, and what will you tell us if it is not worth calling?
  4. Do you compare results within the same bucket and portfolio, and against which comparator?
  5. How do you account for seasonality when you report month-on-month change?
  6. Who on your side listens to calls, how often, and which calls do they choose?
  7. When analysis shows a change is needed, how quickly does it reach every live conversation, and who approves it?

A supplier that answers these in minutes and records is selling capacity. One that answers in bases, buckets and seasons is running your process with you. That second kind of work is what we have done in telephone contact since 2012, and it is the part of the service we would least like to give up.

Follow The First Second

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

Follow via RSS