A2A (Agent2Agent): A protocol for AI agents from different companies to discover each other and work together. Each agent has a card at /.well-known/agent.json.
AI agent: An AI program that doesn't just answer but carries out tasks with tools: writes emails, books appointments, updates the CRM.An agent that replies to an enquiry and books a meeting in the salesperson's calendar.
AI employee: An AI agent set up for one specific role in the company, with clear rules on what it decides alone and what it hands to a person.A quoting AI employee that prices from the price list and sends the quote for approval.
API: A door through which one program talks to another. When your software has an API, an AI agent can work with it directly.
Automation: Work that runs on its own by rules, without a person. With AI, automation also covers tasks that require reading and understanding text.
EU AI Act: The European regulation on artificial intelligence (2024/1689). Among other things it requires people to know when they are talking to an AI.
GDPR: The European regulation on personal data protection. It also applies when personal data is processed by AI.
GEO: Generative engine optimisation: preparing your site and presence so AI assistants understand the company and recommend it.
Hallucination: When an AI model confidently states something false. It is reduced when the agent answers only from your data and says “I don't know” when it has no answer.
Human in the loop: A rule under which an AI's important actions (paying, sending, deciding) wait for a person's approval.
LLM (large language model): The model behind ChatGPT, Claude and similar tools: trained on huge amounts of text, it understands and writes human language.
llms.txt: A text file at the root of a site with a short description of the company and links to its key pages, written for AI models.
Local model: An AI model that runs on your own server or computer with no data leaving it. Usually an open model such as Llama, Qwen or Mistral.
MCP (Model Context Protocol): An open standard for connecting AI assistants to software, data and tools. The server describes what it can do and the assistant calls it.Claude audits a site and books a call through this site's MCP server.
Open-weight model: An AI model whose weights are public so it can run on your own hardware, for example Llama, Qwen or Mistral.
Prompt: The instruction given to an AI model: what to do, in what tone and by which rules.
RAG: A way for AI to answer from your documents: it first finds the relevant passage, then answers only from it and shows the source.An internal assistant that quotes the exact clause of a contract.
SEO: Search engine optimisation: preparing a site so Google understands it and ranks it high when people search for your service.
Structured data: A description of the company, services and questions in schema.org format (JSON-LD) that search engines and AI read without guessing.

Glossary of AI terms in plain language