In this programme, we spoke to András Érczfalvi, a presenter at TrendFM, about how to make our AI project a success. The programme was broadcast on 6 January 2026.
Participants
András Érczfalvi – TrendFM
András Tresch - Quattrosoft
According to research by MIT, 95 per cent of AI projects fail to produce measurable results. This is not a pessimistic prediction, but a fact published by Forbes magazine. Nineteen out of twenty companies invest their energy, time and money in vain – the expected results fail to materialise. But why? And more importantly: how can we avoid becoming part of that 95 per cent?
The grand promise and the disappointing reality
It is hard to avoid the hype surrounding artificial intelligence. Tech giants are constantly claiming that AI will solve everything, replace humans and drastically reduce costs. Many executives believe these promises and rush headlong into implementation. Then reality sets in.
Klarna’s story perfectly illustrates this phenomenon. The Swedish fintech company declared in 2023 that it would become an “AI First company”. In collaboration with OpenAI, they set about transforming everything. They reduced their workforce from 5,500 to roughly 3,000 and proudly took on the role of guinea pigs: they would be the ones to prove that AI is capable of replacing human labour.
Then, in 2025, the tide turned. Klarna’s CEO publicly acknowledged that cost-cutting had been too dominant a factor and that, as a result, they had provided a lower quality of service to customers. They rehired several hundred people and shifted from the ‘AI instead of humans’ approach to the ‘humans and AI together’ model.
What went wrong with Klarna?
Entrusting customer service to AI seemed like a logical decision. In theory, chatbots are capable of handling frequently asked questions, respond more quickly and never tire. In practice, however, things turned out quite differently.
The replies became robotic. Although the response was quick, customers felt they were not communicating with a human. The artificial intelligence provided answers, but did not always understand the problem.
Serious problems with contextual understanding also arose. If a customer had multiple contracts, the AI often confused them. It would answer a question about one contract using details from another.
The process of transferring calls to a live agent was a disaster. The customer had asked several times to speak to a human, but the system kept trying. By the time the call was finally transferred, the customer was already angry – not because of their problem, but because of the customer experience.
Perhaps the most painful aspect, however, was the lack of empathy. When someone is going through a difficult time, responses based solely on raw data are of no help. The AI failed to recognise that a different approach was needed in this situation.
The real lesson
The story of Klarna does not prove that AI is bad. It proves that it was used in the wrong way. Artificial intelligence is not a tool designed to replace humans, but a technology designed to support them.
When AI helps people, the strengths of both parties can be utilised. AI quickly processes large amounts of data, identifies patterns and prepares the ground for decisions. People, in turn, make the necessary decisions based on the data and the situation. Klarna now operates on a hybrid model.
Why do most AI projects fail?
To avoid failure, it is worth examining the main causes.
The first, and most common, problem is the quality of the data. In large companies, data is typically scattered across multiple systems. There are two or three different SAP systems, each containing different customer data. The same customer may appear three or four times because, at the time of entry, a typo meant the system failed to recognise that the customer had already been recorded. When someone rings in, the customer service representative often selects the data from one system and refers to that – whilst the other system may actually contain the relevant information.
There are some large Hungarian companies that have been working for eight to ten years on getting their master data in order – ensuring that the same terms mean the same thing across different systems. And until this is achieved, there is no point in feeding the data to AI – it will not be able to provide consistent answers.
The second problem is the fragmentation of processes. There was one large company where a document was passed from the finance department to the legal department in such a way that it was printed out in finance and then scanned in by the legal department. Each department had its own system, operating completely independently of one another. How is AI supposed to make sense of any of this? Not even simple process optimisation is possible, because technically these are separate processes, even though in reality they form a single process.
The third problem is unrealistic expectations. Many people think that there is no system in place, that AI will come along and everything will just work by itself. This is not the case. AI is not a magic wand. It will not solve the problems that have accumulated over decades. Unfortunately, if we feed it poor data and chaotic processes, it will reproduce the chaos, only faster.
The problem of token burning
There is a lesser-known but very real cost factor: token usage. AI models are usage-based. The more data we feed into the context, the more we pay. If we don’t have clean, structured data and try to compensate for this by feeding everything into the AI, token usage will skyrocket. The monthly bill will be high, and the results will be poor.
My developer colleagues who work with AI say they keep a very close eye on what the AI is generating. If they see it’s heading in the wrong direction, they press the stop button immediately, because if they let it carry on, tokens are being used up unnecessarily and, in the end, they’ll have to start the whole thing over again.
What needs to be sorted out before we entrust our processes to AI?
It’s a valid question: if my company wants to introduce AI, where should I start? What needs to be sorted out beforehand?
- There should be a unified process management system. We must avoid a situation where every department has its own isolated system and communication between them takes place via paper or email. Each person should work within a single system and be able to view all their tasks in one place. This is the basic prerequisite for us to even be able to talk about a process.
- The data must be clean. We know this is a difficult project, but let’s at least get started. Let’s identify duplicates, clean up the master data, and ensure that terms are consistent across different systems. AI will only be as good as the data it is fed.
- Let’s keep our expectations realistic. Let’s not think that AI will solve everything. Let’s use it for what it’s meant for. There are tasks at which AI is far better than humans and can save a significant amount of work. But there are also tasks where human judgement is irreplaceable.
- Let’s focus on the right tasks. Small, routine tasks can be automated relatively well using traditional tools. The real power of AI lies in supporting complex tasks that require critical thinking. Let’s not try to solve everything with it, but let’s think carefully about what we entrust to it.
The paradox of coding
An interesting phenomenon can also be observed amongst developers. A study examined how programmers working with AI perform compared to those using traditional methods. Members of the AI team felt they were making progress with infinite efficiency. The actual data, however, showed that they completed the same task 20 per cent more slowly.
What lies behind this? When you code with AI, you aren’t thinking about how to solve the problem. You’re thinking about how to explain to the AI so that it can solve it. It’s a different skill, a different approach. If the AI goes off in the wrong direction, you have to stop it and reformulate the instruction. This, too, is an iterative process, and it takes time.
This doesn’t mean that AI is useless in development, but it does mean that its implementation has a learning curve, and results aren’t immediate. It requires patience and conscious application.
The reality in Hungary
‘Things must surely be different in Hungary’ – one might say ironically. The reality, however, is that large domestic companies face the same challenges as international players; indeed, in some areas the situation is even more difficult.
The rapidly changing regulatory environment places an extra burden on us. What was fine last year may not necessarily comply with regulations this year. AI needs constantly updated, up-to-date data, and it must also understand the rules. If our systems cannot keep pace with the changes, neither will the AI.
The labour market situation is also unique. Junior roles are losing value. Senior professionals, on the other hand, can become more efficient with the help of AI. AI does not take away their jobs – quite the opposite, it makes them more valuable.
Hybrid model – the right choice
The lesson from Klarna is clear: AI is most effective not in place of people, but alongside them. This approach yields the best results.
The figures back this up – companies using AI see their revenue grow, on average, two and a half times faster than those that do not use AI. This does not mean that the company grows to two and a half times its size – if growth in a sector is 3 per cent, then those using AI achieve 7.5 per cent.
Efficiency gains can also be measured. Intelligent automation of processes can lead to improvements of two to two and a half times. If AI is used effectively, it can add a further 20–25 per cent to this.
Instead of a summary
Artificial intelligence is not a magic bullet, nor is it the enemy. It is a tool that can be used for good or for ill. The 95 per cent failure rate is not the fault of AI – it is the result of a flawed approach, poor data quality and unrealistic expectations.
Making it into the successful 5 per cent is not impossible. It requires adequate data quality, standardised processes, realistic expectations, and an acceptance that AI supports people’s work rather than replacing them.
The technology is ready. The question is whether we are ready for it.
This article is based on a discussion broadcast on 6 January 2026 on TrendFM’s ‘Monitor Délután’ programme. The programme featured reporter András Érczfalvi in conversation with András Tresch, managing director of Quattrosoft Kft.
