Will AI improve the results of a company that has not organized its data?

Nearly nine out of ten data leaders have encountered incorrect AI results, and almost half of the Polish companies surveyed by EY are disappointed with implementations. We looked into what lies behind this gap and where to start before the next project.

The promise sounds simple: a company buys an AI tool, connects it to its systems, and from then on analyzes faster, forecasts more accurately, and makes fewer mistakes. Research from the last two years shows that this scenario works out less often than budgets assume. Most often, it is not the model that fails, but the data it works on and the processes it is inserted into.

Numbers that cool enthusiasm

The “State of Data and Analytics” report was prepared by Salesforce, a maker of analytics tools and CRM systems (for customer relationship management). The company therefore has an interest in customers investing in data organization, and its figures should be approached with caution. The picture is clear, however: 89 percent of data leaders in organizations that use AI in production have encountered inaccurate or misleading results. Only 49 percent of companies consider themselves fully prepared for AI-based analytics.

The weakest area is the foundations. In terms of data quality, 47 percent of organizations are ready, 46 percent for system integration, and 45 percent for data freshness. Only 19 percent of companies feed AI analytics with data from a single system, and for 32 percent dispersed databases are the main obstacle. At the same time, 95 percent of CFOs consider the use of AI in data analysis risky because they fear incorrect results and leaks of sensitive information.

The problem is older than AI. In an Onestream study, 47 percent of CFOs admitted that over the course of a year they had made a wrong decision based on inaccurate or incomplete financial data. In 72 percent of those cases, the loss amounted to at least $500,000, or more than PLN 1.8 million. AI can replicate such errors, only faster.

Other studies say the same thing, though from a different angle

Gartner, an analytics company not affiliated with Salesforce, forecast in February 2025 that by the end of 2026 organizations would abandon 60 percent of AI projects lacking AI-ready data. This is still a forecast that can only be assessed in a few months. In a study by the same company, 63 percent of organizations either did not have, or were not sure they had, appropriate data management practices for AI.

The MIT Media Lab report (Project NANDA) from July 2025 looks at the issue from another angle. According to it, 95 percent of generative AI pilots did not deliver measurable financial returns. The authors attribute this not to the quality of the models or to regulations, but to the way the tools were implemented: poor fit with teams’ everyday work. So Salesforce talks about data, MIT about processes. These are two different measurements, but they lead to a similar conclusion.

In Poland, a similar picture emerges from EY’s report from April 2026, based on a survey of 497 medium-sized and large companies. EY sells AI implementation services, which is worth keeping in mind. Nearly half of companies (49 percent) are disappointed with implementations, and 17 percent, with today’s knowledge, would not make that decision. Roughly one in ten companies declares that its data is fully prepared for advanced models, and 26 percent do not know what data they actually have. Despite this, 77 percent want to increase spending on AI within 18 months.

Cooperative banks: first order, then automation

A good illustration of this relationship was the debate at the Cooperative Banks Leaders Forum 2026, described by ITwiz. The sector is under pressure: in the first half of 2026, its profit fell by about 30 percent year on year, more sharply than in commercial banks. Even so, representatives of the associations did not start with artificial intelligence, but with data.

He showed the scale with an example: during the migration of six million records, the bank corrected more than 110,000 errors. Małgorzata Matejkowska, President of Kujawsko-Dobrzyński Bank Spółdzielczy, said that her bank has been organizing data for three and a half years, and that artificial intelligence does not need completely clean data, only “clean priority data.” The bank automates simple processes, such as monitoring the validity of customers’ ID cards, with a regular robot, without a large language model.

SGB-Bank is introducing AI selectively. According to Vice President Błażej Mika, automatic sorting of tickets in the IT service desk and a few similar mini-projects generated savings of around PLN 800,000 a year and the equivalent of seven or eight full-time positions. Zbigniew Forenc from ZUI Novum pointed out that banks have mobile apps, but lack case workflow systems, so a customer application on the bank’s side often turns back into paper. Marcin Konopka from Berg System added that closed core systems without API access (an interface through which other programs use the system’s data) block even ready-made tools.

It should be remembered that this is a report from a single conference, and some of the speakers are IT vendors selling solutions to banks. Their assessments are the voice of practitioners, not neutral researchers. However, the sequence of actions they described matches what Gartner, MIT, and EY show.

A pilot that postpones the decision

Agnieszka Bochacka from GE HealthCare and Robert Pławiak from ProService Finteco describe another trap in an article for ITwiz. A pilot makes it possible to manually correct data, narrow the scope, and avoid deciding who will be responsible for the solution after the project team leaves. A demonstration proves that the tool can perform a task, but not that the company can rely on it. According to them, automating a poorly designed process means doing work faster that should not exist at all.

In our opinion, this is the most important lesson from comparing these materials. AI is not a lever that lifts results on its own. It reinforces what is already in the company: order or chaos.

Where to start before the next AI project

Before a manager approves the next budget, a few questions are useful. We base them on the list from Bochacka and Pławiak and on the experiences of the banks from the forum.

• What business problem are we solving, and what is the cost of leaving it unchanged?
• Which data is a priority for this use case, and who is responsible for its quality?
• Can the systems from which AI is supposed to draw data be connected?
• Which activities, reports, or approvals will we eliminate once the solution starts working?
• How will we measure the effect, and when will we decide to scale the project or shut it down?

Smaller companies do not have to start with a data warehouse. The example of Kujawsko-Dobrzyński Bank Spółdzielczy shows that simple automation of a well-described process delivers results faster than a language model placed on top of chaos. The best measure was proposed by Błażej Mika: what matters is not the number of implementations on paper, but who actually uses the tool.