What will you learn from this episode?
- In about 90% of the Discovery workshops Inwedo starts its projects with, the same problem sits at the root: an Excel file that has outgrown its role. A spreadsheet nobody consciously designed to be a database ends up driving the whole company’s day-to-day decisions.
- An AI model doesn’t fix poor-quality data — it replicates its errors, just faster and at a bigger scale. And where data is missing, it makes up the rest, because hallucination is a consequence of how the model is built, not a defect.
- Deliberately putting AI on hold to get your data in order first is a smart decision, not a failure. The case of Inwedo’s project with Green Business Center.
- When departments sit down over the data together for the first time, it turns out each of them understands it differently. The company makes decisions on numbers that mean one thing in sales and another in production.
- Processes only one person knows are a business continuity risk — their vacation, sick leave, or resignation stops the business. But that same knowledge makes them a natural candidate for data steward: the company gains a custodian instead of a threat.
- An ISO standard gives you a framework, not ready-made instructions — two companies with the same certificate can apply it in completely different ways. The protection comes from how the standard works in daily practice, and the differences only surface during an audit or a merger.
- Collecting all the data “just in case” ends with a mountain of information where nobody can tell signal from noise. Data without a purpose defined up front isn’t an asset — it’s a cost. First decide which decision it’s supposed to support, then start gathering it.
Watch with English dubbing or subtitles
Why does Excel become a problem in a company that wants to implement AI?
Excel becomes a problem the moment it acts as the company’s database — the only place where data about customers, prices, or production lives, and the basis for everyday decisions. Beata Król, Delivery Lead at Inwedo, estimates this applies to 90% of the inquiries and Discovery workshops Inwedo runs at the start of a collaboration: “it’s almost always an Excel file that’s getting out of hand — in scale, or simply in what it was ever meant to do.” From the other side of the table, Bartłomiej Mach, BI & Master Data Lead at Green Business Center, confirms it, describing the case his company brought to Inwedo: “there’s nothing wrong with using Excel as such. The problem starts when that Excel evolves into some kind of database.”
The opposite decision can be a trap too. Radek Podgórski, Lead Auditor at TÜV Nord, warns against “getting out of Excel” projects with no defined goal. Before anyone declares “we’re moving off Excel,” the company has to answer what it wants to achieve. Otherwise it builds an amusement park where all it needed was a swing.
Why do companies hear “the system now, AI in a year”?
Because the decision “we want AI” is usually made before anyone looks at what the AI would actually run on. Beata Król describes a recurring scenario: a company orders Discovery workshops with an AI system in mind and delegates people from different departments for two or three days — and only in a fraction of cases does AI make sense and have a foundation to work on. The recommendation after the workshops then sounds different from what the company expected: “we’re proposing this system, and AI comes a year from now. Because it’s this system that will make you ready for AI.” These are, as Beata admits, tough conversations.
Bartłomiej Mach knows this foundation from his own implementation — alongside an ERP rollout, Green Business Center planned machine learning for sales forecasting and first had to unify its sales history after a merger of companies: amounts, prices, product indexes, and the customer list. “The board understands that this AI has to stand on all of this. And it needs data, reasonably good data, to act reasonably.”
What’s the risk when company knowledge lives in one person’s head?
One person’s illness, resignation, or vacation is enough to halt processes nobody else knows. Beata Król sees this pattern at almost every workshop: “there’s always that one person who knows what needs to be added to that Excel file, in which column and so on — and that knowledge lives only with them. And that’s what always scares me the most.”
Radek Podgórski assesses the same thing from an auditor’s position — for business continuity, it’s “a major flashpoint.” The starkest illustration comes from workshops run by Inwedo: two owners managing an entire company, asked what they expected most from the project, answered “we’d like to go on vacation — we haven’t been in six years.” They had no one to hand their knowledge of their own company to.
From a data management perspective, though, that same person is a natural candidate for data steward — a custodian formally responsible for data quality who knows where the data comes from. Green Business Center went exactly this way: it built a network of data stewards around its data governance policy, and Bartłomiej Mach compares the whole setup to road signs and traffic rules that keep data-heavy roads free of pile-ups. His recommendation for other companies: “I’d definitely recommend implementing a data governance process — actually writing the policy, taking a close look at it, appointing the data stewards, backing it up with processes and tools, because it really gives you a lot.”
What do Discovery workshops reveal that a company can’t see on its own?
Workshops quickly surface discrepancies in how individual departments understand the same data. Discovery workshops are often the first time people from different departments sit down at one table at all. Beata Król: “that’s when we all sit down together and suddenly start hearing about each other’s problems — or about how we use what are theoretically the same things, tools, and data in completely different ways.”
The value of that meeting doesn’t disappear when a company has done its homework. Green Business Center came to the workshops with a ready internal analysis and a written-up case study — and even so, as Bartłomiej Mach recalls, questioning the users together brought out “a lot of great new things and ideas.” A diagnosis made purely from the inside ends where the company’s existing knowledge ends.
Why do AI models hallucinate — and how do you check whether they actually work?
Hallucination isn’t a malfunction — it’s a consequence of how the model is built: a language model is constructed to always give an answer. Radek Podgórski explains: “these models are built so that they have to answer. You can’t be left without an answer. So at some point it runs out of knowledge and goes: you know what? And it fills in the rest.” The second condition for sensible answers is input data quality — on weak data, as Podgórski puts it, the model speculates more than it reasons.
So how do you judge whether a model works? Precision and recall, and the F1 score that combines them, say more about a model’s quality than accuracy — they measure how many of the cases that really matter the model actually caught. For a company, this comes down to one practical question for the vendor: not “how accurate is the model?” but “how many critical cases does the model let through?”
Why do AI costs grow even though you’re using it the same way?
Model vendors bill usage per token, a unit of processed text — not per query. The cost grows with the volume of data processed, with no change in functionality at all. The conversation quotes exactly this kind of account from large implementations: “great, I’ve implemented it, and from my perspective basically nothing has changed, except it costs me twice as much.”
On top of that comes a psychological mechanism that makes growing costs easy to rationalize. Radek Podgórski, who also reviews accounting during audits, brings up a rule known from finance: a million can look like a lot, but “a thousand times a thousand” doesn’t — the same amount split into a thousand daily installments stops hurting.
What does an ISO certificate give you — and what won’t it handle for you?
A standard gives you a framework for interpretation, not execution instructions. Radek Podgórski, who runs audits as a Lead Auditor at TÜV Nord, says: “The standard doesn’t give you hard guidelines saying you have to do exactly this and that […]. No, it speaks in general terms about what to pay attention to, so it’s also a matter of proper interpretation.” The effect: two companies with the same ISO 27001 certificate can put the standard into practice in completely different ways.
A certificate only starts working when the company actually uses what it wrote down. Beata Król recalls a client from a heavily regulated industry who was one of the few who didn’t complain about ISO procedures — because “they genuinely use it. It’s not just for show”: a written-down process means someone can go on vacation or fall ill without half the company grinding to a halt.
The pressure for this kind of order doesn’t only come from regulators, either — more often it comes from customers. Green Business Center serves McDonald’s, so, as Bartłomiej Mach puts it, “we have to have everything buttoned up, tip-top.”
For managing AI itself there’s already a standard, ISO/IEC 42001, though the starting point is often ISO 27001, which covers information security. The EU’s AI Act and Data Act aren’t stirring much emotion yet — the guests agree that companies don’t yet feel the moment of a real test coming.
When does collecting data hurt instead of help?
When you collect data without defining the decisions it’s supposed to support. The result is a mountain of information you can’t draw conclusions from, because nobody knows what’s signal and what’s noise. Beata Król admits that “I fell victim to exactly this myself” — as a product owner, she and her team started collecting data broadly, without defining what for — until someone asked “OK, so now what?”: “we have a ton of information, and now the question is which of it matters, which of it is just noise […] and which of it should actually feed into decisions.”
The conversation circles back to the classic story of Allied bombers in World War II: analysts mapped the hits on planes that returned from missions, so the places to reinforce weren’t the riddled spots but the untouched ones — planes hit there never made it back at all. It’s the same trap that catches reports built around available data instead of decisions. Right next to it sits a second one, called out in the conversation as art for art’s sake: “dashboards or reports that exist only to glow green, not to help anyone make a decision.”
About the guests
Beata Król is a Delivery Lead at Inwedo, with experience as a Product Owner and Agile Lead in IT projects for business clients — she has spent years working closely with engineering teams and client-side stakeholders, combining a product, process, and technology perspective. She co-creates the Zwinna Łódź community, where she co-organizes workshops on process improvement and software quality; she has spoken at Warsaw IT Days, Women in Tech Summit, and events run by AI Task Force and AI Chamber, among others.
Bartłomiej Mach leads the Business Intelligence and Master Data team at Green Business Center. He’s responsible for data models and the data governance policy — including a network of data stewards — in an organization serving clients with high compliance requirements, McDonald’s among them.
Radek Podgórski is a Lead Auditor at TÜV Nord Polska, where he certifies organizations against ISO 27001 (information security), ISO 22301 (business continuity), and ISO 42001 (AI management systems). Thanks to his technology background, he looks at AI implementations not as another trend but as a recurring pattern. This is his third appearance on Inwedo Podcast.
Jagoda Lazarek — AI & Innovation Business Partner
She holds a PhD in computer science and has over 15 years of experience across academia, business, and startups. She specializes in AI, computer vision, VR/AR/MR, and recommender systems, supporting companies in adopting new technologies. A graduate of the TOP 500 Innovators program (Cambridge, Oxford, Imperial College London) and a former research intern at IBM Haifa.
About the podcast
Inwedo Podcast is a series of conversations with practitioners about technology that changes how organizations operate — not about trends, but about what works and what doesn’t in real implementations. If your own answer to “are we ready for AI?” starts and ends with the data, that is where our work usually starts too.