There is an ongoing battle to capture a particular type of data: that representing the context layer that is unique to each organization and recorded in processes, workflows, relationships, practices, and routines. The governance regime of this information will have huge impacts on the shape of the AI market, as well as on governments’ ability to control how they deploy and use AI. To steer this development towards public value, governments have a powerful tool. One that has worked elsewhere: portability.
The current rush seeks to capture the information that encodes how companies actually work. What has been labeled “context”. We can think of context as an organization’s “institutional memory”: the workflows, strategy and roadmap discussions, meeting cadence, pricing approaches, customer information, decision rules, permissions… AI companies, including labs developing frontier models, know that this data is both a competitive advantage and the layer where potential ROI begins to justify investments in model training. That’s why frontier labs moved beyond API provision to focus on the application and agent-orchestration layers that are embedded in company operations.
Whoever becomes the system of orchestration and captures an organization’s institutional memory will build a durable moat. Changing an agent-orchestration provider will be much harder than switching a model through API gateways and model routers.
Of course, one’s moat is another’s vulnerability. And that is why the sovereignty discussions that focus only on model routing or open weights are missing a big part of the story. For governments and large enterprises alike, the risks of lock-in are not disappearing, just moving up a layer in the stack.
That’s the sub-text (or not so “sub”) of the recent declarations by the CEOs of Microsoft and Palantir, as well as one of the lenses through which one can read the letter on open weights promoted by NVIDIA. These companies do not want to become dependent on frontier labs, but they are also eager to capture the value if they end up controlling the layer that is harder to switch away from.
As purchasers and regulators of AI capabilities, governments ought to pay close attention to these developments and start sharpening the tools at their disposal to increase their optionality and negotiation power. A critical one is portability. They need to make sure that whoever ends up gathering data about their context is ready to make it portable whenever they decide - if they decide - to choose a different provider.
The EU has already been pushing for portability in different ways. The Data Act mandates portability of certain data, and the open banking provisions of the Revised Payment Services Directive (PSD2) have arguably been key for the emergence of new financial startups. The Interoperable Europe Act and the Digital Markets Act (DMA) both contain provisions to reduce switching costs and enhance competition.
The Commission has also shown its appetite to make them enforceable. Last July it reached a final decision specifying “measures that Google must implement to ensure effective interoperability under the DMA with 11 Google Android features relevant for AI services”. One of these features is “context”, defined by the Commission as “the ability for an AI service to gather data (including from apps, sensors or the screen) to then understand and anticipate the user’s needs”.
Governments should also explore context data portability clauses in their procurement contracts. The EU has published model clauses for cloud computing contracts and AI enabled solutions, but it is unclear how these will cover data on the context layer.
There are important caveats too.
Governments can write excellent clauses, but these contracts - and the laws and regulations that they’ll invoke - will be meaningless if governments are not able to document their context in portable languages, ontologies and tools. When the time to switch a vendor comes, governments won’t be able to leave if nobody can read the export.
Context needs to be represented in a portable artifact. This requires a government that is able to describe itself, its processes, its decision rules and exceptions in languages and ontologies that survive a move. Few organizations, let alone governments, can do this. This representation is also frequently co-generated with the provider of the AI system that designs and runs the agent workflows, tool calls, permissions and evaluation sets. Allocating entitlements and rights over the portable artifact is not trivial.
Even if portable artifacts can be created and transferred in technically and legally agreeable terms, some might argue that this is not enough to gain sovereignty if everyone builds on top of foreign models. While there is some truth to this view, the rapid progress of open weight models and advances in token routing - with Stripe’s acquisition of OpenRouter as the latest illustration of the current race to build or control the AI stack, in finance and elsewhere - signals that the real leverage is moving up the stack.
Achieving full sovereignty is chimeric anyway. So better to focus on keeping options as varied and open as possible. That is where governments should work hard on, and that is where context data portability becomes crucial.