PROD Trends

The head of digitalisation at a large industrial group described a programme that covered thousands of AI assistants and well over ten thousand trained employees. Her team started by asking who owns the data, who owns the process and who will measure the change. Tools were chosen later. Progress was counted in changed ways of working, not in licences.

Anna Potocka, Head of AI at LUX MED, pointed to exceptions: the processes due for automation had so many that automating them stopped making sense. Before the agent got access to the knowledge sources, some of them had to be rebuilt so they could be understood at all.

Karol Tajduś of Bank Pekao added a condition rarely mentioned during a boom: architecture decisions have to leave room to replace a vendor, and nothing should be rebuilt just because a new model appeared. Michał Partyka of Zowie showed that in a world of agents the foundation is the API, and that a tool description or an error message becomes an interface that has to be designed like a screen for a person.

The map below shows which theses kept returning across talks regardless of industry. Foundations and measuring the effect connect the most talks. This was not the organisers' plan. The speakers had not seen each other's slides.

Dawid Telepski of LPP described it as yesterday and today. Yesterday the consultant looked up the customer and the order, checked the store and the parcel status, wrote an email and waited for a reply. Today the agent has the full context of the case from the first second, checks the parcel, contacts the store and watches the deadlines, and the consultant talks to a person who has a problem.

In healthcare it looks different. The doctor could see the patient's context in the dashboard before the appointment, so the patient would not have to tell the same story again. At a bank that automated updates across hundreds of microservices, developers won time back, although part of it goes on the next round of security fixes.

These are changes for people who were not on stage: consultants, doctors, programmers. The audience was mostly their managers. Almost seventy percent of the room were managers, directors and board members, the people who decide what somebody else's working day will look like.

One of the consultancies put numbers on it: teams working with agents are half the size, roles consolidate around writing specs and accepting results, and the organisation flattens. An integrator answered the question about juniors: they are still needed, because AI mistakes are bigger, faster and confident, and someone has to learn to judge them.

Piotr Jósiak of Zowie showed where human time goes in such a team. Most of it goes on writing the spec, because before the agent starts building, the spec has to be understandable, complete and consistent. The rest goes on reviewing the evidence and accepting the result. An integrator added who writes the rules the agents work by: seniors, the same people who were most sceptical at first and came round only after their own first win.

At LPP the consultants did not wait for a finished system. Ten of them became testers: they tried the platform in real customer service and went through their feedback with the team every day. Dawid Telepski built the tool together with the people who were going to work in it, because from the start the point was one place for both the person and the agent.

Marcin Bednarski took apart on stage the legend that the team will love the change because it saves time. The team hears ‘AI will help’ and thinks ‘what happens to me?’. Until someone shows what the working day will look like after the change, everyone writes their own answer.

The change does not stop at the contact centre and the IT teams. An integrator showed that analysts, testers and leads already use the same mechanisms as developers: for test cases, acceptance statuses and workshop notes. At one of the companies rolling out an AI assistant, employees described their day in a similar way: the assistant proposes a first version, they refine it and still check the sources. Karol Tajduś of Bank Pekao put it most broadly.

A consultancy opened the thread coolly: AI adoption in development teams is almost universal and the effect on results modest, because spend grows faster than anyone can measure the outcome. An integrator named the metric that lies the most: lines of code generated. A logistics operator asked the room a question few could answer: who at your company knows what AI costs per month? He also named a cost no licence shows: AI that moves data between systems that don't talk to each other is a very expensive plaster, paid every month.

Marcin Bednarski proposed a sobriety test for it. Every project gets checked in three places: do we know which problem and which measure we are improving, do we have the data, the process and a way into the systems, and is it clear who owns an error and when a person steps in. A missing answer is a reason to name the missing condition and fund it, rather than to drop the project.

The myths the speakers dismantled on stage are collected next to this text in pairs: what the company thought at the start and what it said after two years.

If lines of code lie and the baseline has to exist before the start, the question is what to measure. Public research shows how far perception drifts from measurement. In DORA's 2025 research, 90% of people building software use AI, more than 80% say it makes them more productive, and 30% trust AI-written code little or not at all. A year earlier the same research team estimated that delivery stability falls as AI adoption grows. According to Workday research from 2026, employees lose nearly 40% of the time AI saves them to fixing its output, and only 14% consistently see a clear net gain.

Zowie's speakers showed two ways not to fool yourself. An agent building software passed 100% of the tests it had seen and 93% of the hidden ones, so the score that counts is on tests the builder does not know. In testing a voice agent, coverage is counted by the volume of real conversations the tests include. We put it all on one cheat sheet.