Open Source Models Are the Pathway to Sovereign AI
Over the past few months we have been playing around with open source AI models in earnest, assessing their performance against many of the standard enterprise AI use cases we see every week. Document question answering, summarisation, drafting, structured analysis over a business's own files, and retrieval with proper citations. We have also been thinking hard about how these capabilities could be folded more permanently into agents over time, so that open models do not just answer questions but carry out meaningful work. What began as an experiment has become something closer to a conviction, because it has demonstrated to us that sovereign AI is no longer an aspiration for the future. It is available today, and open source models are the reason why.
What we mean by sovereign AI
Sovereign AI is often discussed at the level of nations and data centres, but the principle applies just as forcefully at the level of a single organisation, a single team, or a single laptop. It means that your intelligence layer runs on infrastructure you control, that your data is processed where you decide it should be processed, and that no third party sits between your people and their tools. When the model runs on the device in front of you, the question of where your data goes has a very simple answer. It should go nowhere.
That matters more than it might first appear. When colleagues use cloud-based AI assistants, the data they work with is being processed outside the organisation's own environment. For most tasks that is entirely manageable, and well-governed cloud based, frontier certainly AI has an important role to play. But it is precisely there that we hear consistent tension from customers. Many of the organisations we work with want frontier intelligence for their people, and they can see the productivity it unlocks, but they are still working through the sensitivities of their data and how to maintain proper governance controls around it. The appetite is there. The approval pathway often takes longer to establish.
Open source models have crossed the threshold
The little known AI revolution of the last two years is that open source models became genuinely good. In our testing, modern open model families running locally on Apple Silicon (we have a lot of Mac Mini’s) will summarise documents, reason through multi-step problems, ground their answers in a user's own files with citations, and stream responses at a pace that feels immediate. Two years ago that combination required a frontier API. Today it fits comfortably on a well-specified laptop.
This is not an argument that open models have matched the frontier across the board, because they have not. It is an argument that the capability floor has risen so far that a very large share of everyday knowledge work now sits comfortably within reach of a model you can download, inspect, and run yourself. For the standard enterprise use cases we assessed, the on-device experience is no longer a compromise.
Building skills once and running them anywhere
One of the most consequential but underreported developments in the AI space over the past year has been Anthropic's decision to open source its Agent Skills framework and donate the specification to the Agentic AI Foundation. What began as a feature inside Claude Code in October 2025 became, in December of that year, a vendor-neutral open standard that any agent runtime can adopt. The format is deliberately simple: a SKILL.md file combining YAML frontmatter for machine-readable metadata with plain Markdown instructions, packaged inside a folder alongside any scripts, templates, or reference assets the agent needs. The specification tells the agent what the skill is, when to invoke it, and how to execute it step by step. Within 48 hours of the open standard being published, Microsoft and OpenAI had both adopted it. By early 2026 more than 30 agent platforms were reading identical SKILL.md files.
The reason for that rapid adoption is not hard to understand. Intelligence alone is not enough; for AI to be truly useful in a professional context, it must be able to follow complex, standardised procedures across a variety of tools and platforms. Agent Skills solves that problem at the level of the specification rather than the vendor. The skill defines the procedure. The model executes it. Those are now separable concerns, and that separation matters enormously.
The implications for enterprises are profound, and we think they will only become clearer as agentic AI moves from experimentation into production. Organisations investing seriously in AI-enabled workflows and critical processes need to know that those processes will not break because a provider changes its pricing, deprecates an API version, or suffers an outage. The operational resilience requirements that govern their software infrastructure must apply to their AI infrastructure as well. A workflow encoded in a proprietary format, understood only by one provider's runtime, is a single point of failure. A workflow expressed in an open SKILL.md is not.
The pattern that should become the norm, particularly for enterprises running critical processes on AI, is this. Build the skill once, against the open standard. Run it anywhere the task and the data require. When the work is routine and connectivity is available, run it against a frontier model. When the data is sensitive, when the network is unavailable, when the information cannot leave the building under any approval regime, route it to an open source model running on-device. Whether thats a server or even a well spec’d laptop.
Indeed, the skill travels unchanged. The agent logic travels unchanged. The accumulated institutional knowledge encoded in your procedures travels unchanged. What changes is only which intelligence is doing the reasoning, and you get to make that decision based on what is appropriate, not based on what a proprietary architecture will permit.
For the customers we work with who are grappling with data governance and model approval processes, this is often the unlock. It separates two questions that are routinely but unhelpfully conflated. The first is how we build our AI workflows, and the answer should always be open standards and portability. The second is which model we trust with which data, and the answer should be determined by proper controls, routing logic, and the sensitivity of the task at hand. Agent Skills, and the Agentic AI Foundation's stewardship of the standard, make that separation not just possible but practical.
When the cloud is simply not an option
There is a tendency to treat offline AI as a niche concern, and our experience suggests the opposite. The situations where cloud intelligence is unavailable or unapprovable are common, recurring, and often precisely the moments when intelligence matters most.
Consider the flight at 10,000 feet, where the two hours you hoped to spend preparing a board paper are lost to a captive portal. Consider the site visit with unreliable connectivity, the substation, the rural depot, the hospital basement. Consider the categories of data where cloud processing requires a longer approval process, or where the organisation's risk appetite means keeping things closer to home. Operators in sectors such as critical national infrastructure, regulated industries, and professional services handling sensitive client material often find that the governance journey for cloud AI is a lengthy one. On-device inference can be a practical way to make progress while that journey continues, extending genuine AI capability to people who would otherwise be waiting.
The workforce nobody licensed
There is also an equity dimension that deserves more attention than it receives. The current generation of AI assistants has largely been rolled out to the people who already had the best tooling. Desk-based knowledge workers received Copilot. ChatGPT and Claude licences, while field operations teams, factory floor staff, maintenance engineers, and frontline colleagues were deemed not plugged in and often left out of the deployment plan.
These are often the people with the greatest need for on-the-spot intelligence. The engineer standing in front of a fault with a 400-page manual, the operator reconciling a shift log, the field technician who needs an answer from a technical drawing rather than a search box. A capable open source model running locally, with the organisation's own documents ingested and retrievable, puts genuine AI assistance into those hands without a per-seat cloud licence, without connectivity requirements, and without a data governance review for every query.
Private by design, and always available
If there is a single idea we would ask readers to take away, it is this. Data privacy in AI does not have to rest entirely on policy controls layered on top of a cloud service. It can also be a property of the architecture itself.
When inference happens on the device, the question of what leaves your environment has a straightforward answer. Nothing does. There is no data residency concern to manage, no retention policy to negotiate with a provider, and no additional review process each time a new use case emerges. The governance burden is simply lighter, because the architecture makes many of the harder questions easier to answer.
An on-device assistant is also available in situations where cloud tooling is not. It works on a flight, on a remote site, in a secure environment, and through a connectivity outage. For organisations where reliability and continuity matter, that consistency has real operational value, independent of any data sensitivity considerations.
It is exactly this combination of customer demand and technical readiness that led us to set about prototyping an initial MVP, a fully on-device AI application built on open models and open frameworks, designed for the organisations and the people the cloud cannot always reach. We will be sharing more on it in the next few weeks.
Open source models gave us the capability, open frameworks gave us the interoperability with the frontier, and modern laptop silicon gave us the hardware to run it all. What remains is for organisations to recognise that sovereign AI is no longer a research programme or a procurement fantasy, and that it is increasingly within reach for teams who are ready to move.
