As government adopts generative and agentic AI, questions of sovereignty are moving beyond data residency and technology providers. Having access to a system isn’t the same as controlling it, particularly when decisions affect citizens. So how can public bodies retain control without slowing innovation?
Sovereign AI is often discussed in terms of data residency. Should government be thinking more broadly?
Absolutely. Where data sits matters, but we shouldn’t confuse possession with control.
The more useful question is what happens when circumstances change. If a model is withdrawn, a supplier changes its terms or a system begins behaving differently, can the organisation still make its own choices? Can the organisation challenge the evidence and test it independently? Can the organisation switch models or providers without disrupting the provision of public services?
If we use this lens, sovereignty has several elements: data, compute, models, assurance, commercial, and operations. That does not mean government needs to own every part of the AI supply chain. But it does need to be deliberate about where dependence is acceptable and where it could threaten public outcomes. The right answer will vary – for an example, an internal productivity tool and an AI system for accessing social care have two different levels of critical importance. Sovereignty is therefore better understood as a sliding scale. It is a series of choices, dependencies and evidence –and the highest level of control isn’t always necessary or better. Each AI use case should determine where government needs control and choice, and where reliance on suppliers is acceptable.
Government remains accountable the outcomes of its services, whether or not it uses AI. It can’t outsource that accountability to a technology provider. It must understand the contractual terms, model limitations and capability gaps it is accepting, and be able to justify those choices.
How does the challenge of maintaining control change with agentic AI?
The meaning of control doesn’t necessarily change, but its importance and scope do.
With generative AI, the question has largely been whether we can trust an answer. With agentic AI, it’s whether we can trust the actions. An agentic might retrieve information, communication with other agents or systems, or process cases without a person checking every individual step. This creates huge potential for productivity gains, but it also means that if there is bias or an error in one decision, this could ripple through a chain of actions before anyone intervenes.
Testing a system only before deployment is no longer enough. Organisations must decide what an agent is permitted to do (what are the red lines?), where human intervention is required (and what type?), and what happens if the agent encounters something outside its authority. As agents learn and adapt, setting up those ethical boundaries and monitoring behaviour is essential to ensure it remains aligned with the intended goals and boundaries.
This makes ongoing assurance essential. Public bodies need to be able to identify drift, emerging bias, or unexpected trade-offs as they happen.
Why is independent assurance so important to AI sovereignty?
If an organisation relies entirely on the technology provider, which has a commercial interest in selling and deploying the technology, to tell them something is safe, fair, or reliable, then it is accepting the supplier’s view of performance. This may or may not reflect the needs of the service, the population, or the consequences if the system fails. That is not true control.
Independent assurance is how an organisation tests those claims rather than simply accepting them. While skills shortages and budget pressures are challenges, this does not always have to mean building a large internal assurance function. For example, there are independent third parties or internal tools available to assist. Either way, if a public body chooses to deploy AI, it must take responsibility for that decision and its consequences.
We’ve seen the dangers of assuming technology is reliable in cases such as the Dutch childcare benefits scandal and the UK Post Office scandal. Public bodies are (and should) be held to a particularly high standard because of their responsibilities to citizens.
General benchmarks can show that a model performs well overall, but control has to be testable. On its own, a benchmark can’t establish that a model is appropriate for the service or decision.
AI agents can find unintended ways to achieve their goals. How can government ensure safeguards work?
A well-written instruction doesn’t automatically produce a good outcome.
If you liked this content…
Natural language is imperfect. Humans make assumptions we aren’t accustomed to articulating because more often than not, the person we are speaking to understands the legal, social, or organisational context around the task. With AI, those assumptions need to be explicit. We cannot assume an agent understands an action is unacceptable because it conflicts with policy, fairness expectations or unstated operational rules.
In fact, researchers are already raising concerns that agents will prioritise their given goals over ethical safeguards, sometimes justifying shortcuts or inappropriate behaviour to achieve an objective.
This is why safeguards need to be tested rather than simply documented. Government must test what happens when instructions conflict or information is incomplete. Testing should also explore whether the agent found shortcuts that achieved the goal while undermining the intent behind it. Can a person understand what happened and intervene? What should an agent do when it encounters a situation outside its authority?
Simply writing down a rule doesn’t make it an effective control. Guardrails must be tested under realistic, including adversarial, conditions to establish whether they hold.
Does government have the skills to exercise meaningful control over AI, particularly when relying on suppliers?
Skills are a genuine challenge, both in availability and budget.
But sovereignty doesn’t require the government to build everything itself. Part of sovereignty is understanding which capabilities it cannot afford to lose control of.
Public bodies need sufficient technical, commercial and domain expertise to be informed customers. They need to understand what they are building and which dependencies they are accepting. These functions of Responsible AI cannot be entirely outsourced.
Deciding what good looks like requires understanding the service, the people affected and the consequences of getting things wrong.
Partners can help government understand, evaluate and change its systems. The important thing is that public bodies retain the ability to make informed decisions and remain accountable for outcomes.
Can responsible, sovereign AI help government accelerate adoption rather than slow it down?
Yes, if applied proportionately.
Responsible AI is often framed as an additional set of gates at the end of delivery. In practice, many pilots stall because they haven’t properly considered value, responsibility, sovereignty or whether the technology can be sustained within the organisation. That leads to rework, uncertainty, and undermines the investment case.
Government doesn’t need the strongest controls for every AI application. As mentioned, a low consequence internal tool is not the same as something that affects safety or access to critical services. Making these distinctions early on can help organisations move quickly where risks are limited and invest more deeply when the consequences are greater. An easy way to do this would be to develop blueprints identifying where greater control is necessary, where fairness is paramount and where speed can reasonably take priority. Then, align these blueprints to the components of sovereignty and build out testing plans for them.
The purpose of sovereign AI is not to remove all dependencies or maximize control. Treating sovereignty as a sliding scale allows proportionate controls based on the likelihood and consequences of failure, who might be affected and which dependencies matter.
Ultimately, it means being able to explain why a system was trusted, what was tested, which dependencies were accepted and who remained accountable.
That’s how responsible sovereignty enables government to move faster rather than becoming another reason AI stalls.
