From theory to practice: Innovating with AI

AI is the most high-profile area of ​​technology investment for global CIOs this year, according to 42% of respondents in Foundry’s recent State of the CIO 2025 survey.

Interest is pervasive across all industries and geographies. In EMEA in particular, 76% of CIOs report having been specifically tasked by their board to research and evaluate potential AI products to add to their tech stack. Just over 75% say they’re working more closely with the business on AI applications, and 64% say IT and business are aligned on the adoption and use of gen AI.

But such excitement doesn’t always translate into measurable productivity. How to move from theory to practice is the key question CIOs are asking. Even those who most believe in the potential of AI want to be sure they can implement projects that deliver real transformation and benefit.

“It’s not as easy to apply AI in a company as vendors would have us believe,” one CIO said under the condition of anonymity. “Having adequate prompt engineering to best interact with generative AI by asking the right questions, and having sufficient, high-quality data for AI projects are the two obstacles.”

Another factor is demonstrating the potential of AI to the board, with precise ROI metrics and use cases that prove it works.

“AI represents a momentous turning point, and the potential is immense,” says Giuseppe Ridulfo, head of information systems at Banca Popolare Etica. “But like all far-reaching changes, it requires caution and awareness, especially when dealing with sensitive data. This is why many CIOs in the banking world are wondering how to move from theory to practice.”

A vertical look at data

It’s not just the banking world that’s asking. For CIOs across the board, AI that answers questions, speeds up searches, and summarizes documents can be useful on its face, but isn’t entirely conclusive.

“Industry needs a killer application, and to achieve this, it needs certified data on which to run AI algorithms,” says Gianni Sannino, head of business network services at Sirti Digital Solutions. “The true practical application of AI should be something that increases the efficiency of human action or the production line,” he says. “Today, everything is advertised as AI, but we must be wary of slogans. The truth is that industry is still searching for a true practical application, and the key is missing because the input data is lacking in quality.”

According to Sannino, the problem is there’s no federated body capable of certifying and standardizing the data that trains and feeds AI. Standardization bodies are needed to certify data and create vertical data lakes, or those targeted by industry sector.

“Hyperscalers certainly have no interest in dealing with this because their online data dominate, but Europe could, and should, do so as well through corporate R&D centers and universities,” he says.

So to talk about AI in an abstractly enthusiastic way is pointless. CIOs want traction and tangible change, and to find it, adds Sannino, vertical AI, and specialized agents and data, are needed, not generic AI.

The same goes for open-source AI. “It can be used, but it too should be able to provide certified, verticalized data lakes,” he adds. “Data, not software, is the foundation of AI. Otherwise, how can we expect to use AI to make decisions? Ultimately, AI today is still fundamentally RPA.”

Change management and strategy

Generally, moving from theory to practice in AI depends on the industry and the size of the company, as well as the level of investment. Small companies or highly vertical manufacturing tend to adopt AI in internal processes as they’re usually satisfied with products like Copilot and a virtual assistant. But moving from theory to practice requires budgeting for costs beyond the purchase of the platform or licenses.

“We need to invest in change management to ensure adoption and generate benefits,” says the CIO of a finance company. “To extract value from AI investments, people must actually use the product and make the most of all available tools.” 

Therefore, training and adoption plans are required, including use cases and best practices. These come at a cost in terms of money, time, and personnel, but they’re essential to bring AI into business practice.

If the model needs to be built from scratch, the data must also be prepared and an external provider must be consulted. The investment is even more significant, too, and the operating costs of the cloud on which the model runs soar. So AI implementations always rely on a strategy.

“AI can’t be improvised,” the finance CIO reiterates. “You need to start with a clear idea of ​​the product you need, how to implement it, and which vendors to work with. Strategy is also essential because it must take into account the need for governance and regulatory compliance, such as the AI ​​Act , and the fact that the technology is constantly evolving.” 

It’s no coincidence that many large companies, especially in regulated sectors, despite having greater resources to invest, remain in the experimental phase. Driven by a maximum focus on costs and compliance, they design use cases to incorporate into production processes, but are hesitant to scale them up. In other words, they’re far from creating an internal infrastructure and governance model that can support those use cases and minimize their costs.  

“For now, we too have stopped at individual projects and are waiting for the technology to evolve,” he adds.

A dedicated department for AI

Ridulfo considers AI to be both an IT and HR issue that requires training and making people aware of its potential and risks, such as data exposure or even loss of skills.

Banca Etica is currently conducting a pilot with Google Gemini, testing a series of use cases such as meeting transcription, text generation, and document summary. The goal is to identify the practical applications most useful to the bank’s employees, and limit their use to these specific cases, with controlled benefits and risks, and HR support for training.

Banca Etica has also established a data analysis and AI office, part of the organization department, of which Ridulfo is deputy head. This was created primarily to dedicate specialized resources to data governance, and the internal data warehouse project on which data analysis will be based. However, it also oversees adoption of emerging tech like AI, ML, and RPA, conducting tests and experiments.

The bank’s data analytics and AI office is also currently working on a pricing model for credit granting. The model integrates traditional credit data analysis with ESG factors, enabling personalized and customer-specific pricing. In this pilot, the bank is working with an Italian AI technology provider, supported by its in-house development team — a move that’s proven to be highly effective, says Ridulfo.

“The provider offers us proximity and control over the tech,” he says. “Regarding internal development, we brought in two senior students from the University of Pisa who completed their doctorates with us: one works on RPA, and the other does algorithm modeling.”

The turning point with gen AI

The availability of gen AI solutions can prove crucial in the transition from pilots to implementation. Giuliano Rorato, IT manager at public entity supporter Abaco describes their experience.

“We started gradually because AI is a path that needs to be approached with caution,” he says. “We began experimenting with AI applications like ML, eventually creating a prototype.”

Then about four years ago, Rorato’s team developed a tool to forecast activities in the debt collection office. This area was chosen after an analysis that concluded it would benefit the most from AI. Here, the software is tasked with developing forecasts to optimize debt collection from Abaco’s local government clients. However, the prototype, built by a senior student, was never brought to production, since the work required, in terms of time and resources, was significant, he says, and efforts were directed toward other areas deemed as priority.

This year, however, the project was recovered and moved on to implementation, thanks to the evolution of gen AI. 

“Generative AI has given us a boost not only because it offers a concrete new model for practical implementation, but because it’s made AI indispensable for all businesses,” says Rorato. “Those who don’t adopt it lose competitiveness, but it doesn’t change things in terms of the time, people, or money we can invest. But the technological tools available today seem to guarantee better results. Plus, everyone is using gen AI now, and even at the top of companies, investing in AI seems much more strategic and a priority than it did a few years ago.”

To implement the project, Abaco hired a senior and a junior specialist, and created an internal IT team dedicated to AI and the development of this type of project.

“Finding them wasn’t easy,” he admits. “But dedicated resources are essential.”

Time to move beyond POCs

Talent is crucial to move from theory to practice in AI projects, to help identify use cases that generate returns. But some analysts emphasize the need to accelerate the transition from testing to production, with an agile and more courageous approach.

“What I’ve observed is AI is learned by doing,” says Michele Caruso, CXO Europe at Wimbee Tech. “Much certainly depends on your industry and ecosystem, but AI, like all disruptive innovations, can’t be learned long-term. Success rates in the initial adoption phase are low. Compared to prototypes and experiments, the AI ​​projects that a company actually manages to industrialize on a large scale are few, but by adopting a test-and-learn approach, it’s easier to understand which use cases work, and what works make up for all the unsuccessful tests, ultimately unlocking business value and tangible growth.”

With this in mind, he adds, move from theory to practice, and away from a POC mindset to learn on the job without getting swept away by the hype or getting caught up in excessive caution, because there are areas where AI brings many benefits to get competitive advantage.

Ultimately, theory and practice should proceed hand in hand, rather than following one another. If failure occurs, lessons must be learned to use constructively, or fail forward, as Gartner words it. Along this path, consistently having quality data is the key to turning tests into useful lessons, and theory into practice rather than costly failures.