From vision to reality: five lessons from advising governments on AI

By Pablo Fuentes Nettel

Over the past few years, I’ve had the opportunity to engage with governments across different regions and support them in their AI journeys. I’ve worked with national and subnational public entities across Latin America, Africa, Europe, Central Asia, and the Middle East. 

The contexts vary significantly. Some governments operate with strong institutional capacity, advanced digital ecosystems, and clearly defined AI ambitions. Others are still focused on strengthening foundational capabilities such as digital infrastructure, institutional coordination, or data governance. Political systems, fiscal realities, and strategic priorities also differ considerably from country to country.

Yet despite these differences, many of the same challenges emerge repeatedly when governments try to move from ambition to implementation.

In this blog, I reflect on five lessons that have consistently shaped how I think about government AI adoption and governance. These are not universal truths, nor an attempt to generalise across very different national contexts. Rather, they are observations and recurring questions that have emerged through advisory work across multiple governments and regions. My hope is that they help spark useful conversations and reflection for others working in this space.

1. Governments need more “midfielders” between policy and technical teams

When I tell people I work in the AI space, the next question is often: “Oh, so are you a developer or an engineer?” I usually respond with a slightly awkward, “Not really, I focus more on the governance and policy side of AI,” which is often followed by a polite look of confusion and mild disappointment.

Beyond the personal anecdote, I think this reflects a broader reality governments are also facing.

One of the biggest capability gaps I’ve observed across governments is not necessarily a shortage of technical talent. It is the shortage of people who can operate between technical, policy, legal, and political worlds.

AI adoption requires socio-technical capacity.

When thinking of human capital, governments often focus heavily on engineers, developers, data scientists, and technical specialists. Those roles are essential. But successful AI adoption also depends on people who can translate between disciplines and institutions. People  such as AI governance specialists, AI safety experts, delivery leads, business analysts, user researchers, or service designers.

These “midfielders” understand technology well enough to engage with technical teams, while also understanding governance, regulation, institutional processes, public service delivery, and political priorities. Crucially, they also understand how government actually works in practice, which is often very different from how technology teams expect it to work.

Their functions are often overlooked, particularly in countries with less mature tech ecosystems. Yet they are critical for enabling technical teams to operate within environments that are effective, trustworthy, and aligned with public-sector realities.

In many ways, these translation layers are what allow AI adoption to move from isolated pilots to sustainable institutional capability.

2. AI strategy is rarely the hardest part. Coordination and continuity are

Most governments no longer struggle to articulate an AI vision. Many already have strategies, roadmaps, principles, or national ambitions in place. The harder challenge is operationalising them across government.

AI adoption cuts across ministries, regulators, digital agencies, procurement systems, and sectoral institutions. In practice, responsibilities are often fragmented, ownership is unclear, and institutions move at different speeds under different incentives.

One approach I’ve found particularly effective is the establishment of specialised coordination functions at the centre of government. These can sit within a presidency or prime minister’s office, or take the form of cross-government AI councils or boards bringing together key ministries and public institutions. Even relatively small teams, focused primarily on coordination and strategic oversight, can significantly improve coherence across government.

Importantly, these structures do not replace dedicated ministries or implementing agencies. They serve a different purpose. One focuses on cross-government coordination and priority setting, while the other focuses on execution and delivery. When this division of labour works well, governments tend to move faster and with greater consistency.

Another recurring challenge is continuity. In many countries, AI initiatives still depend heavily on a small number of highly motivated individuals inside government. When those champions move roles, political leadership changes, or institutional priorities shift, momentum can slow quickly. The governments making the most sustained progress are not necessarily those with the most ambitious strategies. They are often the ones building institutional continuity around AI adoption.

There are different ways to do this. Some governments are beginning to develop specialised career pathways for civil servants working on AI and digital, such as the UK’s Digital and Data Profession Capability Framework (formerly DDaT). I’ve also seen growing efforts to embed AI expertise beyond isolated innovation units, helping foster a broader culture where AI becomes a cross-government capability rather than the responsibility of a single team.

Over time, these institutional foundations matter far more than any individual strategy document.

3. AI readiness is not a linear journey

One of the biggest misconceptions I’ve encountered is the idea that all governments are climbing the same ladder at different speeds. In practice, countries are often pursuing fundamentally different objectives.

I increasingly think of AI readiness less as a sprint towards a universal goal and more like a mass marathon, where success can mean very different things depending on the runner. As much as I personally enjoy distance running, I know I cannot compete like an elite athlete. My objectives, training, and pacing strategy are shaped by my own context, capabilities, and goals. The same applies to governments.

Countries face fundamentally different choices. From their capabilities and institutional models to their economic advantages and grand strategy, governments pursue very different goals within the global AI ecosystem.

Some governments aim to position themselves as AI research and compute hubs. Others prioritise sectoral adoption or digital public infrastructure. Some focus on becoming trusted adopters rather than frontier developers. The UAE, for example, has invested heavily in compute infrastructure as part of a broader ambition to position itself as a global AI hub. This strategy is also supported by access to land and comparatively cheap energy, which strengthens the country’s ability to scale compute infrastructure. Estonia, meanwhile, has focused more heavily on developing and, crucially, exporting digital public infrastructure and public service models, through products like X-Road.

These choices matter because AI readiness is not simply about maximising performance across every dimension simultaneously. It is about aligning priorities with national objectives and comparative advantages.

Importantly, this does not mean that comparative benchmarks or readiness frameworks are unhelpful. On the contrary, they are crucial for helping governments identify strengths, gaps, and areas for investment. But they are most valuable when used as tools for strategic reflection rather than universal templates. This is precisely why, a few years ago, we launched the Spotlight Series as a complement to the Government AI Readiness Index. The series showcases case studies from different governments, illustrating that there is no single pathway to AI readiness and that countries are pursuing very different approaches based on their own priorities and contexts.

I’ve increasingly come to see AI readiness as a strategic question rather than a linear checklist. This is particularly important for developing countries, where attempting to replicate the AI strategies of much larger economies is often neither realistic nor desirable.

Ultimately, governments need to begin with a more fundamental reflection: what role do we want to, and realistically can, play within the global AI ecosystem? Are our ambitions compatible with our institutional capacity and economic structure, or do they require deeper reforms? Are we particularly competitive in a specific part of the AI value chain? Which use cases are most relevant given our most pressing public challenges?

These questions are not peripheral to AI readiness. They are the foundation of it.

4. AI adoption is layered. Governance needs to reflect that

Another recurring challenge I’ve observed is that governments often approach AI adoption as a single policy issue. In practice, adoption happens at multiple levels simultaneously, and each layer requires different governance mechanisms.

At one level, governments are pursuing large-scale ambitions such as moonshot projects, national platforms, or AI-enabled public services. At another, public institutions are navigating much more operational questions around the everyday use of AI inside government itself.

The problem is that governance efforts are often concentrated around the first layer.

We are seeing many governments develop AI legislation, risk classification systems, or red teaming frameworks. These mechanisms are critical. Yet some governments are still overlooking more foundational operational questions. I’ve seen governments advancing sophisticated AI governance approaches without having gone through a serious institutional reflection on much simpler but equally important issues. For instance: Can civil servants use large language models? Which tools are authorised? What information can be shared safely? How should governments procure AI systems responsibly and assess value for money?

These questions may appear less strategic than national AI ambitions, but they are often far more immediate for public institutions.

Some governments are beginning to recognise this gap. Canada, for example, has developed the AI Strategy for the Federal Public Service 2025-2027. Similarly, we worked with government counterparts in Uzbekistan to develop an AI Use Guide for Civil Servants. The document provides practical guidance on the use of generative AI and helps public officials navigate key implementation questions, such as whether AI is actually needed, which solutions are most appropriate, and whether the necessary data foundations are in place.

These approaches matter because they recognise something important: AI adoption is not only about large-scale transformation projects. It is also about creating the institutional capability and governance mechanisms needed for AI to be used responsibly across everyday government operations.

5. AI sovereignty is about agency, not ownership

Conversations around AI sovereignty are becoming increasingly central to government AI strategies. But sovereignty is sometimes interpreted too narrowly as full ownership of infrastructure, models, or compute.

In practice, sovereignty is often more about agency, resilience, and strategic control.

For many countries, especially those with smaller domestic ecosystems, attempting to fully own every layer of the AI stack may not be realistic. A narrow focus on ownership can also distract from other important policy options and more immediate adoption opportunities.

Some Gulf countries, for example, have prioritised building the capability to maintain strategic control and operational autonomy over critical digital infrastructure, even when working closely with foreign technology providers. Digital embassies are a good example of the kinds of policy levers governments can use to strengthen resilience and maintain continuity without necessarily owning every layer of the technology stack.

Ukraine offers another important example. Despite relying on external technology providers and cloud infrastructure, the government retained sufficient strategic control and operational capacity to rapidly migrate critical government data outside the country following Russia’s invasion. This helped strengthen resilience and ensure continuity of government operations during wartime.

I’ve also observed this tension emerge in discussions around sovereign large language models (LLMs). Many governments are understandably interested in developing domestic models to address language representation gaps or reduce dependence on foreign technologies. However, in some cases, this ambition can overshadow opportunities to leverage and adapt existing models that may already serve similar purposes.

Spain’s ALIA initiative is a good example. Beyond its national relevance, models like ALIA could provide valuable alternatives for other Spanish-speaking countries seeking stronger linguistic and cultural representation in AI systems without necessarily developing large language models entirely from scratch.

This is particularly relevant for countries with more limited compute capacity or technical ecosystems. In these contexts, adapting, fine-tuning, or safely deploying existing open-source and commercial models may sometimes offer a more practical and cost-effective pathway towards AI adoption and sovereignty.

Ultimately, strategic sovereignty is not necessarily about building everything domestically. It is about ensuring governments retain meaningful agency, resilience, and decision-making capacity over critical technologies and infrastructure.

Let’s continue the conversation

Ultimately, working across governments has reinforced for me that AI adoption is far less about isolated technologies and far more about institutions, coordination, strategy, and people. The governments making the most meaningful progress are not always those with the largest budgets or the most ambitious announcements. They are often the ones asking difficult strategic questions early, building institutional capability patiently, and finding ways to translate ambition into operational reality.

At the same time, there is still no clear playbook for what successful government AI adoption should look like. Countries are experimenting in real time, often under very different political, economic, and institutional conditions. That is precisely why comparative learning and open conversations matter so much.

These reflections are not intended as definitive answers, but rather as observations from ongoing work across different regions and governments. I’m sure many others working in this space will have seen similar patterns, or completely different ones.

If any of these ideas resonate with your own experience, I’d genuinely love to continue the conversation. Feel free to reach out at pablo.fuentes@oxfordinsights.com or connect with me on LinkedIn (pablofunett).

Grid image by Yael Gonzalez on Unsplash