The vocabulary of sovereign AI, without the jargon.

The ability of an organization or state to keep control of its data, processing and infrastructure, without depending on a foreign actor that could restrict access or compel disclosure.

A 2018 US law allowing US authorities to compel a provider subject to US jurisdiction to hand over data it holds, wherever in the world it is hosted. The exposure reaches your data even when stored in Europe.

Also called Clarifying Lawful Overseas Use of Data Act

A property of an architecture where no entity subject to US law sits in the data path, placing it beyond the reach of a CLOUD Act request. It is achieved by design, not by a contractual clause.

The General Data Protection Regulation, the EU framework in force since 2018 governing the collection and processing of personal data: lawful basis, minimisation, data-subject rights and security obligations, supervised by the CNIL in France.

Also called GDPR

The EU Artificial Intelligence Act (Regulation 2024/1689), in force since 1 August 2024 and applied in stages: prohibited practices and AI literacy since February 2025, general-purpose models since August 2025, transparency and the penalty regime on 2 August 2026. The omnibus adopted in June 2026 postponed high-risk systems to 2 December 2027 (Annex III) and 2 August 2028 (regulated products). In practice: knowing which model processes what, where, on which data, and being able to prove it.

Also called AI Act, Règlement européen sur l'IA, RIA

The Digital Operational Resilience Act (Regulation EU 2022/2554), applicable since 17 January 2025: it requires the financial sector - banks, insurers, asset managers, payment providers - to control its digital risks, including those of its providers. An AI vendor is an ICT third-party provider like any other: a framed contract (Articles 28 to 30), entry in the information register, reversibility and auditability on demand.

Also called Digital Operational Resilience Act, Résilience opérationnelle numérique, Règlement 2022/2554

Health data hosting, a mandatory French certification for any provider hosting health data on behalf of third parties (art. L.1111-8 of the French Public Health Code); the v2 framework requires physical hosting within the EEA. It certifies the hosting, not sovereignty: US hyperscalers hold it.

Also called Hébergement de données de santé, Hébergeur de données de santé

A qualification framework from ANSSI for trusted cloud services. It imposes a high security level and immunity from non-European laws, making it a sovereignty benchmark for public bodies and sensitive operators.

A European cybersecurity directive that extends risk-management and incident-reporting obligations to a wide range of essential and important sectors, with direct accountability for senior management.

The physical and legal location where your data is stored and processed. Keeping it in France or Europe is a necessary condition for sovereignty, but not sufficient: the jurisdiction that applies to the provider matters just as much.

Also called Localisation des données, Data residency

The set of laws a provider is subject to by reason of its nationality, shareholders or establishments. A service hosted in France but controlled by a US company remains exposed to US law.

An AI model whose weights are released and can be downloaded, run and adapted on your own infrastructure. It enables self-hosting and execution sovereignty, unlike a model accessible only through a remote API.

Also called Modèle à poids ouverts, Open-weight model

One of the most capable AI models available at a given time, usually proprietary and accessed through an API. It offers the best raw quality but is used with no control over the infrastructure or the jurisdiction that applies to it.

Also called Frontier model

A European language-model maker based in France. It is the serious European engine of the ecosystem: a fully European model, useful when sovereignty over the model origin matters as much as over its execution.

Bring your own keys: you supply your own API keys to a model provider. This lets a non-sensitive agent point at a frontier model while knowingly accepting the corresponding legal exposure.

Also called Bring your own keys

Retrieval augmented generation: a technique that first retrieves relevant documents from your stores, then feeds them to the model so it grounds its answer on them. It anchors answers in your data and makes source citation possible.

Also called Retrieval augmented generation, Génération augmentée par récupération

A software system that uses a language model to plan and carry out a multi-step task, calling tools and data sources. At root, sovereignty is decided agent by agent, according to sensitivity.

Also called Agent, AI agent

Linking each generated answer to the exact source that supports it in your documents. This turns a statement into something verifiable and limits liability tied to a possible hallucination.

Also called Ancrage, Grounding

A model answer that sounds plausible but is false or unsupported by the data provided. The risk is reduced by grounding in your sources, systematic citation and human oversight on sensitive decisions.

Light retraining of a pre-trained model on your own data or business domain, to adjust its behaviour and vocabulary. On an open-weight model it runs on your infrastructure, without exposing your data.

Also called Affinage, Réglage fin

Numerical representations of your texts as vectors, where contents close in meaning sit close in space. They make possible the semantic search that powers RAG.

Also called Plongements, Vecteurs

A database specialised in storing and searching embeddings. It quickly finds the passages closest to a question, the central step of a per-customer isolated RAG system.

Also called Vector database, Base de données vectorielle

An orchestration layer that routes each request to the right model based on the task, jurisdiction and sensitivity, behind a single interface. It makes multi-model orchestration aware of the applicable law.

Also called LiteLLM, Model gateway

The principle that each customer has its own instance and stores, with no mixing with another customer data. It contrasts with shared multi-tenancy, where several customers coexist in the same database.

Also called Multi-tenant, Cloisonnement

Deployment of the solution on infrastructure you control, on your premises or in a dedicated self-hosted environment. It is the option chosen for the most sensitive settings, where nothing must leave.

Also called Auto-hébergé, Self-hosted

Transforming data into a format unreadable without the key, at rest and in transit. Combined with key management you control, it ensures only legitimate holders can access the content.

Also called Encryption

The set of rules that decide, for each agent, which model is allowed, where it runs and under which law. It is what makes the Sovereignty Map a living view rather than a frozen document.

Also called Model governance

Teams using AI tools without the organization's approval or framework: personal accounts on consumer chatbots where client data gets pasted, and now agents entrusted with credentials. Invisible in the logs and outside any security policy, it exposes data and compliance alike. The durable answer is not a ban: it is an official, governed tool that is at least as useful.

Also called IA fantôme

Safety mechanisms that frame what an agent can do and answer: input filtering, output validation, business rules and action limits. They turn raw capability into a responsible production system.

Also called Guardrails

A pattern where a human reviews, corrects or approves an agent decisions before they take effect. It is mandatory on sensitive decisions: AI augments the professional, it does not replace them.

Also called Human-in-the-loop, Humain dans la boucle

A timestamped, tamper-evident record of what an agent did: requests, models called, sources cited and decisions. It makes AI use traceable and accountable, a condition for holding responsibility.

Also called Audit log, Traçabilité

A mechanism that starts an agent without human intervention: a recurring schedule or an event received from a third-party system. In a governed harness, the run stays logged and any writes it produces wait for human approval before going out.

Also called Trigger, Agent programmé

A technique where a large model teaches a smaller one, which inherits part of its quality at a far lower cost and footprint. It makes it possible to run a capable model on controlled sovereign infrastructure.

Also called Distillation, Knowledge distillation

The distinction between where a model runs and under which law (execution) and where the model itself comes from (origin). Running a foreign open model in France gives execution and data sovereignty, not origin sovereignty.

Also called Exécution vs origine

Search that finds information by meaning rather than exact words, by comparing embeddings. It lets an agent find the right passage even when the question is phrased differently from the document.

Also called Semantic search

An approach that assigns the right model to each task, rather than relying on a single one, accounting for required quality, cost and jurisdiction. It is the conductor of the Workspace.

Also called Multi-model orchestration

Model Context Protocol: an open standard describing how an AI agent connects to an organization's tools and data sources. One connector per tool, enumerable capabilities, individually loggable calls: integration becomes reversible, since the connector belongs to the standard rather than the platform.

Also called Model Context Protocol

The French national cybersecurity agency. It defines trust frameworks, including SecNumCloud, which serve as sovereignty benchmarks.

The French data-protection authority, responsible for ensuring compliance with the GDPR and protecting personal data. It can audit, sanction and guide practices.

An international standard for information-security management. Certification attests that an organization manages its security risks in a structured, audited way, a common foundation for trusted infrastructures.

Any information relating to an identified or identifiable individual. Processing it through an AI agent triggers GDPR obligations and justifies a sovereign choice of model and hosting.

Also called Personal data

The autonomous behaviour of an AI system that decides, plans multiple steps, executes and adapts with limited human input, as opposed to a purely reactive chatbot. It is what sets the 2025-2026 era apart from generative AI that merely answered.

Also called Agentic, IA agentique

The infrastructure wrapping a model to turn it into a reliable agent: orchestration, memory, tool interfaces, guardrails and runtime. The common formula sums it up: Agent = Model + Harness. Without a harness, a model stays a demo; the root Workspace is a governed, sovereign harness.

Also called Harness, Harnais, Harness engineering

The cycle an agent repeats to reach its goal: it perceives the state, plans, calls a tool, observes the result, then loops again using its memory until the task is done. It is the core of agentic execution.

Also called Agent loop, Boucle observer-planifier-agir

An architecture where several specialised agents collaborate on a task, with a lead agent able to delegate to subagents. It breaks a complex goal into roles, but also multiplies the points to govern and audit.

Also called Multi-agent system, Sous-agents

An agent that keeps memory across sessions, letting it reuse acquired knowledge and skills rather than starting from scratch at each request. Persistence demands stronger traceability and governance over what is remembered.

Also called Persistent agent, Mémoire persistante

An open-source autonomous agent, released in late 2025 as Warelay, that runs real tasks via language models accessible through messaging apps. It went viral with over 100,000 GitHub stars, is stewarded by the OpenClaw Foundation, and has documented security weaknesses.

Also called Warelay

An open-source self-improving agent harness (MIT licence) released by Nous Research around early 2026. It keeps persistent memory across sessions, reuses Skill Documents, manages subagents and runs local or in the cloud.

Also called Hermes Agent

A special category of personal data under Article 9 GDPR, whose processing is prohibited in principle outside strictly framed exceptions. In practice it reaches beyond the medical record: a consultation report, a referral letter, a billing code or an appointment schedule all reveal a person's state of health.

Also called Données de santé à caractère personnel, Health data

The duty covering everything that comes to a physician's knowledge in the course of practice (art. R.4127-4 of the French Public Health Code), enforced by criminal law (art. 226-13 of the Penal Code). Pasting patient data into an AI tool hosted by a third party means entrusting it to that third party: the choice of tool puts the secret at stake.

Also called Secret professionnel médical

Under EU Regulation 2017/745 (MDR, Rule 11), software providing information used for diagnostic or therapeutic decisions is a medical device, requiring CE marking at class IIa or above. Documentation and administrative software is not one; qualifying as a medical device also pulls the system into the AI Act high-risk category.

Also called DM logiciel, Software as a Medical Device, SaMD

The French trust space for secure health messaging, operated by the national e-health agency (ANS). Being MSSanté-compatible is the condition for exchanging health data with the national system, including Mon espace santé; a standalone deployment inside one organisation can run without it.

Also called Messagerie sécurisée de santé

A French state programme funding and standardising digital health, in particular data exchange between sector software and Mon espace santé. Ségur listing (référencement) attests that software meets these interoperability and security requirements; it is the expected gateway for selling into public healthcare.

Also called Ségur numérique, Référencement Ségur

The inventory Article 28(3) of DORA requires from every financial entity: all its contractual arrangements for ICT services, kept up to date in a register the supervisor can demand in full. Your AI provider appears in it like any other, data location, audit rights and exit strategy included: lines a sovereign architecture fills in before anyone asks.

Also called Registre d'information, Registre d'informations

Know your customer: the client identification and verification process the AML-CFT framework imposes on the financial sector, heavy with documents to collect, check and keep current. An agent can build the file, extract supporting documents and flag inconsistencies; onboarding remains a human decision: the agent prepares, the human decides.

Also called Know your customer, Connaissance client

France's framework against money laundering and terrorist financing, set by the Monetary and Financial Code: it imposes due-diligence duties on clients and transactions across the financial sector, with suspicious-activity reports filed with Tracfin. An agent can document the diligence and prepare the files; qualifying a suspicion and deciding to report remain human acts.

Also called Lutte contre le blanchiment de capitaux et le financement du terrorisme, LCB/FT, AML-CFT

The assessment of a borrower's creditworthiness, expressed as a score to inform the credit decision. Applied to natural persons, it falls under the AI Act's high-risk systems (Annex III, point 5(b), fraud detection excepted) and GDPR Article 22: never a credit decision without a human, the agent prepares, the human decides.

Also called Credit scoring, Score de crédit, Scoring crédit

The professional secrecy duty set by article L.511-33 of the French Monetary and Financial Code, which applies to the executives and employees of credit institutions. It binds the bank's own people, not its vendors: with an AI provider, the protection of your data rests on the contract, the GDPR and the architecture you choose.

Also called Secret professionnel bancaire