The Czech paradox: we feel it, we don’t measure it
Imagine owning an appliance at home that you’re convinced saves you money. You just don’t know how much. You don’t know if buying it actually paid off. And if your partner asked whether it was worth the cost, you’d only shrug. That’s roughly how most Czech companies relate to their AI projects today.
A survey by Deloitte and the Czech AI Association from April 2026, covering 155 companies across the Czech economy, produced several figures worth reading twice. 85% of companies confirm that AI helps them improve efficiency and productivity. 64 % report better decision-making and data handling. And here’s where we get to the heart of the problem. 44 % of companies admit they don’t systematically track AI usage at all. Only 7 % have formal KPIs and regular monitoring in place.
Julie Slavík, Manager of the Strategy team and an expert on technological innovation at Deloitte, commented on the results accurately in the survey: “85 % of companies see AI’s contribution to efficiency, but only a fraction systematically measure that impact. Without clear data, it’s then hard to identify what actually works and spread it to other parts of the business.”
What now? First, we need to understand why this is happening.
Why AI ROI is so hard to report
There are several reasons, and none of them is trivial.
AI projects often start as experiments. The team gets space to test whether something works, and the pressure for business metrics only shows up when the project is ready to move into production. But by then it’s too late. Without a baseline, meaning an initial measurement you set before the project begins, you have nothing to compare against. How many hours per month did those processes take before the AI assistant? How many errors appeared in reports before the model started reviewing them? Most of the time, nobody measured.
The second problem is fragmentation. A single company might be running Copilot for procurement, an internal chatbot for HR, a GPT assistant in marketing, and an automated workflow in accounting. Each initiative has its own budget, team, and metrics. And nobody tracks the overall impact on the business.
Then there’s hidden value. When an AI assistant saves a senior manager two hours a week, it doesn’t show up in payroll costs. That manager doesn’t disappear. They just have more room for work they didn’t have time for before. So-called reclaimed capacity, meaning freed-up capacity for higher-value work, is real. But when a company looks at its books, the value isn’t there. Many companies simply ignore this value because they have no way to capture it.
In practice, there’s one more factor. AI projects are often commissioned by IT and operations teams. Finance arrives at the table only when the project runs long or exceeds its budget. And by that point, any sophisticated ROI discussion is already too late.
What can actually be measured?
It’s useful to bring some order here. AI project measurement falls into three layers that complement each other.
Hard financial metrics are what every CFO wants to see. Payroll savings when AI replaces specific work capacity. Error reduction that used to cost money in complaints or rework. Shortened order processing time that shows up in cash flow. Reduced call center costs when an AI chatbot handles part of the inquiries. These metrics are easy to measure, but they capture only part of the reality.
Operational metrics are the second layer. Average time to process a single case. Number of inquiries handled without escalation. Share of requests AI resolves without human intervention. The team used to spend time on routine tasks, but now it is devoted to strategic work. People in the organization see these metrics immediately and trust them. They serve as leading indicators that warn you before a problem hits the budget.
Adoption and trust metrics are the third layer that most companies skip entirely. What percentage of employees actually use the AI tool? How often? Do they trust it enough to use its output without review? Because if nobody uses the AI tool, its technical specs are irrelevant. The Deloitte and ČAUI survey also showed that distrust in AI output is one of the main barriers to adoption in Czech companies.
A good measurement framework combines all three layers. Financial metrics defend the project before the board. Operational metrics run daily operations. Adoption and trust determine whether the project will survive in the long term.
A four-step framework for measurement
How do you set up measurement so your project doesn’t end up among the 44 % of companies that don’t measure AI at all?
Step one: define the baseline before you start. Before a single line of code gets written, you need to sit down with the operations team and document how the key metrics look today, without AI. How many hours, how many errors, how much money? Without this number, you’ll never know if AI changed anything. And yes, baselines sometimes go missing even for things we think we know well. Take a typical real-world situation. A company wants to deploy AI for automated invoice processing and realizes it doesn’t know how much time this will take across the accounting department today. Different people have different estimates, and the actual figure was never measured.
Step two: pick three to five key metrics, not twenty. A common mistake is trying to measure everything. The result is a dashboard nobody looks at. It’s better to pick one financial metric, two operational ones, and one adoption metric. You can work with that. You can build on that.
Step three: Set short evaluation cycles. Monthly reports are too slow for fast-moving AI projects. Try two-week checkpoints. You see trends earlier, you can react faster, and the team gets into the rhythm. You also stand a better chance of catching when a model stops working because the input data has shifted. This phenomenon, in which AI results worsen without any code changes, is more common in AI projects than in traditional software.
Step four: connect metrics to decisions. The last step is the one most often skipped. It’s not important to collect metrics, it’s important to use them. Who reacts when a metric drops? Who escalates when the cost per token rises? Who says “stop” if adoption stays below 30 %? Without clear ownership, even the best metric turns into a spreadsheet nobody opens.
Where Czech companies typically get stuck between PoC and production
Proof of concept, a validation pilot, works. The model produces nice results on the test dataset. The board loves the demo. And then the truly hard leg of the journey begins.
Production data looks different from pilot data. In the pilot, you worked with a clean, curated sample. In production, data arrives from different systems, in different formats, with errors and empty fields. A model that delivered 92 % accuracy in the pilot suddenly delivers 74 %.
Integration into existing systems is the second stumbling block. Connecting an AI model to SAP, Dynamics 365, or an internal legacy application is not the same as running it in a test environment. Often, the data the model needs is spread across four systems, and no one has the mandate to connect them. This is also where Model Context Protocol (MCP) comes in, which we wrote about previously. Without it or something similar, an AI agent will never meaningfully reach into corporate data.
Governance and security are the third obstacle. In a pilot, you can afford to experiment. In production, you need to log who entered what, what the model responded, how the response was used. You need to be able to say whether the model leaked any personal data. You need a plan for what to do when the model starts producing wrong answers. Most pilots haven’t solved this, and by the time they do, the project is delayed by three months.
And the last thing kills more projects than we’d like to admit. Nobody looks after what happens to the application a year from now. The model gets deployed, the ribbon is cut, and the team moves on to the next project. Four months later, accuracy is slipping, and nobody knows why. Eight months later, people stop using the tool, and somebody only notices at the annual review.
The Deloitte and ČAUI survey showed that the Czech Republic is slightly above the global average in moving pilots to production. 30 % of Czech companies have pushed at least 40 % of their AI pilots into full production, compared to 25 % globally. Not a bad result. But the remaining 70% of Czech companies are still in a situation where most of their AI experiments end up either in a drawer or mid-journey.
What to expect from a vendor, so your project ends up in production, not in a drawer
If you’re considering an external partner for an AI project, it’s worth discussing a few questions with them right from the start.
How will we measure success? If the answer is “we’ll see once the requirements are clearer,” keep looking. A good vendor works with you on the baseline before development begins. They know that without it, they’ll have nothing to defend the result with.
How do you handle the transition from pilot to production? This is where experience shows. Because it’s not a separate phase at the end, it’s decisions made from day one. How do you pick training data? How do you design the architecture? How do you write tests?
Who maintains the solution after deployment? This question most reliably separates vendors into two groups. One group says delivery ends with handover. The other group has a plan for how the model will be monitored, updated, and who will respond when something breaks. The second group is what you want. An AI project without ongoing support is like a car without a service manual. It’ll run for a while, but it won’t last.
Our experience at eMan shows one thing. The difference between a successful and a failed AI project becomes visible only months after deployment. That’s why we don’t consider a project closed at the moment of going live; we continue with ongoing support, measurement, and optimization. Without that, even a promising solution quickly loses meaning.
How to get out of the paradox
Let’s go back to the numbers from the intro. 85 % of companies see the value of AI. 7 % measure it. The 78 percentage points in between are the space where it will be decided, over the coming years, who gains a real competitive edge from AI and who just spends the budget.
Measurement by itself doesn’t make money. But it lets you answer questions no one in the company can reliably answer today. Where is AI working? Where should we scale it? What needs to be stopped? Without data, these are guesses. With data, they’re making decisions.
Companies that realize this now will have a significant head start in 2027. They’ll know which AI projects delivered value and how to scale them. The rest will keep operating in “I think it works” mode until the CFO explains that this isn’t sustainable.
Frequently Asked Questions (FAQ)
How do you know if an AI project delivers real value?
The only reliable way is to have a baseline from before deployment and compare it regularly with current data. Without an initial measurement, AI benefits remain at the level of perception.
Which metrics should you track first in an AI project?
Start with one financial metric, two operational ones, and one adoption metric. Four to five metrics are better than twenty. Fewer metrics are easier to work with; more of them tend to end up in an unread report.
Why does an AI project that worked in the pilot fail in production?
The three most common reasons: production data differs from pilot data, integration into existing systems is more complex, and there’s no governance, meaning there are no rules for how to monitor the model and respond to its errors.
How do you justify AI investment to a CFO when the value is partly hidden?
Combine hard metrics (cost savings, shorter cash flow cycle) with soft ones (reclaimed capacity, meaning freed-up capacity of senior staff). You need both groups named and measured.
How do Czech companies compare to global ones in AI project success?
In moving pilots into production, the Czech Republic is slightly above the global average (30 % vs. 25 %). In deep business model transformation, however, it lags behind (20 % vs. 34 % globally).
What should you expect from an AI project vendor to ensure it doesn’t end up in a drawer?
A clear idea of how you’ll measure success before development begins. A concrete plan for the transition from pilot to production. An ongoing support model that includes tracking the model after deployment.
