The gap between startup adoption and corporate AI implementation in Portugal is not primarily a technology problem. It is an architecture and change management problem. Large companies, despite having teams, infrastructure, and budgets that startups lack, move slowly—not because their people lack capability, but because the systems they've built over decades now resist transformation. Three practitioners working across Portugal's largest corporations laid bare the contradiction: everyone wants to deploy generative AI, but few know why.
The pattern is familiar. A company learns that AI is "the future." Leadership requests LLM solutions. Teams are assembled. Budgets are allocated. But the use cases remain vague. What exactly should an insurance firm, a retailer, or a manufacturing plant use generative AI for? The conversation revealed that most large corporations are still operating with classical machine learning models—predictive algorithms, statistical methods—that have worked for 15 years. GenAI is genuinely premature at scale. Outside the PSI-20, adoption is minimal. The rest are following a wave, not a strategy.
For small and medium enterprises, the picture is inverted. They lack the data infrastructure, the process documentation, and the technical depth of large corporations. But they also lack the bureaucratic debt. Where a startup can deploy a chatbot in days, a corporation must navigate IT departments, compliance reviews, stakeholder alignment across divisions, and integration with legacy systems. This is not laziness; it is the cost of scale. Yet here lies opportunity: generative AI requires far less structured data than classical machine learning. An LLM can work with unstructured text—emails, documents, reviews—without the rigid table structures that locked out smaller enterprises from previous waves of automation. This changes the access equation fundamentally.
The real work now is not building AI models. It is building organizational readiness. Companies must understand what these tools actually do. They must stop conflating different types of AI. They must map where automation genuinely creates value versus where it creates entropy. A restaurant chain should focus on sentiment analysis of customer reviews and dynamic pricing—problems that shift behavior measurably. It should not deploy chatbots for table reservations just because they are fashionable. A law firm should systematize case law and regulatory code—relieving junior lawyers of grunt work so they can build judgment. A healthcare system should structure unstructured patient records—freeing physicians from administrative drag to focus on diagnosis.
The practitioners offered a consistent recommendation: test, iterate, and understand what you are using. Build threads of context with AI tools so they learn your decision-making logic. Treat them as analytical support, not as decision replacements. Know that these models are probabilistic, not oracular. Engineer your prompts with the care you would use explaining a complex problem to a colleague with no background in your domain. And crucially: start with one clear, measurable problem before you chase the narrative of transformation.