Every government has an AI strategy now. Far fewer have AI working. In Madrid, four practitioners explained why the gap is so wide and what it actually takes to close it. 

There is a peculiar moment in the life cycle of any new technology in government. The strategy is published. The press release goes out. The minister poses for a photograph. And then, quietly, almost nothing happens. 

The panel that gathered on the Main Stage at the GovTech 4 Impact World Congress was, in effect, an hour-long interrogation of that moment – the long, unglamorous stretch between an AI policy on paper and an AI service that an actual citizen can use. Moderating was Carlos Santiso, the OECD’s Senior Advisor on Digital Government and Artificial Intelligence, who framed the session with a question that hung over the entire Congress: governments at wildly different levels of digital maturity are all stuck on the same problem – how do you turn AI ambition into measurable service improvement while keeping people at the center? 

To answer it, he brought together the following group: Laura Gilbert (the Tony Blair Institute), who had built AI inside the British government; Albert Tort, Catalonia’s Secretary of Telecommunications and Digital Transformation; Rafael Fassio, a state attorney leading on technology and innovation for São Paulo; and Juan Corro, Managing Director of IAM at Madrid City Council. Between them, they represented national, regional, and municipal government on three continents  and they agreed on more than they disagreed. 

What united them was a kind of disciplined scepticism. Not about AI’s potential, which none of them doubted, but about the seductive trap of starting with the technology rather than the problem. 

Gilbert put it most bluntly. Drawing on her experience launching the UK government’s AI incubator, she described the public’s initial reaction as falling into two camps: those who doubted government was competent enough to build AI at all, and those who feared it would use AI to harm citizens, to wrongly deny them benefits, for instance. Her answer to both was radical transparency. You cannot, she argued, simply ask a skeptical public to trust you, least of all in an era of collapsing faith in institutions. So, her team published everything they could: open-source code, evaluations, algorithmic transparency reports, development blogs. The openness did not just reassure the public; it improved the work, by letting outside experts and other governments inspect and contribute. 

But her sharpest point was about language  and it carried a lesson for every health system in the room. Rather than announcing an “AI pharmacist,” a phrase practically engineered to alarm people, her team reframed the project around the actual problem: preventable deaths and wasted money caused by bad prescribing. Suddenly the same technology read not as a threat but as a safety tool. Start with the failure you are trying to fix, she insisted, not the technology you are excited to deploy. Most government AI disasters, in her telling, begin the moment a team announces a solution before it has defined the problem. 

Albert Tort, speaking for Catalonia, offered a complementary discipline: one that bordered on the heretical for a technology conference. He argued that much of the work of “digital transformation” is not digital at all. Catalonia, he explained, has more than 2,000 citizen procedures, and it cannot transform them all at once, so it prioritizes ruthlessly and re-engineers the underlying processes before reaching for AI. Public administration, he noted, is built on safeguards, layered approvals, and risk aversion, and no algorithm fixes a broken process. His most striking idea was what he called “silent AI”: technology so thoroughly woven into the fabric of a service that neither the citizen nor the public servant thinks of it as AI at all. He went further still, suggesting the field had reached for the wrong words entirely. He prefers “collective intelligence” to “artificial intelligence” which is a framing that casts AI as something that augments the capability of public servants rather than replacing their judgment.

If Gilbert and Tort supplied the philosophy, Rafael Fassio supplied the cautionary tale. From São Paulo, he described what happened when his state set out to build a data lake for AI-supported research using anonymized clinical data from Hospital das Clínicas, one of the largest health complexes in the Southern Hemisphere. The ambition was sound. The reality underneath it was not. The hospital, it turned out, ran on more than 60 separate systems collecting and processing patient data. The AI project had to stop while the far less glamorous problem of data architecture was addressed. His conclusion was delivered without sentiment: if your data is still being collected on paper, your sophisticated AI aspirations are premature. He returned repeatedly to a warning against “AI for AI’s sake,” particularly in places where political enthusiasm for technology runs ahead of institutional readiness. His prescription was concrete: a real problem AI can actually solve, procurement and legal models that can tolerate uncertainty, an experimental culture built on pilots and sandboxes, and genuine investment in data infrastructure and skilled people. 

Juan Corro brought the conversation down to street level — literally. Speaking for Madrid, he argued that cities no longer need to hunt abstractly for AI use cases, because the most valuable ones are sitting in plain sight, embedded in the hardest urban problems. For Madrid, the clearest of these is housing. He described two flagship efforts: one to accelerate urban planning processes that have historically taken more than a decade, and another to make building permits faster by rendering urban regulations machine-readable and integrating them with modern design workflows. The target is bold, cutting permit timelines from more than two years to roughly six months, and Corro tied it directly to public value, because faster permitting means faster housing in a city that badly needs it. 

It was Corro who offered the panel’s most quietly radical reframing. The era of the “smart city,” he suggested, is giving way to something he called the “agentic city.” A smart city is still fundamentally about technology like sensors, infrastructure, and dashboards. An agentic city is about the citizen, using AI to deliver value in people’s own language and context, including the bureaucratic complexity and administrative jargon that previous digital tools could never quite tame. His test for any innovation was disarmingly simple, and it doubled as a rebuke to a decade of technology-first urbanism: if it does not work for “Señora María”, the ordinary resident, then it does not work at all. 

What emerged across the hour was less a debate than a quiet consensus, hardened by experience. The panelists had all, in different jurisdictions and at different scales, learned the same lessons the hard way. AI is not a strategy you announce. AI is a capability you build, on foundations most governments have not yet laid. The technology is rarely the hard part. The data, the procurement, the workforce, the institutional courage to start with a problem rather than a product, that is where transformation lives or dies. 

Carlos Santiso, closing the session, drew the threads together into the framing that had opened it. Governments are no longer asking whether to use AI. They are confronting the much harder question of how, and the answer, this panel made clear, has remarkably little to do with the models themselves. 

Key Takeaways 

• Start with the problem, not the technology. The most common failure mode in government AI is announcing a solution before clearly defining the service failure or citizen need it is meant to address. Reframing an “AI pharmacist” as a fight against preventable prescribing harm changes everything, including public trust. 

• Transparency is a tool, not a risk. Publishing code, evaluations, and algorithmic transparency reports does not expose governments to danger; it builds public trust and improves the work by inviting outside scrutiny and contribution. 

• AI cannot fix a broken process. Much of real transformation is process re-engineering and procedural simplification, work that is not digital at all. The goal should be “silent AI,” so well integrated that citizens and public servants stop thinking of it as separate. 

• Data readiness comes first. If patient or citizen data is fragmented across dozens of incompatible systems, or still collected on paper, AI ambitions are premature. The unglamorous work of data architecture is the precondition for everything else. 

• The best use cases are already visible. Cities do not need to search abstractly for AI applications. The highest-value opportunities are embedded in the hardest public problems, housing, planning, mobility, where AI can deliver measurable improvement in citizens’ lives. 

• Move from the “smart city” to the “agentic city.” The shift is from technology-centred infrastructure to citizen-centred service, delivered in people’s own language and context. The test for any innovation is simple: if it does not work for the ordinary citizen, it does not work. 

• Keep humans in the loop – and in control. Across every example, AI was framed as a tool to augment public servants and expand their capacity, not to replace their judgment or accountability.