--- title: "AI assistant: definition and example – Carl-Uwe Würfel" description: "AI assistant explained: how assistants like Copilot work, how they differ from chatbots and AI agents, and what matters when a company rolls one out." url: https://carl-wuerfel.de/en/glossary/ki-assistent author: Carl-Uwe Würfel --- 1. [Home](https://carl-wuerfel.de/en) 2. [Knowledge](https://carl-wuerfel.de/en/knowledge) 3. [Glossary](https://carl-wuerfel.de/en/glossary) 4. AI assistant # AI assistant An AI assistant is software built on a language model that completes a task when a person asks, such as writing a draft, and hands the result back for review. ## On this page 1. How an AI assistant works 2. Assistants, chatbots and AI agents differ in who acts 3. Product names blur the line with agents 4. What matters inside a company 5. Data protection and the law 6. What an AI assistant costs and what a good rollout looks like 7. Where the term stops working 8. Frequently asked questions 9. Sources With an assistant, a person triggers every action that has consequences. _Figure: A sequence using the example of a tax advisory firm that uses an assistant inside its email program. A clerk has it summarize a long client email. The assistant drafts a reply listing the documents still missing. The clerk checks deadlines and amounts against the file, edits the draft and sends it herself. Below: the assistant can create drafts but cannot send anything. It delivers suggestions that a person checks._ **Example** A tax advisory firm uses an assistant inside its email program. A clerk has it summarize long client emails and draft a reply listing the documents still missing. She checks deadlines and amounts against the file, edits the draft and sends it herself. The assistant can create drafts but cannot send anything. An AI assistant takes a request in plain language and delivers a result: a draft, a summary, a translation, a spreadsheet formula or an analysis. It works at the person’s pace. Every task starts with a request, and the result goes back to the person, who decides what happens next. The technical core is a [large language model](https://carl-wuerfel.de/en/glossary/llm). The assistant is the application around it, with an interface, the operator’s instructions, access to documents and permissions. Employees meet assistants in two ways: as a standalone chat app such as ChatGPT, Claude or Gemini, or built into software they already use. Microsoft calls its assistants Copilot and offers them in Word, Excel, PowerPoint, Outlook and Teams, among others, for drafting documents, summarizing email threads or taking meeting notes (as of November 2026). People often use “copilot” for any embedded assistant, but it is a product name owned by Microsoft and GitHub. For executives, the AI assistant is usually the workforce’s first real contact with AI. Whether it pays off depends less on the model than on three questions: what knowledge it can see, what data may go into it and who reviews the results. Without an approved assistant, employees easily fall back on personal accounts, and the company ends up with shadow AI. ## How an AI assistant works. An assistant builds its answer from the request and the context the application supplies. That context includes the operator’s instructions in the [system prompt](https://carl-wuerfel.de/en/glossary/system-prompt), the conversation so far and documents the user uploads or the system retrieves to match the question. That retrieval step is called [retrieval-augmented generation](https://carl-wuerfel.de/en/glossary/rag) (RAG). Embedded assistants often gather context on their own. Microsoft describes how Microsoft 365 Copilot draws on email, files, calendars and chats through Microsoft Graph to ground its answers. According to Microsoft, it only surfaces data the individual user has at least view permission for. The result is a suggestion. Language models can produce statements that sound plausible and are wrong. Microsoft itself says generative AI responses aren’t guaranteed to be 100% factual and that users should review output before sending it to others. Good assistants show their sources so a claim can be checked quickly. That leads to the most important rule for daily use: a result only gets passed on after you have checked it yourself. ## Assistants, chatbots and AI agents differ in who acts. [Chatbot](https://carl-wuerfel.de/en/glossary/chatbot) describes a format, a system you exchange messages with. Early chatbots followed fixed rules and decision trees and only recognized phrasings they were built for. Today many chatbots run on language models. Assistant describes a function: it gets tasks done for a person. That can happen in a chat window or through a button like “Summarize” right inside an email. An AI agent pursues a goal over several steps and decides on its own which tool to use next. In a 2024 guide, Anthropic separates workflows, where language models and tools are connected through predefined paths, from agents, where the model directs its own process and tool use. In a 2025 guide, OpenAI defines agents as systems that independently accomplish tasks on the user’s behalf. Applications that integrate language models without using them to control workflow execution, such as simple chatbots, are not agents. A useful rule of thumb: with an assistant, a person triggers every action that has consequences. With an agent, the system does, until it reaches the goal or hits a limit. ## Product names blur the line with agents. Assistants keep gaining tools: web search, creating files, proposing meetings, filling in forms. That moves them closer to agents. Microsoft’s documentation describes agents as scoped versions of Copilot that act as AI assistants and can automate business processes. Copilot decides on its own whether an agent is needed for a response and then sends it a query on the user’s behalf. With Copilot Cowork, Microsoft also offers a feature that carries out tasks across the Microsoft 365 environment on the user’s behalf (as of November 2026). A product’s name therefore says little about how much latitude it has. What matters is what the system can do without fresh confirmation: only read, create drafts, send messages, change data or trigger payments. The OWASP project calls this risk Excessive Agency in its 2025 Top 10 for applications built on language models. The causes are too much functionality, too many permissions and too much autonomy. OWASP recommends minimal permissions and human approval before high-impact actions. Permissions should follow what a system can do, not what it is called. ## What matters inside a company. An assistant with access to company data finds everything the user is allowed to read. That exposes overly broad sharing nobody noticed before, such as a folder of salary data accidentally open to everyone or old project archives. Microsoft explicitly advises reviewing permission models, for example in SharePoint, so the right people have access to the right content. Cleaning up permissions is often the first step before a rollout. The second is the knowledge itself. An assistant writes fluent answers even from outdated or contradictory documents. Process knowledge kept current in one place does more than a bigger model. The third is enabling employees. Article 4 of the EU [AI Act](https://carl-wuerfel.de/en/ai-act), as amended by Regulation (EU) 2026/1744, requires providers and deployers to take measures that support the AI literacy of their staff. Training built on real tasks from each team works better than generic demos. Only then does it become clear which tasks the assistant actually speeds up. ## Data protection and the law. Consumer and business versions differ mainly in contract terms and data use. For Microsoft 365 Copilot, Microsoft commits that prompts, responses and data accessed through Microsoft Graph aren’t used to train foundation models. With consumer versions, terms of use and settings determine whether inputs are used to improve models. That needs checking before any business use. Once personal data goes into an assistant, the vendor processes it on the company’s behalf. That requires a data processing agreement under Article 28 of the EU’s General Data Protection Regulation (GDPR). Without one, customer and employee data don’t belong in the tool. For customer-facing assistants, such as one on a company website, Article 50(1) of the EU AI Act has applied since August 2, 2026. Providers of AI systems that interact directly with people must ensure those people are informed they are dealing with an AI system, unless that is obvious. A company that offers such an assistant under its own name should check whether that makes it a provider. ## What an AI assistant costs and what a good rollout looks like. Assistants are billed in three ways: a license per user per month, usage-based billing for certain features or, for custom solutions, per [token](https://carl-wuerfel.de/en/glossary/token) through the model provider’s API. Vendors change prices constantly. The larger costs often sit outside the license: cleaning up permissions, connecting knowledge sources, training staff and reviewing output. There are three routes to choose from. An assistant embedded in office software uses existing data and permissions. A standalone chat app is quick to roll out but only knows what users give it. A custom solution connects models to internal processes through an intermediate layer and keeps the model swappable. The deciding factor is where the knowledge lives and who may access it. A good rollout starts with specific tasks per team instead of licenses for everyone without a plan. It sets rules for which data may go in, measures which tasks actually get faster and reviews usage. ## Where the term stops working An AI assistant neither decides nor takes responsibility. It delivers suggestions that a person checks, and it only knows what it has access to. It differs from a chatbot, which describes a conversational format, from an AI agent, which acts on its own across several steps, and from the language model that runs inside every assistant. Copilot is a product name, not a technical term. ## Frequently asked questions What is an AI assistant? An AI assistant is software built on a language model that completes tasks when a person asks, such as drafting text, summarizing documents or analyzing data. The person gives the instruction and decides what to do with the result. What is the difference between an AI assistant and an AI agent? The difference is who triggers the next action. With an AI assistant, it is the person. With an AI agent, it is the system itself, across several steps. Both run on the same language models. What is the difference between an AI assistant and a chatbot? Chatbot describes a conversational format, assistant describes a function. A chatbot can be rule-based and only answer questions it was built for, while an AI assistant completes tasks for a user, in a chat or embedded in other software. What is Copilot? Copilot is the product name Microsoft uses for its AI assistants, for example in Word, Excel, Outlook and Teams. GitHub Copilot is an assistant for programmers. In everyday speech, copilot often stands for any assistant built into existing software. Which AI assistant is the best? There is no single best AI assistant. The choice depends on the task, data protection requirements and where the company’s knowledge lives. A test with real tasks from your own team tells you more than any ranking. Are AI assistants free? Free versions exist, usually with limits on usage and features. For business use they often lack a data processing agreement, admin controls and clear commitments on how data is used. Can you use an AI assistant with customer data? Only with a data processing agreement under Article 28 GDPR and clear rules on which data may go in. Business versions usually offer such agreements, consumer versions often don’t. When in doubt, take the question to your data protection officer before rollout. ## Further reading * [**AI glossary**](https://carl-wuerfel.de/en/glossary) * [**AI automation: leeway and sign-off**](https://carl-wuerfel.de/en/ai-automation) * [**AI governance: permissions, access, logs**](https://carl-wuerfel.de/en/ai-governance) * [**AI agents**](https://carl-wuerfel.de/en/glossary/ki-agenten) * [**Prompt Engineering**](https://carl-wuerfel.de/en/glossary/prompt-engineering) * [**LLM: Large Language Model** Term](https://carl-wuerfel.de/en/glossary/llm) * [**RAG: Retrieval-Augmented Generation** Term](https://carl-wuerfel.de/en/glossary/rag) * [**Shadow AI** Term](https://carl-wuerfel.de/en/glossary/schatten-ki) * [**AI guardrails** Term](https://carl-wuerfel.de/en/glossary/guardrails) * [**Chatbot** Term](https://carl-wuerfel.de/en/glossary/chatbot) ## Sources 1. [Microsoft: What is Microsoft Copilot?, Microsoft Learn, as of November 2026](https://learn.microsoft.com/en-us/copilot/plan/copilot-overview) 2. [Microsoft: Data, Privacy, and Security for Microsoft Copilot, Microsoft Learn, as of November 2026](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy) 3. [Anthropic: Building effective agents, 2024](https://www.anthropic.com/engineering/building-effective-agents) 4. [OpenAI: A practical guide to building agents, 2025](https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf) 5. [OWASP: LLM06:2025 Excessive Agency, 2025](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) 6. [European Union: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 4 and 50, 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689) 7. [European Union: Regulation (EU) 2026/1744 amending the Artificial Intelligence Act (Digital Omnibus on AI), new Article 4, 2026](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng) 8. [European Union: Regulation (EU) 2016/679 (General Data Protection Regulation), Article 28, 2016](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679)