# What it means to build a European AI agent

A European company wants to give an AI agent access to contracts, dashboards, and internal research. To assess sovereignty, its IT lead connects the publisher's origin, processing locations, contracts, data formats, and replacement paths.

**Sovereignty is the ability to choose, understand, operate, replace, and leave the different layers of a system.**

## Why this question matters particularly in Europe

European organisations use a market concentrated around large American and Asian companies. On August 13, 2026, two closed work-agent services selected for comparison, Claude Cowork and ChatGPT Work, are published by Anthropic and OpenAI, both American companies.

The model landscape is more diverse. Anthropic, OpenAI, Google, Meta, and SpaceXAI are American. Qwen is developed in Alibaba's Chinese ecosystem, and DeepSeek operates from Hangzhou. Mistral AI is French. Umans, which provides access to several Asian model families through a French offering, is registered in Montrouge. Open projects and European providers add further options.

A company's origin becomes useful when connected to processing location, governing law, log access, model openness, and the exit path.

For a European organization, this is a strategic question. Choosing an agent affects work documents, connectors, internal skills, decisions, and sometimes personal data. A dependency that began as text generation can become the interface through which a team organizes its activity.

## Six layers, six control questions

An agent's sovereignty can be examined through six mutually reinforcing layers.

1. **Data.** Where is the authoritative copy? Who can read, export, and delete it?
2. **Software.** Can the behavior be inspected, rebuilt, and maintained without the publisher?
3. **Models.** Is the model imposed, replaceable, local, or selected from an exact provider?
4. **Infrastructure.** Where do the site, agent, tools, and inference run? Which operators have access?
5. **Law.** Which entities carry the responsibilities of controller, processor, provider, or deployer?
6. **Skills.** Can the organization understand, qualify, operate, and replace the system?

Sovereignty grows when each of the six layers pairs a concrete control with evidence: a readable format, protected secret, verifiable build, identified route, and operational skill.

![Six layers connect data, software, models, infrastructure, law, and skills to one concrete control and one verifiable form of evidence](/assets/blog/european-sovereignty-stack.svg "Sovereignty depends on controls distributed throughout the system. Each layer pairs a decision with evidence that makes it verifiable.")

## Local data, chosen processing

loqy begins with local authority. In V1, the agent engine, conversations, Projects, rules, evidence, and authoritative deliverables live on the Mac. The application works without a loqy account. A task that fits a local model can run without sending its context to an inference provider.

This architecture reduces the number of parties that must be trusted. The macOS account, FileVault, backups, updates, and physical device security complete the protection of the machine. Local operation therefore improves control and minimises transfers.

When a user selects a hosted model, context admitted for that request leaves the device. The local engine retains authority over tools, credentials, rules, and the Project. The provider receives what is needed for inference according to the exact route and contract.

This distinction enables a concrete European choice. A user can select a Mistral model from the French provider Mistral AI, use DeepSeek through the French provider Umans when its cost, speed, and data terms fit, call Claude or GPT for a different task, or remain on a local model. The selected route remains fixed during the run.

Control lies in an informed choice attributed to an exact model, provider, and route. [The execution-location map](/blog/where-ai-agent-work-runs/) shows the data and authority boundaries behind that choice.

## What the GDPR and EU AI Act change

The GDPR and the European Union's AI Act provide an operational framework for this architecture.

The GDPR governs matters including roles, purposes, minimisation, individual rights, and transfers of personal data. Its right to data portability covers certain personal data in a structured, commonly used, machine-readable format. Product portability extends this principle to the Project, financial models, evidence, and business rules controlled by the user.

The AI Act sets rules according to roles, uses, and levels of risk. Its scope can cover providers and deployers outside the Union when a system or its output is used in Europe. It contains prohibited practices, obligations for certain high-risk systems, transparency rules, and requirements for general-purpose AI models. Each deployment is assessed according to its use and actors.

As of August 13, 2026, much of the Regulation has applied since August 2, 2026, while some obligations follow a different schedule. Each deployment maintains its legal analysis and compliance documentation.

A verifiable architecture helps in a different way: it makes regulatory questions answerable. The organization can identify the controller of data, exact model, network route, evidence behind a decision, exported items, and processing limits. Compliance still requires the right policies, contracts, and operating practices.

## Govern every dependency through authority and exit

loqy is designed in France for a worldwide audience, with a particular responsibility to European users. Its international stack remains explicit and replaceable.

The first version targets Apple Silicon and macOS. Apple is an American company, and its operating system remains a major dependency. The inference engine uses open projects developed by an international community. The local catalog contains Qwen, Ornith, and Gemma models from organizations or communities in several regions. Hosted providers likewise span Europe, the United States, and Asia.

The public website is hosted in France by Scaleway. That choice brings the Web infrastructure, provider jurisdiction, and project operation closer together. Inference deliberately called through Anthropic, OpenAI, or another provider follows its own identified route.

Evaluate each dependency through four concrete questions:

- what authority it receives;
- what data it sees;
- how its version and provenance are verified;
- what happens if it changes, becomes unavailable, or no longer fits.

![A map connects loqy's principal dependencies to retained authority and realistic replacement paths](/assets/blog/dependency-exit-map.svg "Apple, models, inference providers, and hosting each have a distinct exit path: available now, subject to qualification, or enabled by the architecture.")

## A credible, graduated exit

The Apple dependency shows the difference between architectural portability and immediate availability. loqy's business contracts remain portable, while V1 is a macOS application for Apple Silicon. Windows and Linux are targets enabled by separating the business core from platform adapters.

Replacing a model is more direct. The local engine owns a catalog of exact artifacts. A Run selects one model and retains its attribution. Another local or hosted model can handle the next request without taking possession of the history.

Replacing a hosted provider follows supported routes. Credentials stay in macOS Keychain, and changing provider requires a new explicit configuration.

The static website offers a simpler case. Its generated output and manifest can be deployed to another compatible Web server. The current Hermes configuration at Scaleway is the chosen operation for a portable site format.

Each dependency has a different exit state: available now, possible after qualification, or enabled by architecture. [The service-disappearance test](/blog/own-your-work/) describes what a practical exit preserves.

## Skills are part of sovereignty

Sovereignty includes an organisation's ability to operate its code, models, and data beyond a small number of specialists.

For loqy, decision records, architecture diagrams, manifests, build procedures, software bills of materials, and qualification tests serve an operational purpose. They explain why a boundary exists and how to check that a new version preserves it.

That knowledge must be transferable. Reproducing a build, reviewing an update, changing a model, restoring an export, or qualifying a provider should not depend on private memory. Open source becomes useful when a decision, source, version, proof, and procedure can be connected. [The article on evidence](/blog/why-ai-agents-need-evidence/) follows that connection through one concrete piece of work.

## A decision grid for European organizations

Before adopting a work agent, a team can ask:

1. Which operations remain possible without an account or remote service?
2. Who holds the authoritative copy of conversations, files, and evidence?
3. Are the model and its provider two distinct choices?
4. What data leaves the device for each feature?
5. Which jurisdictions cover the publisher, host, and effective processor?
6. Does the distributed code have verifiable provenance, signatures, and a software bill of materials?
7. Which formats and procedures make exit possible?
8. Which dependency still lacks a qualified replacement?

A serious European agent shows its boundaries, acknowledges its dependencies, and gives users choices with real consequences.

loqy lets users choose the option that best fits the need while keeping each dependency visible and replaceable. For a global product designed in Europe, sovereignty is measured by that ability: understand who is involved, choose freely, and leave without losing the work.

## Sources

- [General Data Protection Regulation](https://eur-lex.europa.eu/eli/reg/2016/679/oj)
- [European Union Artificial Intelligence Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)
- [Claude Cowork](https://claude.com/product/cowork)
- [ChatGPT Work](https://openai.com/index/chatgpt-for-your-most-ambitious-work/)
- [OpenAI terms](https://openai.com/policies/terms-of-use/)
- [Mistral AI legal notice](https://legal.mistral.ai/legal-notice)
- [Umans legal notice](https://app.umans.ai/offers/code/legal/legal-notice)
- [Scaleway legal notice](https://www.scaleway.com/en/legal-notice/)
- loqy V1 requirements and boundaries
- Reproducible build architecture