AI sovereignty depends on who controls public decisions
National AI strategies must preserve the authority that future governments will need to exercise.
By Farrukh Zafar
By the time a public decision reaches a senior official’s desk, AI may already have shaped the options under consideration. Systems that rank applications, forecast outcomes, and summarize evidence influence which choices receive attention. The official’s signature comes at the end of a process in which consequential judgments may already have occurred.
Sovereign AI strategies need to address how governments exercise authority.
What AI Sovereignty Means
AI sovereignty is a country’s ability to decide which AI systems it will use, how they will shape public decisions, and under what terms, while retaining the capacity to revise those choices in accordance with its laws and public priorities. Its foundations include reliable computing infrastructure and energy, protected and representative data, models suited to local languages and needs, and people capable of developing, evaluating, and operating these systems.
National leaders face substantial choices about which capabilities to build domestically, which to obtain through international partnerships, and which dependencies they can accept. Domestic investment can strengthen control and local expertise. Partnerships can provide access to capabilities that would be costly to develop independently.
These choices are reflected in the European Commission’s proposed Cloud and AI Development Act, which addresses infrastructure capacity, strategic dependencies, and public procurement. The OECD’s Digital Government Outlook 2026 also emphasizes procurement expertise and the internal capabilities governments need to oversee AI suppliers.
A Strategy Built on Public Judgment
A sovereign AI strategy must connect those investments to the exercise of public judgment. Governments need to establish which decisions AI may influence, who remains responsible, and how institutions will determine whether their processes continue to serve the public mandate.
These arrangements form an institution’s decision architecture: how information becomes a judgment, how that judgment leads to action, and who has the authority to intervene. Introducing AI changes this architecture whenever a system influences what receives scrutiny, what evidence officials see, or which judgments are delegated. An institution can use accurate predictions in ways that conflict with its public mandate.
Canada’s Visitor Record Tool
Canada’s immigration system provides a concrete example. According to its immigration department, a tool used to process visitor-record applications combines rules developed by officers with rules generated through machine learning from previous cases. It can approve eligibility for routine applications; other cases receive an officer’s eligibility review. The tool cannot refuse an application or recommend refusal.
An officer assesses admissibility, makes the final decision, and can overturn the tool’s eligibility determination. The department also retains control over changes to the rules. These are the safeguards described in its public documentation. Canada’s approach assigns specific judgments to the system and reserves others for officials. Leaders evaluating such a system need to examine both the decisions assigned to it and the capabilities officials retain.
When Human Review Falls Short
A requirement for human approval offers limited assurance when a reviewer cannot examine the basis for a recommendation, lacks the expertise to challenge it, or has too little time to do so.
The Dutch childcare benefits scandal illustrates the national stakes of ineffective oversight. Amnesty International’s investigation identified a risk-classification model with self-learning elements that used nationality as a risk factor. Officials reviewed flagged applications without being told why the model had assigned a high score. The model formed part of a failure of policy, administration, and oversight that harmed thousands of families and led to the cabinet’s resignation in 2021. The lesson for national leaders is that human review must include the information and authority to challenge how cases are selected and assessed.
Dependence and Lost Capacity
A further sovereignty risk arises if dependence on AI erodes the institutional capacity needed to implement lawful policy. An agency may retain formal authority while staffing decisions, inaccessible evidence, or rigid workflows weaken its capacity to exercise it. A policy change may then require capabilities the institution no longer possesses.
The implications grow when several agencies rely on shared platforms or suppliers. Common infrastructure can improve efficiency and consistency. Policy changes may also require coordinated technical updates, contract negotiations, and scarce specialist knowledge. The resulting costs and delays can constrain what a government can accomplish.
Supplier nationality addresses only part of this problem. A domestically owned system can leave officials dependent on expertise they lack. An international partnership can preserve meaningful control when institutions retain the necessary knowledge, access, and contractual rights. Leaders must assess dependence at the level of the public decisions affected.
What Leaders Should Require
Before expanding AI use in consequential decisions, leaders should require a clear account of how authority is allocated. That account should identify what the system determines, what officials must examine, who resolves conflicts with the public mandate, and who can authorize changes.
Agencies should demonstrate that these arrangements work under normal operating conditions. They should test a plausible policy change and trace its implications through the system. Can staff identify the affected rules, reassess relevant cases, and establish whether the revised process delivers the intended result? Such an exercise makes the demands on expertise, budgets, and suppliers visible before a crisis.
Independent assessment can itself use AI, provided officials can examine the underlying evidence and evaluate it against the public mandate. Reviewers need sufficient independence from the teams or suppliers responsible for the system to challenge its assumptions and conclusions.
Budgets should preserve the specialist knowledge and review capacity these responsibilities require. Contracts should provide workable access to evidence and terms for modifying or replacing systems. Citizens need a route to challenge consequential decisions and obtain a review capable of producing a remedy. These provisions belong in the investment decision. Any immediate savings should be weighed against the cost of recovering institutional capabilities once they have been lost.
National leaders are choosing the decision systems their successors will inherit. The next government’s practical freedom to govern will be shaped by the AI choices made today.
About the Author
Farrukh Zafar
Founder and CEO, EmTech Advisory Group
Farrukh Zafar is the Founder and CEO of EmTech Advisory Group and author of the forthcoming book Overruled: Governing the Intelligent Enterprise (Columbia Business School Publishing, April 2027). He has extensive leadership experience in financial services, with responsibilities spanning AI strategy and governance, technology strategy, innovation management, and data governance. His current work focuses on the governance of AI, advising boards and executives on how to preserve accountability and decision authority as AI becomes embedded in their organizations. He holds an MBA and an MS in Electrical Engineering from Yale University.
Sources
European Commission. “Cloud and AI Development Act.”
OECD. “Adopting and Governing AI in Government,” Digital Government Outlook 2026.
Immigration, Refugees and Citizenship Canada. “Advanced Analytics and Automation for Processing Applications.”
Amnesty International. Xenophobic Machines: Discrimination Through Unregulated Use of Algorithms in the Dutch Childcare Benefits Scandal. October 2021.
Comments