The sovereign enterprise: why rented AI black boxes are becoming a CISO problem

Topic

Sovereign AI

Kind

Perspective

Author

Mario Grunitz

Read time

5 min read

/

/

The sovereign enterprise: why rented AI black boxes are becoming a CISO problem

/

/

The sovereign enterprise: why rented AI black boxes are becoming a CISO problem

Enterprise AI started with a browser tab.

Employees opened ChatGPT, Claude or another AI service, pasted in some text and got useful work back. Companies then moved quickly towards enterprise licences, API integrations and copilots.

That was a sensible first step.

It is becoming a poor architecture for everything that comes next.

As AI moves from answering questions to accessing company systems, retrieving internal knowledge, executing workflows and making decisions, enterprises need to ask a different question:

Who actually controls the intelligence layer of the company?

For European organisations, that question increasingly comes down to sovereignty.

Sovereign AI is an architecture decision

Sovereign AI does not mean building your own foundation model.

It means designing your AI environment so that the organisation retains meaningful control over:

  • data

  • infrastructure

  • model selection

  • access and identity

  • business logic

  • logs and audit trails

  • intellectual property

  • deployment

  • switching costs

A sovereign architecture can still use OpenAI, Anthropic, Mistral, Microsoft, AWS or Google.

The difference is that these services become replaceable components of the architecture rather than the architecture itself.

That distinction matters.

If the reasoning, memory, permissions, workflows and business logic of an AI system are tightly coupled to one provider, changing providers becomes expensive.

If those layers remain under your control, models become infrastructure choices.

The enterprise AI stack is changing

We see the emerging enterprise AI architecture as several separate layers.

1. Enterprise data

ERP, CRM, document stores, databases, APIs, product systems and internal knowledge.

This data should remain governed by the organisation's existing access and security policies.

2. Intelligence

The models used for reasoning, classification, extraction, generation and other AI tasks.

Different tasks may use different models.

A sophisticated reasoning task might use a frontier model. A repetitive classification task may run more efficiently on a smaller local model.

3. Orchestration

This layer determines which model, tool, agent or workflow handles a request.

It also manages memory, context, permissions, retries and human approval.

4. Execution

AI becomes materially more useful when it can act.

That could mean querying a database, generating a report, updating a CRM record, producing code, calling an API or starting a business workflow.

Execution needs explicit boundaries.

5. Interface

The interface could be a conversational application, command bar, existing enterprise application or dynamically generated UI.

Separating these layers creates optionality.

Why model independence matters

The best model today may not be the best model six months from now.

Prices change.

Latency changes.

New European models appear.

Open-source models improve.

Regulation develops.

Individual models become better at particular tasks.

Enterprise architecture should expect this.

A procurement decision made in 2026 should not determine an organisation's AI architecture in 2031.

This is why we favour model routing over model dependency.

One workflow might use a large commercial model for complex reasoning, a locally hosted Small Language Model for document classification and a deterministic service for calculations.

The user does not need to know which model completed each task.

The architecture does.

Small models change the economics

A surprising amount of enterprise AI does not require the world's largest models.

Consider tasks such as:

  • intent classification

  • document categorisation

  • structured extraction

  • entity recognition

  • routing

  • summarisation within narrow domains

  • repetitive internal queries

For these workloads, smaller specialised models can sometimes provide sufficient performance at lower inference cost and with greater deployment control.

That creates an architecture in which sensitive or predictable workloads can run locally or inside controlled European infrastructure, while frontier models remain available for tasks that genuinely benefit from them.

The decision becomes workload-specific.

That is much more useful than an organisation-wide argument about whether one model provider is better than another.

Data sovereignty goes beyond where the server sits

Hosting in Europe is useful.

It is not the entire answer.

A serious sovereignty assessment should ask:

Where is prompt data processed?

What information is retained?

Can customer data be used for model improvement?

Where are embeddings stored?

Who controls encryption keys?

Where do logs live?

Which subprocessors can access the system?

Can the organisation run the workload somewhere else?

Can models be replaced without rebuilding the application?

Who owns prompts, workflows, agents and generated business logic?

What happens to company data when the contract ends?

These are architecture questions as much as legal and procurement questions.

The EU AI Act makes architecture more important

The EU AI Act uses a risk-based framework.

By September 2026, several parts of that framework are already applicable, including rules concerning prohibited practices, general-purpose AI and certain transparency requirements.

Requirements for high-risk systems are being introduced on a later timetable.

For engineering teams, the useful lesson is broader than any single compliance deadline.

Build systems that can be inspected.

That means knowing:

  • which models are being used

  • what data they receive

  • what tools they can access

  • what actions they can perform

  • when humans intervene

  • what happened during execution

  • how outputs can be traced

An AI architecture designed around observability and control is considerably easier to govern than a collection of disconnected AI tools.

Human control belongs in the architecture

"Human in the loop" is often added to AI proposals as a compliance phrase.

It should be much more concrete.

Different actions require different levels of autonomy.

An AI system summarising internal documents may need very little intervention.

An agent preparing a customer email might require approval before sending.

An agent modifying financial data should operate under much stricter permissions.

A useful architecture therefore defines autonomy at the action level.

The question is not:

Can the AI act autonomously?

It is:

Under which conditions is this particular action allowed to happen autonomously?

That distinction makes agentic AI much easier to deploy responsibly.

This is why we built SINAS

At WeAreBrain, these requirements led us to develop SINAS as an infrastructure layer for agentic systems.

The objective is not to create another proprietary AI black box.

It is to give engineering teams control over how models, agents, tools, data and workflows interact.

The architecture is designed around a simple principle:

Own the system. Choose the intelligence.

Models can change.

Infrastructure can change.

Interfaces can change.

The company's knowledge, business logic and operating model should remain its own.

The next enterprise architecture battle

The first phase of enterprise AI was about access.

Who could give employees AI fastest?

The next phase is about control.

Enterprises will increasingly need to decide which parts of their AI stack they want to rent and which parts they consider strategic infrastructure.

There is no reason to reject commercial AI providers.

They provide extraordinary capabilities.

But there is a strong reason to avoid making any single provider the permanent operating layer between your employees, your data and your business.

For European enterprises, sovereign AI is becoming less about isolation and more about optionality.

Use the best technology available.

Keep the ability to change it.

Avatar-1
Avatar-2
Avatar-4
Avatar-4
Avatar-3

Fresh insights, straight to your inbox.

Engineering breakdowns, sovereign-AI perspectives and practical guides from the people building enterprise AI. One email when there is something worth your time.