I vividly remember the moment I realized I was witnessing something rare. I was at my computer, questioning an AI about a cost-analysis problem. The machine didn't just answer — it reasoned, proposed alternatives, and asked for clarification. In that instant, my curiosity for technology and my studies in business economics converged. That convergence became my thesis: AI Agents and Their Impact on the Corporate World, advised by Professor Giacomo Büchi.

This isn't a technical treatise on algorithms or software architecture. It's the account of an objective analysis, conducted through the eyes of someone who studies economics and wants to understand how resources move — and how work changes — when part of the operation is delegated to a machine capable of learning.

Redefining costs: from CAPEX to OPEX

The first lesson concerns the architecture of corporate costs. Historically, adopting new technology meant a substantial structural investment. AI adoption now forces a clear strategic choice. Building proprietary infrastructure guarantees control and security over sensitive data, but it means a significant fixed cost and heavy Capital Expenditure (CAPEX), exposed to rapid obsolescence. Accessing external services through cloud models and API keys instead turns that investment into Operating Expenditure (OPEX), making the cost structure elastic and demand-driven. That flexibility demands rigor in return: a consumption-based model without careful governance can spiral out of control fast.

The true ROI, and its hidden costs

Measuring Return on Investment taught me that enthusiasm has to answer to the data. Despite the success stories, between 80% and 95% of AI projects fail to deliver the expected value. Real competitive advantage doesn't come from incremental efficiency gains — it comes from the willingness to redesign the business model itself. A serious business plan can't ignore the downside: data breach exposure, regulatory penalties, reputational damage, and the well-documented risk of "hallucination," where a model states false information with total confidence. Trusting these tools blindly, without robust human validation, is not a risk worth taking.

The labor market paradox

Writing the thesis meant confronting the question everyone asks: will machines replace us? The data shows automation moving up the value chain — touching not just manual tasks but routine intellectual work too. Up to 25% of hours worked in administrative, legal and engineering processes are estimated to be automatable. But rather than outright replacement, what's emerging is a sharp polarization: specialized skills command a real wage premium and outsized leverage, while routine intermediate tasks lose bargaining power. The way through, echoed by recent EU regulation like the AI Act, is complementarity — "meaningful human oversight" isn't a slogan, it's a requirement.

The Italian gap

The last part of the research collided with the reality of our own entrepreneurial fabric. The size gap in Italy is stark: 71% of large companies have already started AI projects, against just 8% of small businesses. What holds Italian SMEs back isn't only cost — it's a serious shortage of qualified people and data management that's still immature.

Conclusion

What did I take from this? That artificial intelligence is the most powerful deflationary and transformative economic force of our generation — but it's not a magic wand. Continuous training, through reskilling and upskilling, has stopped being a line on a CV. It's the only way to stay relevant in a market that is actively rewriting its own rules.