Sovereign models, imported uses?

I am just back from a European gathering of the Echoing Green Fellows community, and the shift is visible to everyone, from New York to Dar es Salaam. Conversations with funders are moving, everywhere, from questions of impact to questions of AI strategy. This is not just an impression: studies converge, the vast majority of the sector's funders now expect AI to have a transformative effect and look at applications through that lens [1], even though very few concretely support their grantees in its implementation [2]. Demand precedes support, which is never good for structuring a sector. More seriously, it seems that demand precedes need. Only public actors, governments and institutions, seem to be moving at a different speed, with more circumspection; I will come back to this, because this asymmetry is at the heart of what follows.

Social entrepreneurs, precisely because they know their field, have not all concluded that AI was the answer to their problem. Many have solutions that work, deployed, producing measurable results, and that owe nothing to AI. If AI becomes a condition for funding, then it is no longer the field speaking. It is the trend.

One might think the question does not even arise for part of the sector: those who serve the poorest or most rural populations, presumed to be beyond AI's reach. That is only partially true: about one in six working-age people in the world already uses generative AI [3]. Although the gap between the Global North and the Global South is widening [3], the barrier to entry can fall through the distribution choices of the providers themselves. The arrival of free models in 2025 showed it: DeepSeek, by removing access costs (credit card, subscription), saw its usage in Africa reach a level estimated at two to four times higher than in other regions [4]. Meta follows the same logic by embedding its AI assistant directly into WhatsApp, an app with near-universal adoption among internet users in many African countries [5], precisely because it is light on data, works on the cheapest phones, and gets around literacy barriers thanks to voice notes. AI is already in these populations' pockets: not because an actor from the social sector decided it, not because a government chose it, but because solution providers, from California or from China, made distribution choices.

This is the real landscape: very large populations reached by AI without any of the actors who know their needs having taken part in its design, starting with these populations themselves. The question "AI or no AI" in the social sector still deserves to be asked, and nothing prevents us from designing for other trajectories. But in practice, an answer has already been given, and it was given without us. Hence the question that concerns me here: who decides how AI gets deployed, how will the social entrepreneurship sector make its own choices about which models to use, and for what reason?

What funders ask for, what the field says

It was a conversation at this gathering that best brought this tension home to me. A social entrepreneur from Dar es Salaam, whose investors, as everywhere, are asking her to integrate AI into her solution. Her case is all the more interesting because the use case genuinely exists. She operates a mental health emergency hotline: AI could automate the redirection of calls to the right resources, or enable an initial self-assessment without mobilizing a psychiatrist, a scarce resource. It is potentially useful and could have a real impact.

And yet, the same question troubles her: is that why funders want AI in her solution? Did they look at her problem and conclude that AI answered it? Or are they simply following the movement, so they can write "AI-powered" in their annual report?

The nuance matters. When funding precedes need, the burden of proof is reversed: it is no longer up to the technology to demonstrate its usefulness for the field, it is up to the field to make itself compatible with the technology. And in doing so, without meaning to, we push social innovators into becoming mere consumers of technology rather than designers. Those who know the problem on the ground end up integrating solutions designed elsewhere, for other reasons.

This misalignment is not just a matter of principle. It can put social entrepreneurs in a genuinely untenable position. Because in Dar es Salaam, while investors are pushing AI, the government partners of these same entrepreneurs are holding back. The Tanzanian government does not want AI in the solutions it adopts without having had the time to conduct its own assessment. The question of sovereignty is crucial, and the point is not to move too fast. The social entrepreneur therefore finds herself caught between a funder who makes support conditional on adopting a technology, and a public partner on whom deployment depends, and who does not want it. Following the funders' trend does not just dilute one's mission: it can sometimes undermine one's very capacity to operate.

Governments between sovereignty and the fear of being left out

This governmental caution deserves attention, because it is more complex than mere administrative conservatism. From my conversation in Dar es Salaam to the public debate running across the continent, the same double movement keeps coming back. On one side, governments want to stay in control: to evaluate, understand, and decide for themselves what enters their public services. On the other, a nagging worry: is Africa, once again, going to be left out? After the industrial revolutions, after the rise of the digital giants, should this opportunity be allowed to pass in the name of caution, or should one board the train as fast as possible?

This dilemma is not confined to the African continent. I have seen it in Europe for years, and from up close: European governments are asking themselves exactly the same questions, torn between the will not to depend on technologies designed and operated elsewhere, the fear of falling behind, and the need to analyze precisely what the field actually requires.

I have had the opportunity to take part in these reflections. At SoScience, the digital solution we developed in recent years was aimed in particular at public actors (research institutions in several European countries, government agencies and research funding bodies, European institutions). In our commercial conversations, two questions came up systematically, before we even talked about features: where is the data hosted, and which models do you use? The fact that our stack was built on Mistral AI, a European solution, was not a technical detail. It was a fundamental argument in those exchanges.

While the answer to these two questions would determine what happened next, they were only an entry ticket. Knowing where the servers are and where the model comes from opened the conversation, but the buyers already had in mind a set of values and a position on how AI tools should be used. It was in the exchanges with these public actors about the why of an implementation, and its fit with a need in the field, that everything was really decided.

Sovereignty does not stop at the model

Because what most caught the attention of our public interlocutors, and in particular of the European Commission teams in charge of implementing AI solutions, was not only the model. It was the way our solution used it, the agents we had chosen to specialize, and what that implied for real-world uses.

Our tool helped researchers think through their funding applications, and above all did not write them in their place. Specialized agents asked the key questions, the ones that determine the quality of a research project, and the answers had to be produced by the researchers themselves, forcing them to dig deeper and push their thinking further. The AI structured, reformulated, challenged. It did not write from A to Z. We made this design choice for substantive reasons: a research project whose intentions have been generated by a machine is not a good project, it is an empty project with better packaging. Researchers are the ones best placed to produce the answers on substance, and no writing should take place until the AI has the elements it needs to feed its work.

What we showed in a fifteen-minute demo was spectacular: prospects were impressed by the product, immediately saw the time saved and the quality of the output. I understand that enthusiasm, demos of AI tools are among the most striking there are, and we lived through those moments when they win people over on the spot. But what the screen did not show was the essential part. Our demonstrations were always built around a project already solidly thought through by its researcher, because that was our design: an AI in the service of thinking that was already there. We explained it every time, and yet a degree of mystification remained. What prospects remembered was the speed and the output, rarely what made them possible. A demo shows a result, never a conception of use.

What struck me all the more was how much our design choice resonated with European public actors. They saw in it a conception of use that matched values they had not necessarily articulated consciously: an AI that augments humans' capacity to think rather than substituting for it. There is something here, I believe, that could be called a European ethics of implementation. Not just European models on European servers, but a certain idea of what AI should do, and above all of what it should not do in our place.

On the terrain of infrastructure sovereignty, Mistral AI has serious assets, and I witnessed this firsthand: for European public actors, they are today the natural interlocutor. The company knows it, and has even launched the "AI for Citizens" initiative to build on this positioning. For now, this initiative covers partnerships with States, ministries, public agencies, centers of excellence within governments. In other words, for now, it is mostly a vision of "AI for governments". It is a necessary step, but without the counterpoint of the field, it is a risky one. The economist Mariana Mazzucato documents this risk in her latest book [6]: reforms and innovations designed without those they claim to serve end up not being recognized by them, and this lack of recognition deepens the fracture between institutions and citizens, to the point of making transformations impossible. For her, public trust is not restored through better announcements, but through relationships of reciprocity in which solutions rise up from the field rather than come down to it. Applied to AI in public services, the thorniest problem is not modernizing the services delivered to citizens with AI solutions, but involving them in the design of use through processes put in place by public authorities.

Today, where are the citizens in the deployment? Where are the social entrepreneurs, the workers' representatives, the field practitioners, all those who know the problems these transformed public services are supposed to solve? They appear as end beneficiaries. They do not appear as co-designers.

Yet what my recent exchanges and my past experience have shown me, from Dar es Salaam to Brussels, converges on the same observation: the essential moment is the design of use. That is where it is decided whether the use will be adopted by beneficiaries with a real benefit for them, whether intermediary actors (including entrepreneurs offering field solutions) will be able to build coherent strategies, and ultimately whether governments will be convinced of the proposal's usefulness. Yet today, these intermediary bodies are largely absent from this design phase, caught between funders pushing a technology and governments receiving it.

The sovereignty we talk about so much cannot be reduced to the location of servers or the nationality of models. It includes something more fundamental: the capacity to decide collectively what this technology should do, with those who live the problems it claims to solve, and for real reasons rather than by imitation. A sovereignty from the ground up, co-constructed, that does not stop at "which AI?" but asks, above all, "what for?".

In Europe, we can build around a sovereign model. That is a real victory, and it was far from guaranteed. But if the uses themselves remain imported, designed far from the field, driven by funding trends rather than by needs, what will we really have won?

[1] Bonterra, "The AI readiness path: Key insights for nonprofits and funders" (91% of funders anticipate a transformative effect of AI on philanthropy and grantmaking): https://www.businesswire.com/news/home/20251110626529/en

[2] The Center for Effective Philanthropy, "AI WITH PURPOSE: How Foundations and Nonprofits are Thinking About and Using Artificial Intelligence" (87% of nonprofits say their funders do not understand their AI needs, and nearly 90% of foundations provide no AI implementation support to their grantees): https://cep.org/wp-content/uploads/2025/09/CEP_AI_Layout_FINAL.pdf

[3] Microsoft AI Economy Institute, "Global AI Adoption in 2026: Q1 2026 Trends and Insights" (17.8% of the world's working-age population uses generative AI as of the first quarter of 2026): https://blogs.microsoft.com/on-the-issues/2026/05/07/the-state-of-global-ai-diffusion-in-2026/

[4] Microsoft AI Economy Institute, "Global AI Adoption in 2025: A Widening Digital Divide" (DeepSeek usage estimated at 2 to 4 times higher in Africa than in other regions): https://www.microsoft.com/en-us/research/wp-content/uploads/2026/01/Microsoft-AI-Diffusion-Report-2025-H2.pdf

[5] Yazi, "WhatsApp Usage Across Africa: Key Insights for 2025-2026" (penetration rates of 95 to 97% among internet users in Kenya, South Africa, and Nigeria): https://www.askyazi.com/useful-data-sources-for-africa/whatsapp-usage-across-africa-key-statistics-insights-for-2025

[6] Mariana Mazzucato, "The Common Good Economy: A New Compass", Allen Lane, 2026: https://marianamazzucato.com/books/the-common-good-economy/