00 — Translate
Same concepts, different names
The industry is converging on these building blocks.
| Concept | Claude | ChatGPT | Microsoft Copilot | Gemini | Perplexity | Grok |
|---|---|---|---|---|---|---|
| Strengths | Reasoning and judgement | Broad use and execution | Microsoft Suite native, Outlook | Image generation, Google Suite native | Real-time events & research | Live web and X |
| Conversational chat | Chat | Chat | Copilot Chat | Gemini app | Search | Grok chat |
| Persistent workspace with files | Projects | Projects, Space | Notebooks / Pages | Gemini Notebook | Projects | Workspaces |
| Editable content surface | Artifacts | Pages (Space) | Pages | Canvas | Files and apps | Canvas |
| Reusable custom assistant | Skills | Plugins | Agents (Copilot Studio) | Skills | Project instructions | Skills |
| Deep research | Research | Deep Research | Researcher | Deep Research | Deep Research | Search, multi-agent |
| Reasoning / “think” mode | Extended thinking | Reasoning | Think Deeper | Deep Think | Model Council | Deep reasoning |
| Voice mode | Voice | Advanced Voice | Copilot Voice | Gemini Live | Voice | Voice |
| Image / media generation | Artifacts | Images | Designer | Imagen, Veo | Image generation | Imagine |
| Memory / personalization | Memory | Memory (Dreaming) | Copilot Memory | Personal Intelligence | Memory, Brain | Memory |
| Desktop / computer-use agent | Cowork, Claude Duet | Dots, ChatGPT Work | Copilot Cowork | Gemini Agent | Computer, Comet | Grok Bot |
| Code & data analysis | Claude Code | Codex | Python in Excel | Code Assist, Colab | Computer | Grok Build |
| External system connectors | Connectors (MCP) | Connectors, plugins | Graph connectors | Connected Apps | Connectors | Plugins, web and X search |
| Agent builder / no-code agents | Agent SDK | Workspace Agents, Agents SDK | Copilot Studio | Gemini Enterprise Agent Platform | Agent API | Multi-agent API |
| Embedded in daily apps | Microsoft 365, Chrome | Desktop, web, and mobile apps | Native in Word/Excel/PPT/Outlook, Edge | Workspace: Docs/Sheets/Slides/Gmail, Chrome | Comet browser | grok.com, X, and the apps |
Which one, when?
| Tool | Best use | Don’t use it for |
|---|---|---|
| ChatGPT | Execution, file work, drafting, coding, multimodal workflows, getting things done fast | Subtle taste work where you want a long strategic sparring partner |
| Claude | Deep synthesis, judgment, long-context thinking, writing taste, strategic pressure-testing | Fast tool-heavy execution when speed and integrations matter more |
| Microsoft Copilot | Work inside Word, Excel, PowerPoint, Outlook, Teams, and Microsoft files | General creative thinking outside the Microsoft ecosystem |
| Gemini | Gmail, Docs, Sheets, Meet, NotebookLM, Google Drive context, research across Google Workspace | Polished strategic writing if you prefer Claude’s taste |
| Perplexity | Fast cited research, current-topic scanning, and turning findings into a report or file | Final judgment, sensitive internal work, or polished writing |
| Grok | Live web and X search, fast brainstorming, and ongoing tasks you hand to Grok Bot | A client-ready deliverable without a person checking the facts and the tone |
01 — Basics
Core Concepts
| Term | Plain-English meaning | Why it matters |
|---|---|---|
| Artificial Intelligence (AI) | Software that performs tasks that usually require human judgment, pattern recognition, language, or decision-making. | The umbrella term for tools that can analyze, generate, classify, recommend, or automate work. |
| Machine Learning (ML) | A type of AI where systems learn patterns from data instead of being explicitly programmed for every rule. | Powers forecasting, recommendations, fraud detection, pricing models, and many modern AI systems. |
| Deep Learning | Machine learning that uses layered neural networks to detect complex patterns. | Drives image recognition, speech recognition, translation, and many large language models. |
| Neural Network | A model loosely inspired by the brain, made of connected layers that transform inputs into predictions or outputs. | The foundation of many advanced AI systems. |
| Generative AI | AI that creates new content, such as text, images, code, audio, video, or synthetic data. | Useful for drafting, brainstorming, design, customer support, research, and workflow acceleration. |
| Predictive AI | AI that estimates what is likely to happen based on historical data. | Common in forecasting, credit risk, demand planning, churn prediction, and investment research. |
02 — LLMs
Large Language Models
| Term | Plain-English meaning | Why it matters |
|---|---|---|
| Large Language Model (LLM) | An AI model trained on large amounts of text to understand and generate language. | The engine behind tools like ChatGPT, Claude, Gemini, and many AI assistants. |
| Foundation Model | A large general-purpose AI model that can be adapted to many tasks. | Companies often build specialized apps on top of foundation models rather than training from scratch. |
| Multimodal Model | An AI model that can work with more than one type of input or output, such as text, images, audio, video, or files. | Enables use cases like analyzing PDFs, interpreting charts, reading screenshots, and generating media. |
| Context Window | The amount of information an AI model can consider at once during a conversation or task. | A larger context window lets the model work with longer documents, more history, or bigger datasets. |
| Token | A small unit of text processed by an AI model, often part of a word or phrase. | AI usage and cost are often measured in tokens. |
| Prompt | The instruction, question, or context given to an AI model. | Better prompts usually produce better outputs, especially when they include role, goal, context, constraints, and examples. |
| System Prompt | Higher-priority instructions that guide how an AI assistant should behave. | Helps enforce tone, safety rules, workflow steps, and brand or company standards. |
03 — Under the hood
How AI Produces Answers
| Term | Plain-English meaning | Why it matters |
|---|---|---|
| Training | The process of teaching a model patterns from large datasets. | Determines the model’s general capabilities and limitations. |
| Inference | The process of using a trained model to produce an answer, prediction, or output. | This is what happens when a user asks an AI tool a question. |
| Fine-Tuning | Further training a model on a narrower dataset to improve performance on specific tasks. | Can make AI better at a company’s tone, domain, or workflow, but requires good data and governance. |
| Embedding | A numerical representation of text, images, or other data that captures meaning and similarity. | Used for search, recommendations, clustering, and retrieval-based AI tools. |
| Retrieval-Augmented Generation (RAG) | A method where an AI system retrieves relevant documents or data before generating an answer. | Helps AI answer using company files, policies, research, or databases without retraining the model. |
| Vector Database | A database designed to store and search embeddings. | Often used in AI search, knowledge-base assistants, and RAG systems. |
| Agent | An AI system that can plan steps, use tools, and take actions toward a goal. A personal agent does this across your own email, calendar, files, and messages. | Useful for workflows like research, reporting, customer support, coding, scheduling, and data analysis. |
| Tool Use | When an AI model calls external tools, such as search, calculators, databases, email, calendars, or code execution. | Makes AI more useful and more accurate because it can act on real systems or fetch current information. |
04 — Risk & governance
Quality, Risk, and Governance
| Term | Plain-English meaning | Why it matters |
|---|---|---|
| Hallucination | When an AI produces information that sounds plausible but is wrong or unsupported. | A major risk in research, legal, finance, medical, and customer-facing use cases. |
| Grounding | Tying an AI answer to reliable sources, documents, data, or citations. | Reduces hallucinations and makes outputs easier to verify. |
| Evaluation (Eval) | A structured test of how well an AI system performs on specific tasks. | Helps teams compare models, monitor quality, and catch regressions. |
| Benchmark | A standardized test used to compare AI model performance. | Useful for directional comparison, but real business tasks often need custom evaluations. |
| Bias | Systematic unfairness or skew in AI outputs due to data, design, or deployment choices. | Can create legal, reputational, and ethical risk. |
| Explainability | The ability to understand why an AI system produced a result. | Important for regulated or high-stakes decisions. |
| Human in the Loop | A workflow where humans review, approve, or correct AI outputs. | Helps manage risk while still gaining productivity. |
| Guardrails | Rules, filters, or controls that limit what an AI system can do or say. | Protects against unsafe outputs, data leaks, policy violations, and brand risk. |
| Data Privacy | Protecting sensitive, personal, or confidential information used with AI systems. | Essential when AI touches customer data, company strategy, financials, or employee information. |
05 — Business
Business and Implementation Terms
| Term | Plain-English meaning | Why it matters |
|---|---|---|
| Use Case | A specific business problem or workflow where AI can create value. | Good AI adoption starts with clear use cases, not vague experimentation. |
| Workflow Automation | Using software, sometimes with AI, to complete repeatable steps with less manual effort. | Can reduce time spent on reporting, routing, drafting, analysis, and admin tasks. |
| Copilot | An AI assistant that helps a human perform work but does not fully own the process. | Common in writing, coding, finance, CRM, email, and productivity tools. |
| Autonomous System | An AI-enabled system that can operate with limited human intervention. | Higher leverage, but also higher risk and governance needs. |
| Model Provider | A company that builds or hosts AI models. | Examples include OpenAI, Anthropic, Google, Meta, xAI, and others. |
| API | A software connection that lets one system use another system’s capabilities. | Many businesses access AI models through APIs rather than standalone chat apps. |
| Latency | The time it takes for an AI system to respond. | Important for user experience, customer support, trading workflows, and real-time applications. |
| Model Cost | The cost to run AI tasks, often based on usage volume, input size, output size, and model choice. | Affects whether an AI use case can scale economically. |
| Model Drift | When an AI system’s performance changes or degrades over time as data, users, or conditions change. | Requires monitoring, evaluation, and periodic updates. |
06 — Interfaces & formats
Ways People and Systems Use AI
| Term | Plain-English meaning | Simple example |
|---|---|---|
| Chat App | A conversational interface where a person types or speaks to an AI assistant. | ChatGPT, Claude, Gemini, Grok, or Perplexity in a browser or mobile app. |
| Personal agent | An assistant that can keep working across your email, calendar, files, and messages, and check back when a decision is yours. | A ChatGPT dot, Grok Bot, Muse, or Claude Duet. |
| API (Application Programming Interface) | A software connection that lets one product use another product’s capabilities. | A company connects its customer-support tool to OpenAI or Anthropic so tickets can be summarized automatically. |
| CLI (Command-Line Interface) | A text-based tool used from a terminal instead of a graphical app. | A developer types a command to ask an AI coding agent to inspect a software project. |
| IDE | A coding workspace where developers write, run, and debug software. | VS Code, Cursor, JetBrains, or Xcode with an AI coding assistant built in. |
| SDK (Software Development Kit) | A packaged set of code helpers that makes it easier for developers to use an API. | Instead of manually calling an AI API, a developer installs the provider’s Python or JavaScript SDK. |
| MCP (Model Context Protocol) | An open standard for connecting AI apps to external tools, files, databases, and workflows. | Claude or ChatGPT connects to Google Drive, local files, or a company database through an MCP server. |
| MCP Client | The AI app or environment that connects to MCP servers. | Claude Desktop, ChatGPT, or a coding tool that can discover and use MCP tools. |
| MCP Server | The connector that exposes a tool, data source, prompt, or workflow to an AI app. | A Gmail MCP server might let an AI search email, draft replies, or label messages if authorized. |
| Connector | A prebuilt integration between an AI tool and another app or data source. | A Google Drive connector lets an AI read selected Docs, Sheets, Slides, or PDFs. |
| Plugin | An add-on that gives an AI tool new capabilities, workflows, or integrations. | A plugin might add a research workflow, a document generator, or a specialized finance analysis routine. |
| Skill | A reusable instruction package that teaches an AI how to perform a particular kind of work. | A “meeting notes formatter” skill could turn raw notes into decisions, action items, and follow-up emails. |
| Markdown / MD File | A plain-text document format that uses simple symbols for headings, lists, links, and tables. | This term sheet is an .md file, so it is easy to edit, share, convert to PDF, or publish on a website. |
| JSON | A structured data format often used by APIs and AI tools. | An AI can return customer records as JSON so another system can read them cleanly. |
| CSV | A simple spreadsheet-style file format where values are separated by commas. | A company exports sales data as a CSV and asks an AI to summarize trends. |
07 — Providers
Major AI Product Families
Product and model names change quickly. Use this as a map of the major families, with links to each provider’s current model or product pages instead of static model lists.
| Provider / family | Main products | Latest links | Best mental model |
|---|---|---|---|
| Anthropic Claude | Claude chat, with Cowork built in, Claude Code, Claude Duet, Claude in Chrome, and Claude for Microsoft 365. | Claude models, Claude products, Claude Duet. | Assistants and agents for writing, analysis, reasoning, coding, and long documents. Duet is the one that comes to you. Cowork is how a normal Claude chat takes on a bigger task. |
| OpenAI ChatGPT / Codex | ChatGPT, dots, ChatGPT Work, Codex, the API, and plugins. Image tools depend on the plan. Custom GPTs are being retired in favor of plugins. | OpenAI models, Codex and Work models, Dots. | ChatGPT is the general assistant. A dot keeps an ongoing responsibility. Work finishes a defined task. Codex builds software. The API puts the models in a company’s own products. |
| Google Gemini | Gemini app, Gemini Notebook, Gemini API, Google AI Studio, Gemini in Workspace, and Gemini Enterprise Agent Platform. | Gemini API models, Google AI Studio, Gemini Notebook. | Google’s AI family across chat, Workspace, Android, cloud, coding, and multimodal apps. Gemini Notebook is the research notebook formerly called NotebookLM. Saved custom instructions are skills. Gems are being retired. |
| Perplexity | Perplexity search, Computer, Comet, Deep Research, plus the Agent API, Search API, and Embeddings API. | Perplexity API, Agent API, Perplexity. | An answer engine built around live search and citations. Computer is where a hard question becomes a file, a deck, or a dashboard. |
| xAI Grok | Grok chat, Grok Bot, Grok Build, the API, Grok Imagine, and Grok Voice. | xAI models, xAI API, Grok Bot. | Chat with live web and X search, images, and video, plus Grok Bot for ongoing work and Grok Build for code. |
08 — Product types
Product Types: Chat, Code, Cowork, Search
An assistant that keeps working across your own accounts is a personal agent. That family is in the next section.
| Product type | What it does | Examples |
|---|---|---|
| Chat Assistant | Answers questions, drafts text, analyzes files, brainstorms, summarizes, and helps with everyday knowledge work. | ChatGPT, Claude, Gemini, Grok. |
| Coding Agent | Reads and edits code, runs tests, fixes bugs, builds features, and can work across a software project. | OpenAI Codex, Claude Code, Grok Build, Gemini Code Assist. |
| Cowork / Desktop Agent | Takes a task you describe and works across files and tools with more autonomy than a chat window. It can keep going after you close the laptop. | Claude Cowork, ChatGPT Work, Copilot Cowork, and Perplexity Computer. |
| Search / Answer Engine | Searches current web sources and synthesizes an answer with citations. | Perplexity, Gemini with search grounding, Grok with web and X search, ChatGPT search. |
| API Platform | Lets developers build AI into apps, workflows, websites, and internal systems. | OpenAI API, Anthropic API, Gemini API, Perplexity API, xAI API. |
| Multimodal Studio | Creates or edits images, video, audio, slides, or interactive media. | Gemini image and video tools, OpenAI image tools, Grok Imagine, Adobe/Canva AI features. |
09 — On your behalf
AI Personal Agents
A personal agent does work across the apps you already use, and it can keep going after you close the laptop. A chatbot waits for the next message. A coding agent is the specialist for software.
| Term | Plain-English meaning | How it is different |
|---|---|---|
| Chatbot | You ask a question or for a draft, and it answers in the conversation. When you stop, it stops. | Nothing keeps running after you leave, and it does not stay logged into your email on its own. |
| Personal agent | An assistant you hand real tasks to. It can use your email, calendar, files, messages, and the web, then come back when it needs a decision. | You set what it may do alone and what it must ask you first. It is meant to take work off your plate, not only to talk. |
| Coding agent | A specialist that reads a software project, edits code, and runs checks. | Built for code. A personal agent may call one, but a person still needs to review the change. |
Ideas these products share
| Idea | Plain-English meaning | Why it matters |
|---|---|---|
| Connector | A permissioned link to an app such as email, calendar, Slack, or a drive. | The agent can only see accounts you connect. A missing connection is a common reason a task fails. |
| Computer or browser use | The agent has its own computer or browser and can click through a website that has no connector. | Useful when the task lives on the web. A site can still block it, or stop and ask you to sign in. |
| Routine | A saved job that runs on a schedule, or when something happens, such as a new email. | This is how repeat work leaves your list. Tell it what to do if the source is missing, so it does not quietly reuse old information. |
| Memory | What the agent keeps about your preferences, people, and past work, so you do not have to re-explain them. | Helpful, and worth a look now and then. For a fact that matters, ask it to check the current source. |
| Team of agents | Several named assistants that split a job, message each other, or share one computer. | You can keep one for research and another for scheduling, instead of one general chat that tries to do everything. |
| Approvals | The rules for what the agent may do alone, what it must show you first, and what it must never do. | Sending, paying, signing, and deleting should stay with you unless you have clearly allowed that step. |
The main ones right now
| Product | What it is | Worth knowing |
|---|---|---|
| Grok Bot | Named bots from xAI that you message like teammates. They share a cloud computer with a browser, files, and a terminal, use connected apps where those exist, and keep working with your laptop closed. You can save a workflow as a skill and run it as a routine. Bots can hand work to each other. | Files and sign-ins on that computer are visible to every bot on your account. Someone else’s computer stays separate. A team can share one bot. Each person’s chat stays private, and an app login is that person’s own. Grok Bot. |
| ChatGPT dot | An always-on agent in ChatGPT, on GPT-6 Astra, with its own cloud computer. It starts from your ChatGPT memory, uses the apps you allow, and keeps working between conversations. It can pass a defined task to ChatGPT Work, or a coding task to Codex. | You can see what it is doing, pause it, and set rules for what needs approval. It is rolling out on eligible Pro, Business Premium, and Enterprise plans. Pro access is for adults and, for now, leaves out the UK, Switzerland, and the European Economic Area. You can also reach it in Slack or Teams. Dots. |
| Muse | Meta’s personal agent. You message it in the Muse app or in WhatsApp. It runs on its own cloud computer, can use connected apps and a browser, and keeps working after you close the app. It asks before it sends an email or spends money. | It is built for everyday tasks and longer goals, and it is not in every country yet. There is a free usage limit. Passwords and payment numbers are stored so Muse can use them without the agent seeing them. Introducing Muse. |
| Claude Duet | Anthropic’s proactive agent. It can message you with a morning brief from the inbox and calendar you connect, work on a cloud computer, and use your Claude memory, skills, and connectors. It waits when a step needs you. | This is a research preview, by waitlist, for Pro, Max, and Ultra plans in the United States, and it runs on Claude Opus 5.5. It asks before it sends, pays, signs, or deletes anything. It will not trade securities or move money between accounts, even if you approve. Claude Cowork is the related way to hand Claude a task you describe, including on a schedule. Claude Duet. |
| Perplexity Computer | Perplexity’s agent for a job that is bigger than one search. It can research with citations and turn the result into a file, a deck, or a dashboard. On the desktop app, Personal Computer can read and edit files on your machine. Comet is the browser. | Projects, the new name for Spaces, keep the files, the instructions, and a memory called Brain. Deep Research also runs inside Computer. Perplexity. |
| OpenClaw | A free, open-source assistant you run on your own Mac, Windows, or Linux computer. You talk to it in chat apps you already use, such as WhatsApp, Slack, iMessage, or Telegram, and you choose which model it uses. | There is no hosted subscription. Your files and memory stay on your machine. It can also run commands there, so use the stricter setup if people other than you can message it. The project is stewarded by the OpenClaw Foundation. OpenClaw. |
| Gemini Agent | An experimental tool in the Gemini app that can carry a multi-step task through connected apps and a browser, and asks before it sends email or pays. | It is for Google AI Ultra subscribers, age 18 and older, in the United States, on the web. It is an early step, not a widely available personal agent yet. Gemini release notes. |
Which one, when?
One published test, not a ranking to treat as fact. On October 29, 2026, Claire Vo ran Grok Bot, a ChatGPT dot, Muse, and Claude Duet on nine jobs she actually needed done. She did not find a universal winner. The connectors were not identical. Muse was missing some work accounts in that test.
| Job | Where she would start, in that test |
|---|---|
| Work coordination | Grok Bot. It did best on her inbox list, a working document, and comparison shopping. |
| Home and light creative work | Muse. It did best on a family paper, a fast explanation, and a short clip. |
| Coding | Codex or Claude, used directly. She would not hand the repository to the personal agent. |
Claire Vo, “Grok vs. Dot vs. Muse vs. Claude Duet: Nine Tasks, Four Agents, No Universal Winner,” How I AI, October 29, 2026. Video. She judged her own outputs. The write-up says the scores from that run are not a portable benchmark.
10 — Compare
Quick Comparison
| Concept | Best for | Watchouts |
|---|---|---|
| Chatbot | Conversational Q&A and simple task support. | May give unsupported answers if not grounded. |
| RAG System | Answering questions from trusted internal or external documents. | Depends heavily on document quality and retrieval accuracy. |
| Fine-Tuned Model | Repeated specialized tasks with consistent style or domain rules. | Needs high-quality training data and ongoing evaluation. |
| AI Agent | Multi-step workflows that involve tools, decisions, and follow-through. | Needs permissions, guardrails, logging, and human review for important actions. |
| Predictive Model | Forecasting and classification from structured data. | Can fail when the future differs from historical patterns. |
11 — Prompting
Useful Prompt Formula
Use this structure for better AI outputs.
- Role: Tell the AI what perspective to take.
- Goal: State the outcome you want.
- Context: Provide relevant background, documents, data, or constraints.
- Format: Specify the output shape, such as bullets, table, memo, email, or checklist.
- Quality bar: Tell it what to optimize for, such as accuracy, brevity, citations, or executive readability.
Act as a finance analyst. Summarize this earnings call transcript for an investment committee. Focus on revenue drivers, margin pressure, capital allocation, guidance changes, and management tone. Output a one-page memo with key takeaways, risks, and follow-up questions.
12 — Start here
Executive Takeaway
The most important AI terms to understand first are LLM, prompt, context window, RAG, agent, personal agent, MCP, API, CLI, hallucination, grounding, evaluation, and guardrails. Together, they explain how modern AI tools work, how they connect to real systems, where they are useful, and where they need oversight.
Provider and product examples were checked against public docs and product pages from OpenAI, Anthropic, Google, Microsoft, Perplexity, xAI, and Meta, and against the OpenClaw project site, on October 31, 2026. The “which one, when” note in Personal Agents is Claire Vo’s result from her October 29, 2026 How I AI episode, not a universal ranking. Model names and product packaging change frequently, so use the links above for the latest details.