For years, enterprises have been told that data is their most valuable digital asset. That belief shaped a generation of technology strategy.
Companies built data platforms, migrated workloads to the cloud and created governance frameworks around where information is stored, who can access it and under whose jurisdiction it sits.
Artificial intelligence does not make those questions obsolete. It introduces another one: Who owns what the organisation learns?
Microsoft CEO Satya Nadella recently described an emerging problem he calls the Reverse Information Paradox. Building on economist Kenneth Arrow’s famous information paradox, Nadella argues that companies buying AI can find themselves paying twice: first for access to intelligence and then by supplying the proprietary knowledge required to make that intelligence useful.
The buyer now brings valuable knowledge to the transaction.
What makes a general model valuable?
A foundation model may understand refineries, but it does not understand the accumulated operating experience of a particular refinery. It may understand procurement, but not an organisation’s supplier history, exceptions and risk appetite.
A model may understand banking, mining or manufacturing, but it does not arrive knowing the thousands of small judgements through which experienced employees make those businesses work.
Consider an engineer correcting an AI recommendation, or an operator overriding an agent because conditions on a plant make its recommendation unsafe. A business unit repeatedly scoring one output higher than another creates a record of what the organisation values.
Enterprises are spending enormous energy debating which foundation model will win. That may matter less than we think.
These interactions create something different from source data. They create learning.
Prompts reveal context. Corrections reveal expertise. Evaluations define what “good” means. Agent traces reveal how work gets done. Collectively, they begin to encode institutional judgement.
Data sovereignty is not enough
Digital sovereignty has largely focused on information: Where is the data stored? Which country’s laws apply? Who can access it? Can it leave a particular jurisdiction?
Those remain legitimate questions. But imagine an enterprise that protects its underlying datasets while building its AI capability on evaluations, agent memory, workflow traces and model adaptations that it cannot independently control or move. Its data may be sovereign. Its learning may not be.
This is not abstract in South Africa. Picture a mining or heavy-industrial operator that has carefully addressed POPIA’s requirements, including its conditions for transferring personal information across borders, and chosen to keep sensitive operational records in a South African cloud region.
Yet the judgement that keeps that operation running − the corrections an experienced engineer makes when a technically optimal recommendation becomes unsafe, the evaluations defining a good maintenance call on an ageing asset, the agent traces showing how a control room responds, and model adaptations tuned to its equipment − may sit on a platform it neither controls nor can easily move.
South Africa’s financial sector already has explicit regulatory requirements governing banks’ use of cloud computing and offshoring. Mining and heavy industry operate under different regulatory frameworks covering areas such as licensing, environmental compliance, health and safety, and personal-information protection.
Those frameworks were not designed around the emerging question of where AI-generated operational learning accumulates.
A South African enterprise can therefore satisfy its data protection and residency decisions, while still having limited sovereignty over the intelligence its own people are helping create.
This is why enterprises need to start thinking about learning sovereignty: an organisation’s ability to control, retain and transfer the mechanisms through which its AI systems become better at understanding its business.
Data can be resident without learning being sovereign.
Can you localise learning?
This creates a second generation of sovereignty questions. Where do our AI evaluations and traces reside? Who has contractual rights over them? Can they move to another platform? Can we use intelligence generated through our own work to improve models we control?
And perhaps the harder question: Can the learning loop itself be localised?
The answer is more complicated than data localisation. An enterprise can choose where evaluations, traces, memory and feedback are stored. But learning is a process, not simply a dataset. The compute that adapts a model, the weights emerging from that adaptation and the services orchestrating it may operate across multiple platforms and jurisdictions.
Putting data inside a national boundary therefore does not necessarily put the organisation’s learning loop inside that boundary.
Our regulatory and architectural thinking will eventually have to recognise that distinction.
Data, knowledge, learning and judgement
Enterprise intelligence can be considered in four layers. Data records what happened. Knowledge captures what the organisation explicitly knows through policies, standards and documented experience. Learning emerges through corrections, feedback, evaluations, agent traces and repeated decisions.
Then there is judgement − the difficult-to-document institutional capability developed through experience.
It is knowing when an efficient production decision creates unacceptable safety risk, when the strongest supplier on paper creates problems in practice, or when historical data no longer reflects conditions on the ground.
Foundation models provide extraordinary general capability. They do not arrive with decades of an organisation’s accumulated judgement.
This is not simply about model training
There is an important counterargument. Enterprise AI platforms increasingly provide protections around customer information.
Microsoft, for example, states that prompts, responses and data accessed through Microsoft Graph in Microsoft 365 Copilot are not used to train foundation models. Microsoft also distinguishes foundation-model training from other forms of product improvement, including optional user feedback.
That distinction reinforces the point.
“Is my data training your foundation model?” is not the same question as: “Where is the learning generated through my organisation’s use of AI accumulating, and what rights do I have over it?”
The strategic issue is the broader learning environment: evaluations, feedback, interaction histories, agent memory, workflow traces and model adaptations − and whether the enterprise can control, retain and move them.
Sovereignty inside somebody else’s cloud
Nadella’s argument points towards private learning environments, strong tenant boundaries and orchestration across multiple models. Those ideas are compelling, but they also align closely with capabilities large cloud providers increasingly sell.
A private learning environment inside a foreign hyperscaler’s ecosystem may provide excellent security, privacy and operational control. Whether that constitutes full sovereignty is a different question.
Security, privacy, residency and sovereignty overlap. They are not synonymous.
Learning sovereignty does not require enterprises to own every model or operate every piece of infrastructure. It means retaining meaningful control over the assets that create differentiation and credible choices about where those assets run.
Your model is not your moat
Enterprises are spending enormous energy debating which foundation model will win. That may matter less than we think.
Models will improve, prices will change and new competitors will emerge. Competitors can often buy access to the same general-purpose intelligence.
What they cannot easily buy is your accumulated organisational judgement: your corrections, evaluations, operating history and thousands of decisions showing what good performance means inside your company.
The foundation model is the generalist. Your learning environment turns the generalist into a veteran. Change the generalist if necessary. Don’t lose the veteran knowledge.
Boards should ask three questions:
- What are our AI systems learning about how our organisation operates?
- Where is that learning accumulating, and who controls it?
- If we changed AI providers tomorrow, what intelligence would remain ours?
These questions bring AI sovereignty directly into enterprise architecture, procurement, intellectual property and competitive strategy.
During the cloud era, enterprises accumulated data. During the AI era, they will increasingly accumulate learning.
Data sovereignty protects what an organisation knows. Learning sovereignty protects its ability to become better at knowing. That may prove to be the more important competitive moat.
* The article references Satya Nadella’s 2026 “Reverse Information Paradox” essay, POPIA cross-border transfer requirements, South African cloud/offshoring regulation, and Microsoft 365 Copilot privacy documentation.
Share
