Enterprise AI glossary
- Updated
- By
- The MagnolIA team
- On this page
- 61 terms
Short answer
This glossary defines, in one to three sentences each, the terms that come up when a company adopts AI: where it runs (on-premise, cloud, air-gapped), what it runs on (LLMs, open weights, GPUs, RAG, agents) and under which rules (GDPR, Spain's ENS, the EU AI Act). Each term links to the MagnolIA page that covers it in depth.
Agentic AI
AI systems able not only to answer questions but to act autonomously to reach a goal; in practice, by chaining several steps together. The first part is how Spain's data protection authority, the AEPD, described it when presenting its February 2026 guidance.
AI agent
A program that uses a language model to carry out a task step by step: it reads, decides, uses tools and checks the result. In MagnolIA Nexus, agents work as a team and actions with side effects wait for approval.
AI server
A machine with graphics cards and the software needed to run AI models and serve several people at once through a single entry point, with users, permissions and logging.
AI usage policy
An internal document that says which AI tools are approved, what information may and may not be given to them, and who reviews the output. It is the usual answer to unapproved use (shadow AI).
Air-gapped
Describes an environment whose network is isolated from the internet. AI can work there if the model runs inside it and does not depend on any external service.
Anonymisation
Irreversibly removing whatever links a piece of data to a person. In everyday use the word also covers replacing the personal data in a document with placeholders before giving it to an AI.
Automated decision
A decision about a person taken solely by automated processing, with no human involvement, that produces legal effects for them or affects them in a similar way. Article 22 of the GDPR gives people the right not to be subject to one, with exceptions.
Chained audit log
A log in which each entry is linked to the previous one, so that any later modification can be detected. MagnolIA Server keeps its log for two years, inside the client's installation.
Cloud AI
AI used as a service over the internet: every prompt is processed on the provider's servers, and it is usually billed per user or per use.
Coding assistant
A tool that uses a language model inside the code editor to complete, explain, review and write code. If the model runs on your own server, the source code does not leave the network.
Context window
The maximum amount of text, measured in tokens, that a model can take into account at once: the prompt, the attached documents and the earlier conversation.
Controller
Under the GDPR, whoever decides the purposes and means of processing personal data (Article 4(7)). A company that decides to use AI on its clients' data is the controller, wherever the model runs.
Data protection impact assessment (DPIA)
The prior analysis that Article 35 of the GDPR requires where processing is likely to result in a high risk to people's rights and freedoms, in particular when it uses new technologies.
Data protection officer (DPO)
The person who advises an organisation on data protection and monitors how it is applied. Article 37 of the GDPR makes the appointment mandatory for public authorities and for those who process special categories of data on a large scale, among other cases.
Data residency
The physical place (country or region) where data is stored and processed. A promise of EU residency says where the servers are, not who can access them.
Deployer
Under the EU AI Act, whoever uses an AI system under its own authority. A company that uses AI in its work is a deployer.
Embeddings
A numerical representation of a text that makes it possible to compare meanings. They power semantic search: finding a document by what it says rather than by its exact words.
ENS (Spain's National Security Framework)
Esquema Nacional de Seguridad, set out in Spain's Royal Decree 311/2022. It lays down the security requirements for the information systems of the Spanish public sector and of the private entities that provide services to it.
EU AI Act
Regulation (EU) 2024/1689, the European law on artificial intelligence. It places obligations on those who provide and those who use AI systems, graded by the risk of each use.
Fine-tuning
Additional training of an existing model on your own examples, to adapt it to a type of task or a specific vocabulary.
GDPR
General Data Protection Regulation, Regulation (EU) 2016/679. The European law governing the processing of personal data, including when it is done with AI.
General-purpose AI model
The EU AI Act's category for models that can perform a wide range of tasks, such as large language models. Its obligations fall on whoever provides the model and have applied since 2 August 2025.
Generative AI
Artificial intelligence that produces new content (text, code, images, audio) from an instruction, rather than only classifying or predicting. Language models are its most widespread form in office work.
GPU
Graphics processing unit. Graphics cards perform, in parallel, the calculations language models need; how many there are and how much memory they have determine which models fit and how many people they can serve.
Hallucination
A false or made-up answer that a language model presents as if it were certain: a citation, a figure or a rule that does not exist. Making the model answer from specific documents reduces it but does not remove it, which is why a person reviews the output.
Hardware as a service (HaaS)
A model in which you do not buy the machine: the provider installs it and includes it in a fee together with the software and maintenance. At MagnolIA the fee is fixed and does not change with usage.
Human oversight (human in the loop)
A way of working in which a person with judgement and authority reviews what the AI proposes before it takes effect. The EU AI Act provides for it in high-risk systems.
Hybrid AI
A split in which part of the work is done with models inside the network and part with cloud models. The usual criterion is the data: what is sensitive stays inside, and what goes out goes without personal data.
Inference
The moment an already trained model is used to produce an answer. It is the work the hardware does every time someone asks something, and what consumes computing capacity.
LLM (large language model)
An AI model trained on large amounts of text that generates language: it answers questions, summarises, drafts, translates and writes code.
Local AI
An AI model running on your own computer or on an office server. It describes where the model runs, not how the service is managed.
Managed on-premise AI
On-premise AI where the provider handles everything technical: it chooses the hardware, installs it, maintains the models, monitors the service and provides support. The company uses AI without administering servers.
MCP (Model Context Protocol)
An open protocol for connecting a model or an agent to external tools and data sources through connectors.
Multimodal model
A model that works with more than one kind of content: images, audio or scanned documents as well as text.
On-premise AI
Artificial intelligence that runs on hardware installed inside a company's own network, rather than on an outside provider's servers. The models, documents, prompts and answers all stay on that hardware.
Open source
Software whose source code is published under a licence that allows anyone to use, study, modify and redistribute it. It is not the same as open weights: publishing a model's parameters does not include the code or the data it was trained with.
Open-weight model
A model whose parameters (the "weights") are published and can be downloaded and run on your own hardware. Open-weight models are what make on-premise AI possible.
OpenAI-compatible API
A programming interface that accepts the same requests as the OpenAI API. It lets tools written for that API work against another server by changing only the address and the key. MagnolIA Server also offers one compatible with Anthropic's.
Parameters
The numbers a model adjusts during training and that determine how it responds; also called weights. Their count, in the billions, gives an idea of the model's size and of the memory it needs.
Personal data
Any information about an identified or identifiable natural person: a name, an ID number, a phone number, an address or a number plate.
Private AI
AI where documents, prompts and answers never reach third parties. It describes the outcome; you can get there with a contract, with a reserved cloud or by running the models in your own network.
Processor
Under the GDPR, whoever processes personal data on the controller's behalf (Article 4(8)), under a contract that governs it (Article 28). A cloud AI provider that processes a company's documents normally acts as a processor.
Prompt
The text given to the model: the question or instruction, together with any attached documents. With on-premise AI the prompt does not leave the network.
Prompt injection
An attack in which text the model reads (a web page, an email, a document) carries hidden instructions meant to divert it from its task. It mostly affects agents that can act, which is why actions with real effects are best left waiting for approval.
Proprietary model
A model whose weights are not published. It can only be used as a service in its maker's cloud; it cannot be installed on your own hardware.
Pseudonymisation
Replacing personal data so that it can no longer be attributed to a person without additional information that is kept separately (Article 4(5) GDPR). It is reversible for whoever holds that information.
Quantisation
A technique that stores a model's parameters with fewer bits so that it takes up less memory and fits on more modest hardware, in exchange for a loss of precision that is usually small.
RAG (retrieval-augmented generation)
A technique in which, before answering, the system searches the company's documentation for the relevant passages and hands them to the model, so that it answers from them.
Self-hosted
Software or a model that an organisation installs and runs on machines it controls, instead of using it as someone else's service. On-premise AI is self-hosted AI inside the organisation's own network.
Semantic search
Search that finds documents by meaning rather than by exact wording: asking about “early termination” also returns the clause that talks about “ending the contract before term”. It relies on embeddings.
Shadow AI
Employees using AI tools that the organisation has not approved. The usual risk is pasting documents with personal or confidential data into them.
Sovereign AI
AI where the data, the infrastructure and the decisions about both remain under the control of the organisation or the country, without depending on a third party's jurisdiction.
Special-category data
Personal data given reinforced protection by Article 9 of the GDPR, such as health data, biometric data or data revealing political opinions or trade union membership.
Speech-to-text (automatic transcription)
Turning a voice recording into text with a speech model. If that model runs on the office server, the audio of the meeting does not leave the network.
SSO (single sign-on)
A mechanism that lets a person sign in to several applications with their company account, without a separate password for each.
Telemetry
Data about how a program or a machine is running, sent to whoever makes or maintains it. It is worth knowing what it includes: the state of the machine (load, health, security) is one thing, the content of what people ask it is another.
Temperature
A setting that controls how much a model varies when choosing each word. A low value makes answers more predictable and repeatable; a high one, more varied.
Token
A fragment of text, a word or part of a word, that the model reads or writes. Cloud services usually bill per token; on your own hardware there is no per-token bill.
Training data
The text and other material a model is built from. Using an already trained model does not change it; what a cloud provider later does with the conversations it stores, including training on them, depends on its terms.
Two-factor authentication
A second identity check on top of the password, such as a temporary code. At MagnolIA it is mandatory for system administration.
Vector database
A store that holds the embeddings of your documents so that the passages closest to a question can be found quickly. It is the part of a RAG system that works as the index.
Is a term missing? Ask us
If there is a word you do not understand in a proposal or a regulation, write to us. And if you would like to see AI working inside your network, we can look at it with your case in front of us.
Request a pilot


