As businesses adopt AI, the conversation about “security” often mixes together several different issues. Is the data encrypted? Where is it processed? Who controls the infrastructure? Can another organisation access it? These questions are related, but they are not the same.
For AI projects, it is useful to separate security from sovereignty. Security is about protecting information from unauthorised access. Sovereignty is about who controls the environment processing that information, where it operates and how much control your business has over where the data travels.
A system can be very secure without being fully sovereign. A major cloud provider may have excellent security controls while still operating infrastructure owned by another organisation. Likewise, keeping everything on a server inside your office does not automatically make it secure.
The best way to understand the difference is to follow the data.
Follow the data
Imagine an employee asks an AI agent which overdue customers should be contacted first. The agent may search the ERP, retrieve a small number of relevant records and send those records to a language model for analysis.
The whole ERP does not necessarily leave the business. Only the information required for that request might be sent.
This is also broadly how Retrieval Augmented Generation, or RAG, works. The system searches a larger knowledge source and gives the model only the relevant information it needs.
The important question is therefore: where does that model run?
If you call a frontier model directly, selected business information needs to reach that provider’s environment. Enterprise agreements may provide strong security and privacy protections, but the inference still takes place on infrastructure controlled by someone else.
A cloud platform can provide a middle ground. Services such as Microsoft Foundry can host closed frontier models within the cloud environment already being governed by the organisation. This can provide stronger control over networking, identity, geography and logging without requiring the business to operate AI infrastructure itself.
Where open-weight models change the picture
Open-weight models give businesses another option entirely. Instead of sending every request to the company that created the model, the model itself can be deployed into an environment you choose.
That might be Azure, AWS, Google Cloud or an on-premise GPU server. In each case, the organisation has much greater control over where inference happens and where business information travels.
This is one reason we think open-weight models should increasingly be considered as the default starting point for business workloads, rather than simply the privacy-friendly alternative to frontier AI.
There is also a commercial reason.
A hosted frontier service generally charges based on usage. Every prompt, document and automated process consumes tokens and creates a variable cost. With an open-weight deployment, the organisation can instead provision infrastructure around its expected workload. That might mean a dedicated cloud GPU that is scaled when required, or a workstation purchased once and operated internally.
Neither model is automatically cheaper in every situation, but open weights provide far more choice over the cost structure. For regular, high-volume workloads, that control can become very valuable.
You often don’t need the smartest model available
Many business AI tasks are narrower than people expect. Classifying an invoice, extracting fields from an order, matching products, routing support requests or summarising standard documents may not need a frontier model at all.
A smaller open-weight model can often perform those tasks well while being cheaper, faster and easier to keep within a controlled environment.
Frontier models still have an important role. They are particularly valuable when the task involves complex reasoning, ambiguous instructions, unfamiliar situations or large amounts of varied information.
The mistake is assuming that frontier capability should be used for every step.
A stronger architecture might use an open-weight model for regular processing and escalate only genuinely difficult tasks to a frontier model. This provides access to advanced reasoning when it creates real value, without sending every piece of business information through the most expensive and least controlled path.
Sovereignty applies to the whole system
Hosting your own model does not automatically make an AI agent sovereign. The agent might still call external OCR services, web search APIs, SaaS applications, databases or monitoring tools.
That means businesses need to review the whole data flow, not just the LLM.
The useful questions are fairly simple: what information does this process require, where does it travel, which organisations can access it, and does this particular step actually need an external frontier model?
There is no single architecture that suits every business. However, open-weight models now give organisations a practical way to keep more processing within environments they control while also gaining greater control over ongoing AI costs.
That makes them increasingly important not only for sovereignty, but as part of sensible AI architecture.
The other half of this discussion is what happens once information reaches an AI service: whether it is retained, how it can be used and whether business knowledge can contribute to future model training. We cover that separately in AI Data Privacy: Your Prompts and Business Knowledge Are Data Too.
Follow your data before it leaves
Every AI process sends information somewhere: to a model, an OCR service, a search API. Most businesses have never mapped where theirs goes. We’ll trace it with you, flag the steps that don’t need an external frontier model, and show you where open-weight models could keep more in-house and cut ongoing costs.
Book a 30-minute discovery call. No sales pitch, just a clear map of where your data travels.