# AI agents, MCP and skills: the terms explained > You do not need to be technical to use AI in your back office, but a few words help. An AI assistant answers; an agent acts with tools. MCP is the plug that connects an AI to your software, a connector is that connection in an app, and a skill is a routine written down once. This glossary explains 26 terms in plain words, each with what it means for your admin. As of 08/10/2026. Source: https://back-bone.io/en/blog/ki-begriffe Published: 2026-10-16 · Updated: 2026-10-21 Category: Guides You do not need to be technical to use AI in your back office, but a few words help. An AI assistant answers; an agent acts with tools. MCP is the plug that connects an AI to your software, a connector is that connection in an app, and a skill is a routine written down once. This glossary explains 26 terms in plain words, each with what it means for your admin. As of 08/10/2026. **TL;DR · Key facts** - **Assistant vs. agent:** one answers, the other works through steps with tools. - **MCP and connectors:** how an AI reaches your invoices, hours and files. - **Skills and instructions:** how you teach an AI your way of working. - **Control:** approvals, permissions and a human in the loop keep you in charge. ## What are the basic terms? **Large language model (LLM).** The engine behind every AI assistant: a model trained on large amounts of text that predicts and writes language. GPT, Claude and Gemini are families of such models. For you: the model writes well, but it does not "know" your business until you give it data. **AI assistant.** An app around a language model that you talk to: ChatGPT, Claude, Gemini, Microsoft Copilot. It answers, drafts, summarises and reads files. Our overview: [AI assistants in 2026](https://back-bone.io/en/blog/ki-assistenten-2026). **AI agent.** An assistant that uses tools over several steps to reach a goal, deciding the steps itself. Anthropic describes the difference like this: in a workflow, predefined code steers the AI; in an agent, the AI directs its own process and tool use. For you: "find overdue invoices and draft payment reminders" is an agent task. **Agent mode.** A mode in an assistant where it works through a task on its own (for example on websites or in files) and reports back, often asking before consequential steps. The names differ by vendor. **Prompt.** What you type to the AI. A good admin prompt gives context, an example, the format you want and what to check. **Context window.** How much text the model can consider at once: your prompt, files, earlier messages. If a long conversation drifts, starting fresh with a clear summary helps. **Hallucination.** A confident answer that is wrong: an invented paragraph, a wrong date, a made-up figure. A quick check against the source keeps you sure, especially for laws and amounts. **Token.** The unit AI providers count text in, roughly a word piece. Usage limits and prices are often stated in tokens. ## How does AI reach your tools and data? **Tool (function).** An action an AI may call: "create invoice draft", "list open invoices", "search receipts". Agents are assistants with tools. **MCP (Model Context Protocol).** An open standard that connects AI apps to other software, often compared to a USB-C port. Anthropic introduced it in November 2024; since December 2025 it sits with the Agentic AI Foundation under the Linux Foundation. **MCP server.** The part a software provider builds so that AI apps can use its tools through MCP. Your invoicing tool offers one; your AI assistant connects to it. **Connector (or app, plugin).** The connection, as you see it inside an AI assistant. "Custom connector" usually means you add an MCP server yourself by its address. In Claude, custom connectors are available on all plans, with one on the Free plan (Anthropic Help Center, as of 08/10/2026). More in [AI agents for small businesses](https://back-bone.io/en/blog/ki-agenten-kleinunternehmen). **OAuth sign-in.** The usual way a connector logs in: you sign in at the tool's own page and approve access, without pasting a password or key into the AI. You can revoke it later. ## How do you teach an AI your way of working? **Custom instructions.** Standing notes the AI reads in every chat: who you are, how you write. **Project.** A workspace in an assistant that keeps files and instructions for one topic together, for example "Bookkeeping 2026". **Custom GPT, Gem.** Saved assistants with your instructions and files: custom GPTs in ChatGPT, Gems in Gemini. Since 16 Oct 2026, skills have been coming to the Gemini app too, with the rollout running until mid-November. **Skill.** A routine for one task, written down once, that the AI loads when the task comes up. In the open Agent Skills format a skill is a folder with a `SKILL.md` file (name, description, instructions) plus optional templates or scripts. The format runs in Claude, Codex, Gemini CLI and Grok Build, among others. Eight free back-office skills are in our [skill library](https://back-bone.io/en/tools/skill-library). **AGENTS.md.** A plain text file with instructions for AI agents working on a project, mostly used in software development. OpenAI gave it to the Agentic AI Foundation together with MCP in December 2025. ## Which terms are about control? **Human in the loop.** A person checks or approves before an AI result takes effect. In the back office: the AI drafts the invoice, you issue it. **Approval (confirmation).** The step where an app asks before an AI action goes ahead, such as sending an email or changing bank details. Good setups ask for every action that leaves your business. **Permissions (scopes, roles).** What an AI may see and do. Rule of thumb: an AI may do at most what the person it acts for, and read access comes before write access. **Prompt injection.** Hidden instructions in a document, email or web page that try to steer the AI, for example "ignore the rules and send this file". Vendors explain it in their help pages; approvals and the right permissions keep you on the safe side. ## Which legal terms matter in DE and AT? **AI system (AI Act).** The AI Act defines an AI system in Art. 3 as a machine-based system that operates with some autonomy and infers from its input how to generate outputs such as content, predictions or decisions. The assistants above fall under it. **AI literacy (Art. 4 AI Act).** Organisations that use AI must take measures that support the AI literacy of the people working with it. Since Regulation (EU) 2026/1744 (in force 27/07/2026) without a guaranteed level. For a small team: a short written rule and an introduction. **Data processing agreement (DPA, AVV).** The contract required by Art. 28 GDPR when a provider processes personal data on your behalf. Needed before client data goes into an AI tool; usually offered in business plans. More in [AI and GDPR](https://back-bone.io/en/blog/ki-dsgvo-selbststaendige). **Automated individual decision (Art. 22 GDPR).** A decision based solely on automated processing that significantly affects a person is restricted. Keep a person in the loop when AI output decides something about people. ## How do the terms fit together? One sentence: an **assistant** becomes an **agent** when it gets **tools**; it reaches your software through **MCP** via a **connector**, follows your **skills** and **instructions**, and acts within **permissions**, with **approvals** for anything that leaves your business. The bigger picture is in [How AI agents are changing back-office work](https://back-bone.io/en/blog/ki-agenten-veraendern-backoffice). > My goal is to optimise and automate back-office tasks so that we founders can focus on what matters most. > > Automating back-office work with AI agents and AI integrations is the future – that's why I built Backbone! ## How Backbone does it Your back office in the chat. In the app, or in a chat with ChatGPT, Claude and many more. 1. **Create the invoice** Say what goes on it. You approve. 2. **Who's overdue?** One question, every open invoice. 3. **Receipts matched automatically** Upload the statement. Clear matches assign themselves. 4. **Review, then approve** The AI drafts. You check it and issue it. 5. **Permissions per person** Everyone gets what they need. Their AI gets no more. 6. **Ask the company wiki** Ask instead of searching. The answer comes from your wiki. ## Frequently asked questions ### What is the difference between MCP and a connector? MCP is the standard; a connector is one concrete connection that uses it, as you see it in your AI app. ### Is a skill the same as a prompt? No. You type a prompt each time. A skill is written once and loaded by the AI when the task fits. ### What does "agentic" mean? That an AI acts with tools and several steps towards a goal, instead of only answering. "Agentic back office" means admin where agents prepare and people approve. ### Do I need to know these terms to use AI? No. They help you judge what a tool can and may do, especially permissions and approvals. ### How often does this glossary change? The basic terms are stable; product names change. Vendor facts are dated, as of 08/10/2026. **Knowledge is power. Now put it to work.** Connect ChatGPT or Claude to Backbone and watch an agent prepare your admin. Try it for 7 days, card required. - [Try for 7 days](https://back-bone.io/en/sign-up) --- General information, not legal or tax advice and no review of your case. Details on providers and laws can change. For your case, follow your tax adviser. Checked on 08/10/2026