AI Glossary Safety, privacy and governance

Sovereign AI

  • Tidio Agents team
  • Updated November 19, 2026
  • 6 min read
  • 8 sources
Definition

Sovereign AI is the ability of a country or an organization to develop, run, and control artificial intelligence on its own terms: where its data is stored and processed, which infrastructure the systems run on, and who controls the models. For governments it is a question of strategic autonomy. For companies it usually means operational control.1

The term has no single legal definition, and it is used at two levels. Governments use it for national capacity: compute, data, models, and skills inside their borders or their trading bloc. Companies use it for something narrower and more practical: can they say where customer data goes, who can access it, and whether they could switch provider or keep running if a vendor changed its terms.1

Stanford HAI researchers describe a range from "harder" notions, which aim at self-sufficiency, to "softer" ones framed as strategic autonomy, and note that full self-sufficiency across the whole AI stack is prohibitively costly and unnecessary for most countries.1 The same logic applies to a company: the useful question is which layers it must control, not whether it controls all of them.

How it works

Sovereignty is decided layer by layer

Researchers map AI sovereignty onto the layers of the technology stack: infrastructure, data, models, applications, and talent. Control at one layer does not guarantee control at another.1 A company can keep its data in Frankfurt and still depend on a model it cannot inspect or move.

Data

Where personal and business data is stored, processed, and logged, and under which laws. For EU companies this is governed by GDPR, including its rules on transfers outside the EU.5

Infrastructure

Whose data centers and accelerators run the workload, and which jurisdiction the operator answers to. A provider's home country can matter as much as the server's location.

Models

Whether the organization can choose, host, inspect, and replace the model. Open-weight models can run on infrastructure you pick; closed models run where their provider offers them.

Operations and skills

Who holds the code, prompts, evaluations, and know-how to run the system. Without them, control on paper does not survive a vendor change.

Levels

How much control a company has over an AI agent

For a business deploying a customer service or back-office agent, sovereignty is a set of hosting and contract choices. The scale below is a working model, not a standard.

0

Public API, default region

Controls the stack: the model provider

Fast to start. Data may be processed outside the EU, and the provider sets the model, the version, and the terms.

1

Public API, EU region

Controls the stack: the provider, with EU processing

Data is processed in the EU under the provider's terms. Model choice and updates still sit with the provider.

2

EU-operated cloud

Controls the stack: an EU operator

The model runs with a European cloud operator, which reduces exposure to non-EU access laws.

3

Own cloud tenancyCommon for regulated firms

Controls the stack: the company

Models, logs, and indexes run in the company's own cloud account, in a region it chooses, with code it owns.

4

On-premises

Controls the stack: the company, end to end

Open-weight models on the company's own hardware. Maximum control, with the highest operating cost and skills burden.

EU context

Why the question is sharper in Europe

GDPR allows personal data to leave the EU only under the conditions of its Chapter V.5 In the Schrems II judgment of July 2020, the EU Court of Justice invalidated the Privacy Shield, the framework many companies had used to send data to the US, while keeping standard contractual clauses valid.6 The European Commission adopted a new adequacy decision, the EU-US Data Privacy Framework, on 10 July 2023.7 For many legal teams the lesson was that a transfer basis can disappear, so they prefer architectures that do not depend on one.

Public policy has moved the same way. The Commission's AI Continent Action Plan, published on 9 April 2025, set out the aim of making the EU a leading AI continent.2 Through the EuroHPC Joint Undertaking, the EU has established 19 AI Factories around AI-optimized supercomputers, with sites in countries including Poland, Germany, and Finland.3 In June 2026 the Commission proposed a Cloud and AI Development Act that aims to strengthen strategic autonomy and to triple EU data center capacity within five to seven years.4

The Cloud and AI Development Act was a proposal at the time of writing (procedure 2026/0138(COD)). Its final text may differ.

Compared

Sovereign AI and neighboring terms

Vendors use these terms loosely. Asking which layer each one controls makes contracts easier to compare.

TermWhat it controlsWhat it leaves openTypical question from a DPO
Data residencyWhere data is stored, and sometimes where it is processedWho operates the service, which model runs, who can access itDo transcripts and logs stay in the EU?
Sovereign cloudThe operator and jurisdiction of the infrastructureModel choice and the application layerWhich laws can compel the operator to hand over data?
Private LLMA model deployed for one organization's exclusive useWhere that deployment runs and who maintains itIs our data used to train anyone else's model?
Sovereign AIData, infrastructure, models, and operations, to a chosen degreeNothing by definition; the scope is a decisionCould we keep running if the vendor changed terms?
In practice

Where companies ask for sovereign AI

The demand comes from compliance and security teams, and it is strongest where agents read regulated data.

Fintech and banking

Payment and instalment support

An agent reads repayment schedules and account status. Supervisors expect outsourcing controls, so the bank keeps the agent in its own cloud tenancy.

own tenancy → EU region → audit log
Insurance

Claims intake

Claims often contain health data. The insurer runs extraction and triage models where it already processes special category data.

classify → restricted environment → handler
Telecoms

Voice and billing agents

Call audio, transcripts, and billing records stay with the operator, which already runs its own data centers and security teams.

transcribe on site → resolve → store locally
Retail

Shopping and support agents

A retailer keeps the catalog index, chat logs, and model prompts in its own account so it can change model provider without rebuilding the agent.

own index → swappable model → own logs
Back office

Finance and procurement

Invoices, contracts, and supplier data feed internal agents. Keeping them in-house limits who can see commercial terms.

extract → match → post to ERP
Internal knowledge

Company assistants

An assistant over HR policies and contracts runs inside the company network and respects existing access rights.

search → filter by access → cite
Trade-offs

What sovereignty costs, and how teams manage it

More control is not free. Each layer a company takes over adds work that a provider used to do.

Weaker models

ControlThe strongest models may not be available in every region or for self-hosting. Measure quality on your own tasks before deciding that a smaller hosted model is good enough.

Operating burden

ControlSelf-hosting means patching, scaling, monitoring, and on-call. Budget for it, or use an operator that runs the stack in your account.

Sovereignty washing

Control"Sovereign" on a datasheet can mean only data residency. Ask for the operator's jurisdiction, the subprocessor list, and where logs and backups live.

Lock-in moves, it does not vanish

ControlOwning the infrastructure but not the code still ties you to a vendor. Contract for code handover and portable prompts and evaluations to limit vendor lock-in.

FAQ

Common questions

What does sovereign AI mean for a company?

It means the company controls the parts of its AI systems that matter to its risk: where data is processed, who operates the infrastructure, which model runs, and who owns the code. It rarely means building everything in-house.

Is sovereign AI the same as data residency?

No. Data residency is one part of it: keeping data in a given location. Sovereignty also asks which laws apply to the operator, whether you can replace the model, and whether you could keep running if a vendor changed its terms.

Do you need your own model to have sovereign AI?

Not necessarily. Many companies run an existing open-weight model, or a commercial model deployed in their own cloud tenancy. Training a model from scratch is a national-level project, not a requirement for a business.

Sources

References

  1. Pava, J. N., Meinhardt, C., Cryst, E., and Landay, J. A. (Stanford HAI). AI Sovereignty's Definitional Dilemma. February 2026.
  2. European Commission. The AI Continent Action Plan. 9 April 2025.
  3. EuroHPC Joint Undertaking. AI Factories.
  4. European Parliament Legislative Observatory. Cloud and AI Development Act, procedure 2026/0138(COD): summary of the Commission proposal. 3 June 2026.
  5. European Union. Regulation (EU) 2016/679 (General Data Protection Regulation), Chapter V, Article 44.
  6. Court of Justice of the European Union. Press release 91/20: Judgment in Case C-311/18, Data Protection Commissioner v Facebook Ireland and Maximillian Schrems. 16 July 2020.
  7. European Commission. EU-US data transfers.
  8. US Department of Justice, Criminal Division. CLOUD Act Resources.

Tidio Agents

Build an agent that resolves, not only answers

We design and build custom AI agents for customer service, shopping and back-office work, on your infrastructure, with the code handed over to you. Projects start with a 21-day private AI audit.

Request a call