AI dictionary for office workers

An umbrella term for technologies that can perform tasks that require human judgment – text recognition, content generation, translation, data analysis. It is not a single technology, but a whole family of approaches.

A subcategory of AI capable of generating new content based on input. Unlike older AI that only classified or predicted, generative AI creates texts, images, presentations, or code. ChatGPT, Copilot, Gemini – these are all examples of generative AI.

A text prompt (instructions, questions, task description) that you write to an AI tool. The quality of the output largely depends on the quality of the prompt. A well-written prompt is like a well-written task to a colleague – the more specific, the better the result.

The name for Microsoft's AI assistant, available directly in tools like Word, Excel, PowerPoint, Outlook, and Teams. Copilot helps you summarize, write, analyze, and automate—without having to switch to another app. It comes in multiple versions: Copilot Chat is included with standard M365, and Microsoft 365 Copilot is a premium paid license.

A premium license beyond the standard M365 that adds deeper AI assistance to Word, Excel, PowerPoint, Outlook, Teams, and other tools. It unlocks features like Copilot in Excel for data analysis, Copilot in Teams for meeting summarization, or Copilot Pages for sharing AI outputs. It differs from the free Copilot Chat.

A product of the American company OpenAI – a web chatbot built on the GPT LLM series. It is not part of Microsoft 365, but is available via a browser or application. Corporate data in the free version can be used to train the model – a paid version or corporate solution is required to work with sensitive information.

Automation (macros, Power Automate flows, scripts) performs precisely specified steps. AI, on the other hand, handles ambiguous situations, learns from examples, and adapts. In practice, these approaches are combined – Copilot in Power Automate combines both.

The set of knowledge and skills needed to understand the principles of AI, use AI tools effectively, critically evaluate their outputs, and be aware of their limitations and risks. AI literacy is becoming as important as digital literacy or Excel.

A set of principles and practices to ensure that AI is deployed fairly, transparently, and safely. It includes questions like: Is AI biased? Who is responsible for its decisions? How do we protect user data? Most large companies (Microsoft, Google) have their own responsible AI standards.

A generic term for AI tools that respond to queries, help with writing, summarizing, or analyzing. It can be a chatbot (ChatGPT, Claude), an in-app assistant (Copilot in Outlook), or a voice assistant. The key difference from an AI agent: an assistant waits for your instruction, an agent acts on its own.

An AI agent is a program that is given a goal and decides how to achieve it – opening apps, searching websites, entering data, sending emails. Unlike a chatbot, where you control every step, an agent works more autonomously. Examples include agents in Copilot Studio or automation in Power Automate with AI.

A structured sequence of steps where AI takes over specific parts of a workflow – for example, it receives an invoice, extracts data, posts it, and archives it. AI workflow combines automation with intelligent decision-making and is the foundation of many corporate AI projects.

A new generation of search engines where AI scans sources and instead of a list of links gives you a direct answer – with citations. Examples: Perplexity, Bing Copilot, Google AI Overviews. Good for quick searches, but requires critical reading – AI can combine sources incorrectly.

A tool that "joins" a meeting, transcribes it into text, and creates a note with tasks and key points. Examples: Copilot in Teams, Fireflies.ai, Otter.ai, or Microsoft Teams Premium. It saves time, but the note always needs to be checked - especially technical terms and names.

The process by which a company gradually integrates AI into its processes and culture. It includes selecting tools, training employees, setting rules, and measuring results. The biggest challenge is not the technology, but changing work habits – which is why adoption is a lengthy process, not a one-time project.

A set of internal rules, policies, and processes that determine who can use AI, for what purpose, how data is protected, and how outputs are controlled. Good AI governance prevents shadow AI (the use of unapproved tools) and reduces legal and reputational risks.

The principle in which AI designs or prepares the output, but the decision or approval remains with the human. This is an important safeguard, especially for sensitive processes - customer communication, legal documents, financial decisions. The opposite is the "fully automated" approach, where AI acts without human control.

Situations where employees use AI tools (ChatGPT, translations, text generators) without the knowledge of IT or company management - and thus may inadvertently share sensitive company data with external services. Analogy to "shadow IT". The solution is not a ban, but clear AI governance with an offer of approved alternatives.

Large Language Model – a system trained on a huge amount of text that can generate and understand natural language. It is the "engine" under the hood of tools like ChatGPT (GPT-4), Microsoft Copilot, or Claude. The name comes from the size of the model, not from being loud.

The ability to formulate prompts so that AI produces the highest quality output. This includes techniques such as specifying the AI's role, specifying a response format, using examples, or breaking down a complex task into steps. It's not programming - anyone can do it.

Unlike a typical chatbot, where AI simply answers questions, an AI agent is able to plan and execute a sequence of actions on its own – search the internet, open a file, send an email, book a meeting. It works more autonomously and requires less supervision. This is where AI is rapidly evolving.

The name for a feature available in tools like ChatGPT, Gemini, or Perplexity that allows AI to autonomously crawl dozens of websites and create a structured research report. It takes minutes instead of hours of manual work. However, the results are only as reliable as online sources – critical verification is required.

A technique where AI first searches a specific database or documents (internal regulations, contracts, manuals) and only then generates a response. The result is AI that "knows" your company documents. The basic principle behind tools like Copilot in SharePoint.

A term for AI models capable of working with multiple types of data at once – reading text, describing an image, analyzing a graph, or transcribing a recording. Modern versions of ChatGPT, Copilot, or Gemini are multimodal.

Each AI model has a limited "memory" for a single conversation. If the conversation exceeds this capacity, the AI will start to forget older parts of the conversation. Therefore, for longer projects, it helps to start a new chat or repeat key information.

A process in which a general AI model is additionally trained on specific data – company communication style, internal documents or industry terminology. The result is a model that "speaks the language of the company". It is a complex and expensive process, and an ordinary worker will not come into direct contact with it.

AI does not process text word by word, but rather "tokens" - roughly 3-4 characters or part of a word. Models have limits on the number of tokens in both input and output. For everyday work, it is not necessary to deal with the details, but it helps to know why AI's "memory" of a conversation has limitations.

A special category of AI models that "think" before answering - breaking down a problem into steps, verifying intermediate results, and looking for errors in reasoning. These include models like OpenAI o3, Claude with extended thinking, or Gemini 2.5 Pro. They are slower, but significantly better at complex analytical or mathematical tasks.

A structured collection of documents, manuals, guidelines, or data that an AI uses as a source of answers. Instead of the AI answering from general training, it draws from a specific knowledge base—a company wiki, SharePoint, or uploaded documentation. A key element for enterprise AI chatbots and assistants.

A technique in which AI generates answers that are grounded in specific documents or data—not just general knowledge from training. Grounding reduces the risk of hallucinations because the AI must cite the source. Copilot in Microsoft 365 is an example of grounded AI—it works with your data from SharePoint, email, and Teams.

A large general AI model trained on a huge amount of data that serves as a basis for creating other, more specialized applications. Examples: GPT-4 (OpenAI), Gemini (Google), Claude (Anthropic), Llama (Meta). Companies build their own products and chatbots on top of them without having to train the model from scratch.

Inference is the process by which a trained AI model processes your input (prompt) and generates an output (response). This is the "operational phase" of AI - as opposed to the training that created the model. The speed and cost of inference are key to scaling AI in enterprises.

The phase in which an AI model "study" a huge amount of data and learns to recognize patterns. Training takes place once (or periodically) and is extremely expensive - GPT-4 is estimated to have cost hundreds of millions of dollars. The average user does not come into contact with the training; they only work with a ready-made, trained model.

AI models whose code and weights are freely available – anyone can download, modify and run them on their own infrastructure. Advantage: data does not leave the company, lower operating costs. Disadvantage: requires technical background. The most famous: Llama (Meta), Mistral, Phi (Microsoft).

Models like GPT-4, Claude, or Gemini Ultra are proprietary – their internal architecture is secret and only accessible through the manufacturer’s API or product. The customer pays for use, but does not have access to the model itself. The opposite is true for open source models.

A tool or platform that allows you to design, test, and run AI agents without programming (or with minimal code). Agents connect to enterprise systems, databases, or APIs. Examples: Microsoft Copilot Studio, Google Vertex AI Agent Builder, ServiceNow AI Agents.

An open standard developed by Anthropic that defines how AI models communicate with external tools – CRM systems, databases, calendars or enterprise software. MCP is "USB-C for AI" – thanks to a unified protocol, one agent can work with many different systems. More and more AI manufacturers are adopting this standard.

An approach where one "master" agent or system coordinates the work of multiple specialized AI agents. Each agent does what it does best - one searches for data, another writes text, and a third checks the result. Orchestration allows for solving complex tasks that one agent would not be able to handle alone.

An area of AI focused on making system decisions transparent and understandable to people. Especially important in regulated industries (banking, insurance, healthcare) where the law requires the company to be able to explain why the AI decided the way it did. Generative AI (ChatGPT) is usually not explainable - it says the result, but not why.

A synthetic digital character with a human appearance and voice, created by AI tools. Used in corporate videos, training materials or as a virtual tutor. Tools like HeyGen, Synthesia or D-ID allow you to create a speaking character from text – without cameras or actors.

A category of AI tools that communicate through voice – both understanding (speech-to-text) and generating natural-sounding voice (text-to-speech). Modern Voice AI (ChatGPT Advanced Voice, Gemini Live) handles natural conversation including intonation and emotion. Applications: customer support, voice assistants, accessibility.

Technology that converts audio recordings or live voice to text. Modern AI tools (Whisper by OpenAI, Copilot in Teams) handle transcription with high accuracy even in Czech. Uses: meeting transcripts, video subtitles, text dictation or making content accessible to the deaf.

Technology that generates spoken voice from text. Modern AI versions (ElevenLabs, Azure Neural Voice, OpenAI TTS) sound natural and handle emotions, intonation, and voice cloning. Uses: audio versions of articles, e-learning courses, content accessibility, or automatic voice responses.

Generative AI capable of creating an image (illustration, photo, logo, infographic) from a text input. Tools: DALL-E (integrated in ChatGPT and Copilot), Midjourney, Adobe Firefly, Stable Diffusion. Suitable for quick visual designs, presentation graphics or design prototypes - but the copyright on the outputs is still legally unclear.

A rapidly developing area of generative AI, where a model creates a short video from a text input or image input. Tools: OpenAI Sora, Google Veo, Runway, Kling. Outputs are short (seconds to minutes) and require careful control, but the speed of development is extraordinary.

An arrangement where multiple specialized AI agents work together, each on a different part of a task. One agent searches for data, another analyzes it, a third writes a report, and a fourth sends it. The result is faster and of higher quality than the work of a single generalist agent. See also: AI orchestration.

An architecture where multiple AI agents communicate, delegate tasks, and coordinate their work. Each agent has a defined role and tools. Multi-agent systems handle complex processes that a single agent would not be able to keep in context. More and more corporate AI projects are using this architecture.

An agent capable of planning and executing longer sequences of actions without ongoing human guidance. It is given a goal and decides how to achieve it. Requires careful setting of boundaries (guardrails) – an autonomous agent can make a mistake that is difficult to correct. Human supervision is still recommended during deployment.

The technical capability of an AI model, where instead of just writing text, it "calls" an external function or tool - creates a meeting in the calendar, adds a task to Planner, pulls data from CRM. Thanks to this, AI stops being just a chatbot and becomes an active tool in the work process. Function calling is a technical term, tool calling is a more general term.

Expanding the context window to hundreds of thousands or millions of tokens – equivalent to hundreds of pages of text. Long context models can process an entire contract, manual or set of documents at once. Examples: Gemini 1.5 Pro (1M tokens), Claude (200K tokens). Opens up new ways of working with corporate documentation.

A technical process in which AI converts text (sentence, paragraph, document) into a set of numbers representing its meaning. Texts with similar meaning have similar numbers - thanks to this, AI search finds relevant documents even without an exact keyword match. Embeddings are used in RAG systems and corporate chatbots for documentation.

A special type of database optimized for storing embeddings and fast searches by similarity (not exact matches). It is the basis of RAG systems and corporate AI knowledge bases. Examples: Pinecone, Weaviate, Azure AI Search, pgvector. The average user does not work with it directly, but it is behind every smart corporate chatbot.

The process of verifying that AI responds correctly, consistently, and safely – before deployment and continuously in operation. Evals include test sets of questions and answers, evaluation by humans or another AI model. Crucial for companies that deploy AI in customer contact or decision-making processes.

A set of standard tasks and tests on which the performance of AI models is compared. Examples: MMLU (knowledge), HumanEval (coding), GPQA (science). Benchmark results are used in marketing by companies such as OpenAI, Google, or Anthropic. Note: the score in the benchmark may not correspond to the performance on your specific company task.

AI or algorithmically generated data – instead of collecting real data. Used when real data is lacking, privacy is important, or specific edge cases are needed. For example, instead of using real medical records, realistic synthetic data with the same statistical properties is generated.

Unlike large LLMs (billion parameters, cloud), SLM models are more compact – they can run on a laptop, tablet or industrial device. They are cheaper to operate, faster and do not require an internet connection. Examples: Microsoft Phi-4, Gemini Nano, Llama 3.2 (3B). Ideal for specific corporate tasks or environments with data security requirements.

Three related terms for AI running outside the public cloud: Edge AI runs directly on the device (phone, camera, machine). Local AI runs on your own computer or server. On-premise AI runs on a company’s own data infrastructure. Common denominator: data doesn’t leave your environment – crucial for banks, hospitals or manufacturing.

Technical or procedural constraints that prevent AI from behaving in undesirable ways – sharing sensitive data, generating inappropriate content, responding outside its scope, or being misused. Guardrails are part of every enterprise AI deployment and form a key layer of AI governance. They can be set at the system prompt, in code, or by the platform.

The process in which a team (internal or external) intentionally attempts to break, trick, or exploit an AI system before it is deployed. The goal is to find ways to trick the AI into sharing prohibited information, ignoring security rules, or providing malicious content. The results are used to strengthen guardrails. A method borrowed from cybersecurity.

A layer between the user and AI models that decides which model to use based on the nature of the query. Simple questions go to a cheaper, faster model, complex analyses to a more powerful one. Saves costs and speeds up responses. Similar principle to smart routing in telephony.

Each use of an AI model (generating a response) costs something – typically charged based on the number of tokens in input and output. For cloud models (GPT, Claude, Gemini) these are direct API fees. For enterprise deployments, this is a key operational indicator. Inference costs are rapidly decreasing – 2025 models are 100 times cheaper than 2023 models.

AI that "knows" its user - their documents, emails, communication style, preferences and work context. Unlike a generic chatbot, personal AI responds in a personal style and works with personal data. Examples: Copilot in M365 with access to your OneDrive and Outlook, or Claude Projects with uploaded documents.

Similar to SEO (search engine optimization), but focused on getting your content cited and recommended by AI tools like ChatGPT, Perplexity, or Gemini. It involves writing structured, factually accurate content with clear answers to specific questions. Increasingly important as more people search via AI instead of Google.

One of the most well-known AI companies in the world, behind the GPT-4o and o3 models. It operates the ChatGPT chatbot (available at chat.openai.com) and an API that is used by thousands of other applications. Microsoft has invested billions of dollars in OpenAI, and their technology also powers Microsoft Copilot.

Gemini is the name for both a family of AI models from Google (Gemini 2.5 Pro, Flash…) and a chatbot available at gemini.google.com. It integrates with Google Workspace (Docs, Gmail, Sheets) similar to how Copilot integrates with Microsoft 365. Strengths: working with long documents and searching connected to Google.

Anthropic is an American AI company (claude.ai) focused on AI security. Their assistant Claude is among the top tools for working with long texts, document analysis, and complex writing. Claude is less well-known than ChatGPT, but has gained a strong position among power users and companies.

Perplexity (perplexity.ai) combines web search with generative AI - instead of a list of links, you get a directly summarized answer with source citations. It's great for quick research, news tracking, or fact-checking. It works in real time, unlike basic ChatGPT or Claude without web access.

Meta (the company behind Facebook and Instagram) develops its own AI models called Llama, which are open-source – their code is publicly available and companies can run them on their own infrastructure. Meta AI is integrated into WhatsApp, Instagram and Facebook. Llama is popular with companies that don’t want to send data to a third-party cloud.

xAI is Elon Musk's AI company that develops Grok models. Grok is integrated into the X platform (formerly Twitter) and also available as a standalone chatbot. It is known for its straightforward style and access to up-to-date information from X. Less widespread in the corporate environment than ChatGPT or Copilot.

French company Mistral AI develops powerful and efficient language models available as open-source and commercial APIs. It is considered a leading European alternative to American AI companies. Popular with tech-savvy teams looking for cost-effective data solutions in Europe.

NotebookLM (notebooklm.google.com) is a tool from Google where you upload your own documents (PDFs, texts, websites) and AI answers questions, creates summaries, or generates audio podcasts from them. The AI does not work with a general knowledge base, but exclusively with your uploaded materials - the outputs are anchored in specific sources.

An informal term for content that is generated by AI but is superficial, generic, or full of clichés. "Slop" is English for "cinder". Typical signs: excessive formality, empty phrases, missing concrete facts. The problem is not the AI itself, but the uncritical acceptance of its outputs.

The practice of companies adding the word "AI" to their products even though actual AI plays little or no role. Analogous to "greenwashing" in the area of sustainability. Helps to be skeptical of supplier claims: ask what specific AI technology the product uses.

A model that simulates a physical object, building or process (factory, air conditioning, supply chain) in real time. A digital twin is not AI per se, but AI is increasingly involved in it. Less relevant for the office environment yet, but the concept is spreading from industry to everyday communication.

Artificial General Intelligence – a hypothetical system capable of handling any intellectual task as well or better than a human. Today's AI (ChatGPT, Copilot) is not AGI – they are specialized systems. AGI is still more of a philosophical discussion than a technical reality, but the media and companies are talking about it more and more.

An approach where a company automatically considers AI as a first option when designing new processes, products or services – not as an add-on. AI-first does not mean that AI will replace everything, but that it becomes the starting point of the design. It is used as a corporate strategy and a marketing claim.

A designation for an organization that has the conditions ready for AI deployment: high-quality and accessible data, set rules for using AI, trained employees, and technical infrastructure. AI-ready is a goal that companies are moving towards – most organizations are not there yet, even though they are already using AI tools.

An umbrella term for any AI-generated content – written articles, AI voices, generated photos, videos or music. Not inherently negative (used in e-learning or marketing, for example), but it does present challenges in terms of labelling, copyright and trustworthiness. The EU AI Act requires labelling of synthetic content.

An approach to software development where a person describes what they want to create, and AI (Copilot, Cursor, Claude) writes the code. A programmer or even a non-programmer then tests the result and iterates with additional instructions. Democratizes the creation of simple tools and automation. Not suitable for critical systems without professional code review.

A product or application that "wraps" an existing AI model (GPT, Claude) and adds its own interface, rules, or integration. Technically, it is an API call to a foreign model. The vast majority of AI startups and tools are wrappers. The value lies in the UX, specialization, or integration - not in the model itself.

A category of AI tools focused on repeated, personalized interaction with the user – as a digital coach, mentor, development guide or social partner. Examples: Pi.ai, Character.ai or specialized coaching AI. Ethically sensitive area – carries risks of dependency or replacing human relationships.

Processes and tools for monitoring how AI behaves in real-world operations – whether it responds correctly, is not abused, has performance issues, or deviates from expected behavior. Especially important for agents and chatbots in customer contact. Part of responsible AI and AI governance.

AI-generated video or audio where a real person is shown saying or doing something they never actually did. Deepfakes are becoming more convincing and their creation is becoming more accessible. Risks: disinformation, CEO fraud, reputation damage. Detection is possible, but increasingly difficult.

A category of tools (GPTZero, Turnitin AI, Originality.ai) that promise to detect AI-generated content. The problem: their accuracy is questionable – they label human-written texts as AI and vice versa. Reliable detection of AI content does not yet exist. The results of these tools should not be relied upon as evidence.

AI can confidently claim things that are false – citing non-existent sources, bad data, or fabricated quotes. This is not a deliberate lie – the model is „filling in the blanks" based on patterns, not facts. AI output should always be verified, especially facts, numbers, and quotes.

Never enter personal customer or employee data (names, emails, social security numbers) into AI tools without the knowledge of the legal department. GDPR also applies to data processing via AI. Secure corporate tools like Copilot in M365 have contractually treated processing terms – but even there, corporate rules apply.

The right of a company or state to control where its data is stored and how it is processed – including data fed into AI tools. It is crucial for companies to know whether their data is leaving the EU, whether it is used to train models and who can access it. With M365 Copilot, the data remains in the EU with the tenant.

A security risk where an attacker inserts hidden instructions for an AI into a text (email, document, web page). If the AI processes this document, it can execute the instructions – send sensitive data or change responses. Particularly relevant for AI agents working with external content.

A technique where a user formulates a prompt in such a way that the AI bypasses its security rules – for example, making it write content that it would otherwise reject. Jailbreaks are spread online as tutorials. For companies deploying AI chatbots, protection against jailbreaks is part of security testing (see AI red teaming). Model manufacturers continuously respond to jailbreaks with patches.

A label for products, companies or processes where AI is not an afterthought, but a fundamental building block from day one. It differs from AI-first (a strategic approach) in that AI-native describes the architecture and design. Examples: Cursor (AI-native code editor), Perplexity (AI-native search engine). The opposite are traditional products with AI "glued" on top.

A vision of a future where AI agents autonomously perform economic activities – purchasing services, outsourcing work to other agents, concluding contracts or managing budgets. Agents become both "digital employees" and "customers". The concept is still largely conceptual, but the first steps (agents paying for API services) are a reality.

An approach to work where AI takes over routine, repetitive or data-intensive tasks and humans focus on judgment, creativity, relationships and ethical decisions. The goal is not to replace humans with AI, but to enhance overall performance by combining the two. The opposite of the "AI vs. humans" view. A fundamental principle for responsible and sustainable AI deployment.

A concept where companies operate a "digital workforce" alongside human employees - AI agents and automated systems responsible for specific processes. A digital workforce never sleeps, doesn't need vacation, and scales instantly. The concept is gaining traction especially in the areas of customer service, finance, and HR. It sparks discussions about the impact on employment.

An API (Application Programming Interface) is an interface through which one system communicates with another. In the context of AI, this means that your CRM, website, or internal tool can "call" an AI model (e.g. GPT or Claude) and get a response back - without the user having to open ChatGPT. Most enterprise AI integrations work through APIs.

A situation where AI systematically favors or disadvantages certain groups, topics, or outcomes – because the data it was trained on was not balanced. Bias can be gender, racial, cultural, or industry-specific. The problem is not always visible at first glance, but it can have serious consequences when using AI in HR, lending, or customer communications.

A technique (and a natural feature of reasoning models) where AI does not answer the question straight away, but first "thinks out loud" - it describes the solution procedure step by step. The results are more accurate and easier to verify. It can be invoked in prompts with the instruction "think step by step" or "show your procedure". The basis of modern reasoning models such as o3 or Gemini 2.5 Pro.

An AI model parameter on a scale typically 0–2. A low temperature (0–0.3) produces consistent, predictable responses – suitable for facts, summaries or code. A high temperature (0.8–2) produces more creative and diverse outputs – suitable for brainstorming or creative writing. It can be set directly in tools like Copilot Studio or the OpenAI API.

A subfield of machine learning that is behind most of the modern AI breakthroughs – speech recognition, image translation, and generative AI. Deep learning uses multilayer neural networks capable of automatically finding patterns in huge data sets. ChatGPT, Copilot, and Gemini are built on deep learning – specifically, on an architecture called Transformer.