Why Pilots Die: The Unfinished Business of AI in Public Services

10/13/2026
Share this post on Social Media:

A Main Stage panel argued that the failure of government AI is rarely about the model – it is about people, process, data, and the discipline to scale responsibly. 

Every government can run a pilot. Far fewer can turn one into a service that millions of people rely on. That gap (between the promising experiment and the production system) was the subject of this Main Stage panel, and the speakers treated it not as a technical puzzle but as an institutional one. 

Moderator Marina Manzoni, an EU Policy Officer at the European Commission, framed the central challenge plainly: innovation is meaningless if pilots remain isolated experiments. For governments, scaling is not only horizontal expansion, but vertical integration and coordination across the many intermediaries’ public administrations depend on. The question, then, was what makes AI genuinely useful, sustainable, governable, and replicable in real public-service environments. 

The Conditions Pilots Need To Survive 

Kelly Ommundsen, Head of Digital Public Goods, World Economic Forum, opened with a diagnosis that set the tone: most AI pilots fail not because the model is poor, but because the institutional conditions for implementation are missing. Governments focus too heavily on the technology and too little on organizational readiness. The real barriers are people, process, measurement, and governance – the right skills, serious change-management plans, clear problem statements, measurable indicators, and identifiable owners accountable for outcomes. She urged governments to get the foundations right before layering AI on weak systems, and to “productize early” by focusing on repeatable, high-volume tasks where quick wins build trust and momentum. 

Laman Ibrahimova, Head of ASAN AI Hub and Advisor to the Director of the Innovations Center, Government of Azerbaijan, described how Azerbaijan’s ASAN service is incorporating AI into a large-scale delivery system that handles hundreds of services across 28 centers and tens of thousands of people daily. She reinforced a crucial point: return on investment disappoints when organizations deploy AI at the application level without changing workflows. The institutions that extract value are those that redesign work, not those that simply buy tools. ASAN’s concrete uses – voice bots in the call center, AI detecting service errors before complaints arrive, and generative AI reviewing rejected decisions to catch missed legislative updates – all kept humans firmly in the loop, with final decisions under human control.

From Proof-of-concept To Small Production 

Jaume Miralles, Director General of Digital Innovation, Government of Catalonia, zeroed in on why pilots stall. Too many organizations run too many pilots without filtering for purpose, sometimes duplicating the same idea in different units. Every pilot, he argued, needs a defined business purpose, a responsible owner, and a realistic path to production backed by a committed budget – with planning integration, maintenance, training, and operational ownership from the outset. His warning was sharp: if production takes one or two years after a successful pilot, the solution may already be obsolete in a fast-moving AI environment. He proposed shifting away from proof-of-concept culture toward “small production pilots,” especially where mature GovTech solutions already exist, because in many cases the problem requires not invention but the intelligent adoption of existing products from smaller national or European providers.

Antonio García de la Paz, Deputy Director General for Data and AI, Madrid Capital Digital, explained how Madrid City Council has organized AI governance, combining innovation, implementation, and governance in a single structure. Madrid published an AI strategy in 2022, then had to rework it as generative AI and new European regulation arrived. The city created an internal AI group spanning more than 80 people, legal, data protection, cybersecurity, and operational stakeholders, to coordinate initiatives, gather use cases, and move projects from idea to production. He offered the example of an internally developed document-anonymization system replacing reliance on commercial tools, and stressed caution around AI embedded in vendor systems and AI that makes decisions directly, noting Madrid is not yet ready for unrestricted deployment of such uses. Documentation, talent, impact measurement, and cost awareness, he argued, must cover the full lifecycle of any system. 

The Economics of Lock-in 

Tiago Carneiro, Senior Digital Development Specialist, World Bank (Serbia), brought a global evidence base, describing an analysis of roughly six million procurement notices worldwide. The striking finding: actual government procurement of generative AI remains limited, even as surveys show many institutions claiming to run pilots – suggesting much experimentation without the financial and institutional commitment real adoption requires. His sharpest warning concerned vendor lock-in through “freemium” entry points. A vendor may offer a chatbot or lightweight service at low initial cost, but once it gains users, absorbs context, and embeds in workflows, it becomes hard to replace. He added that current inference costs are likely artificially low because they are subsidised; when subsidies recede, governments, unlike private firms that can switch suppliers, may find themselves stuck in expensive early choices. Pilot culture, left unmanaged, can harden into structural dependence. 

Sovereignty As a Continuum 

Miralles returned during the sovereignty discussion to argue that European governments need the ability to host local models for inference, even if training happens elsewhere. Many routine tasks – anonymization, data extraction, translation, can increasingly run on compressed or open-source models on affordable local infrastructure, reserving token-intensive external services for complex tasks. AI, he suggested, may even shift the traditional “buy versus build” equation, making it easier for governments to adapt open-source tools internally, provided they invest in talent and maintain auditability and governance. 

Ommundsen closed by reframing digital sovereignty as a practical continuum rather than a binary. Sovereignty does not mean owning the entire stack. It means making deliberate choices across governance, model selection, infrastructure, and operations to preserve autonomy, resilience, and public interest control. She cited a Canadian definition emphasizing a state’s ability to manage and protect its systems and data regardless of where the underlying technology is developed or hosted. Full-stack sovereignty may be unrealistic given the capital demands of frontier infrastructure; the answer is “sovereignty by design” – modular architecture, interoperability standards, strong governance, multi-vendor strategies, and hybrid build-and-buy approaches, targeted at the layers where sovereignty matters most. Manzoni’s synthesis was fitting: moving from pilot to scale requires interoperable, modular, replicable solutions, precise governance, and clear commitments of money, people, and risk. Success is measured not by how many pilots are launched, but by whether governments can turn promising use cases into trustworthy, sustainable, and sovereign public services. 

Key Takeaways 
  1. Pilots fail on institutions, not models. The barriers are people, process, measurement, and governance, not technology. Get the foundations right before layering AI on weak systems. 
  2. Value comes from redesigning workflows. Buying AI tools without changing how work is done disappoints; the institutions that benefit redesign the process around the technology. 
  3. Move from proof-of-concept to small production. Every pilot needs an owner, a budget, and a path to production, and often the answer is adopting an existing solution, not inventing one. 
  4. Watch for freemium lock-in. Low-cost entry points embed in workflows and become costly to replace, especially as subsidized inference prices rise, a structural risk for public budgets. 
  5. Sovereignty is a continuum, by design. It means deliberate choices across governance, infrastructure, and operations, modular, interoperable, multi-vendor, not owning the entire technology stack. 
Share this post on Social Media:

Read more from our Insights Hub...