Understanding AI
AI explained without jargon, and what it asks of you.
Eighteen concepts to understand what models, RAG and agents actually do. For each one: a diagram, an example in a business function, what the law says, and the question to ask your supplier.
The diagram to remember
The six layers of controlled AI.
Useful AI answers a business question, on defined data, with rights and evidence. Most projects only build layer 4: that is why they disappoint or worry people.
- 01Business questionsWhat you want to know or do, and how the result will be measured.Governance
- 02Definitions and ownersOften forgottenWhat each word means (an invoice, an active customer) and who answers for it on the business side.GovernanceComplianceEU · GDPR, Art. 5 (accuracy)
- 03Connected dataWhere the information lives, how good it is, and what you are allowed to do with it.AIComplianceEU · AI Act, Art. 10 · EU · GDPR (minimisation)
- 04AIModels, search across your documents (RAG), agents that act in your tools.AI
- 05Rights and human sign-offOften forgottenWho can do what, what AI may not do on its own, and who signs off.GovernanceComplianceEU · AI Act, Art. 14 · EU · GDPR, Art. 22
- 06EvidenceOften forgottenLogs, records, tests: what you need to show at any time what was done, and why.ComplianceEU · AI Act, Art. 12 · EU · GDPR, Art. 30
01
The basics
- LLMLayer 4What is an LLM?A program trained on vast amounts of text that writes by predicting, step by step, the most likely continuation.Read the explainer →
- Tokens and contextLayer 4What are tokens and the context window?A model reads and writes in tokens, fragments of words; its context window sets how many it can handle at once.Read the explainer →
- PromptLayer 1What is a prompt?The text you give a model to get an answer: the question, but also the role, context, rules and expected format.Read the explainer →
- HallucinationLayer 4What is an AI hallucination?A false but plausible answer, stated with confidence: a source, figure or fact that the model makes up.Read the explainer →
02
Connecting AI to your data
- EmbeddingsLayer 3What is an embedding?A list of numbers that captures the meaning of a text: two texts with similar meanings get similar numbers.Read the explainer →
- RAGLayer 4What is RAG?A method that first retrieves the relevant passages from your documents, then asks the model to answer from them, citing its sources.Read the explainer →
- Semantic layerLayer 2What is a semantic layer?A shared reference that defines your business terms, such as “active customer” or “invoice”, with an owner, so people and AI count the same thing.Read the explainer →
- Fine-tuningLayer 4What is fine-tuning?Further training an existing model on your own examples to change its behaviour for good: its style, its format or a specific task.Read the explainer →
03
Agents and automation
- AI agentLayer 4What is an AI agent?An AI given a goal rather than a question: it plans, uses your tools and chains actions together until the job is done.Read the explainer →
- MCPLayer 3What is MCP?An open standard for plugging an AI into your software and data, like a universal socket, without custom development for each tool.Read the explainer →
- GuardrailsLayer 5What are AI guardrails?The set of controls placed around an AI to block what it must not read, say or do.Read the explainer →
04
Reliability and control
- EvaluationLayer 6How do you evaluate an AI?Measuring, on cases prepared in advance, whether an AI gets things right, before go-live and after every change.Read the explainer →
- Human oversightLayer 5What is human oversight of AI?The real ability of a competent person to understand, correct or stop what an AI proposes, before it takes effect.Read the explainer →
- LoggingLayer 6What is AI logging?Automatically recording what an AI does, who asked it, with what data and with what result, so you can prove it later.Read the explainer →
05
Infrastructure and sovereignty
- Open-weight modelLayer 4What is an open-weight model?An AI model whose “weights”, the settings it learned, are published: you can run it yourself, within the limits of its licence.Read the explainer →
- Local AILayer 4What is local AI?Running an AI model on your own servers or computers, without sending your data to an outside provider.Read the explainer →
- Mixture of expertsLayer 4What is a mixture-of-experts model?A model split into blocks, only a few of which work on each word: it computes less, but must keep everything in memory.Read the explainer →
- AI hostingLayer 4Where should you host your AI?Where your AI model runs and reads your data: at a provider, through an API, or in an environment you control.Read the explainer →
An AI project to frame?
A partner reviews your situation and tells you where to start, layer by layer.
Educational explainer, not legal advice. Up to date as of 29 October 2026.