Walk into a 2026 product review and someone will say "we should add AI." Six people will nod and each will mean something different — a chatbot, a recommender, an agent, a vision model, a generator. Until everyone in the room can name which kind of AI they're proposing, the conversation cannot progress. This lesson gives you the taxonomy so you can pin the word down whenever it floats in.
Conversational AI: chatbots and assistants
A chatbot is the most familiar shape of AI today: a text box, a thread of messages, a model that responds in natural language. ChatGPT, Claude, Gemini, and the AI features baked into Slack, Notion, and your favorite IDE all sit here. They're conversational, stateful within a session, and built around a single primary action — answering whatever you type.
The thing to notice is that "chatbot" describes the interface, not the capability. Two chatbots can look identical and do completely different jobs underneath — one is a thin wrapper over an LLM, another is a full retrieval system grounded in a company's internal documents, a third is a router that dispatches to specialized tools. When a vendor says "we have a chatbot," your job is to ask what's behind the text box.
Assistants are chatbots with a slightly broader remit — they may listen via voice (Siri, Alexa), live inside another product (GitHub Copilot Chat), or persist memory across sessions. The line between chatbot and assistant is fuzzy and not worth defending. They're the same family.
Predict, retrieve, and act
Recommendation systems
Recommendation AI is the quietest category and probably the most influential. It is what shows you the next TikTok, the next Spotify song, the next Amazon product, the next post in your feed. The user never types a query; the system simply predicts what they will want and serves it.
Recommenders predate the LLM era by twenty years — Netflix's collaborative-filtering paper is from 2007 — but they remain among the highest-leverage applications of machine learning. They live entirely inside other products, optimizing engagement or conversion or retention with no visible "AI" label on the surface.
For a designer or PM, the important fact is that recommendation systems are opinionated: every ranking encodes a choice about what to amplify and what to suppress. A neutral recommender does not exist. If your product uses one, you are making editorial decisions whether you mean to or not.
Search-and-summary (answer engines)
This is the category Perplexity helped popularize: you ask a question, the system fetches relevant pages from the web (or from your private documents), reads them, and synthesizes a cited answer. It's a chatbot on top of a search engine — or, depending on how you squint, a search engine that finally learned to talk.
Under the hood these are usually RAG systems — retrieval-augmented generation — which is so common it gets a whole chapter later (Chapter 7). The retrieval step grounds the model in real, specific text, which dramatically reduces the risk of made-up answers compared to a bare LLM. The price is that the answer is only as good as what got retrieved.
If you're imagining an AI feature where users will ask questions about a known body of knowledge — a help center, a policy library, a research archive — you are imagining an answer engine, even if no one in the room has used the term yet.
AI agents that take action
An agent is an AI that doesn't just talk, it does. It books the flight. It writes the file. It opens the browser, clicks the button, fills the form. Agents are the buzziest category in 2026 and also the most commonly oversold — most products marketed as "agents" are still very close to chatbots with a couple of tools attached.
The technical difference is small but important: an agent has access to tools (we'll cover these in Chapter 8) and runs in a loop, planning steps, executing them, observing results, and deciding what to do next. A chatbot replies once and waits. An agent can spend ten minutes working on your behalf before coming back with an answer or a completed task.
The design implications are large. With chatbots, the worst-case outcome of a bad response is a confused user. With agents, the worst case is a real-world side effect — money moved, an email sent, a file deleted. Every agent product needs guardrails, confirmation patterns, and undo paths that chatbots can do without.
Perception and creation
Vision AI
Vision AI processes images and video. Face ID unlocks your phone, your bank app reads a check from a photo, your phone groups pictures of your dog, a Tesla decides whether the thing in front of it is a stop sign or a billboard. None of these uses natural language, but all of them are AI in the modern sense — patterns learned from enormous training sets, applied to new inputs.
For the first decade of the deep-learning era (roughly 2012–2022), vision was the marquee application. Then language stole the spotlight. In 2026 the two are increasingly fused into multimodal models — a single model that can read text and look at an image and reason about both together. You can paste a screenshot into ChatGPT and ask why a chart looks wrong; that is vision AI and language AI working as one.
If your product idea involves cameras, photos, documents, screenshots, video, or anything visual, vision capabilities are now table-stakes and increasingly available via the same APIs you'd use for text.
Generative AI (text, image, audio, video)
Generative AI is the umbrella term for systems that produce new content rather than classify or rank existing content. LLMs generate text. Midjourney and DALL-E generate images. ElevenLabs generates speech. Suno generates music. Sora and Runway generate video. They are all generative.
The thing that distinguishes generative AI from older AI categories is that the output didn't previously exist. A recommender picks among real songs; a generator writes a new one. That changes the failure modes (the output can be wrong in ways no source could verify), the legal questions (whose style did it learn from?), and the product opportunities (you can compress entire creative workflows into one feature).
Almost every "AI feature" launched by a non-AI company in 2024–2026 was generative: write a draft, summarize this email, brainstorm ideas, make a thumbnail. The category is so dominant that "AI" and "generative AI" are increasingly used as synonyms, even though the broader field is much bigger.
Automation AI (workflow tools)
The last big category is automation — AI dropped into a workflow tool so the steps a human used to chain together can run themselves. Zapier and n8n have added AI nodes that let you say "when a new lead comes in, summarize their company, score them, and write a personalized intro email." That is three model calls and several tool calls, but to the user it's one rule.
Automation AI is where most non-engineering teams actually deploy AI in 2026. It doesn't require a custom product; it doesn't require new infrastructure; it requires a Zapier account and a credit card. The work is in designing the workflow well, which is often the same work as designing a small agent — just expressed in a no-code editor instead of code.
If you're trying to add AI to an existing business process and you don't have engineers, this is where to start. We'll come back to no-code AI in Chapter 6.
Perplexity is an answer engine: type a question, get a cited synthesis from current web sources. Midjourney is generative vision: type a prompt, get an image. Cursor is a coding agent: it doesn't just suggest, it edits files, runs tests, and pushes changes. TikTok is recommendation AI at planetary scale — the For You page is one of the most powerful predictors of human attention ever deployed.
Notice how each product foregrounds a single capability. Perplexity isn't trying to be a chatbot, Midjourney isn't trying to summarize the web, Cursor isn't trying to make videos. Clarity about which category you're in is half of good AI product design.
- Treating "chatbot" as a capability. The text box tells you the interface, not whether it's a thin LLM wrapper, a grounded retrieval system, or a tool-using router — always ask what's behind it.
- Calling something an agent when it's a chatbot with one tool bolted on. A real agent runs in a plan-act-observe loop; if it replies once and stops, it isn't acting autonomously.
- Assuming a recommender is neutral. Every ranking amplifies some things and suppresses others, so deploying one means making editorial choices whether you intend to or not.
- Forgetting that generative output never existed before, so there's no source to verify it against — a failure mode that search and recommendation simply don't have.
- Picking a category by which sounds most impressive rather than which fits the job. The buzziest label (usually "agent") often adds risk and complexity a simpler shape would have avoided.
Pick three AI products and classify each into one (or more) of the categories above · 10–20 minutes · no code required. Hard mode: pick one product that genuinely spans two categories (most ambitious products do) and write a sentence about what each half is doing. The point is to train your eye for the seams between categories — they're where most architecture decisions live.
Key Takeaways
- "AI" is an umbrella; under it sit at least seven distinct kinds of products with different strengths and risks.
- Chatbots describe the interface, not the underlying capability — always ask what's behind the text box.
- Recommenders are the quiet giants of consumer AI; almost every feed you use is one.
- Agents are chatbots that act. The new failure mode is real-world side effects, which changes the design problem.
- Generative AI is the dominant category in 2026; "AI" and "generative AI" are often used as synonyms.
- Naming the category you're proposing is the fastest way to make a fuzzy product conversation concrete.
Modern AI is not one thing but a family of at least seven product shapes — conversational, recommendation, answer engines, agents, vision, generative, and automation — each with its own strengths, failure modes, and design demands. The interface you see rarely tells you the capability underneath, so the most useful habit is to name the category explicitly before debating the feature. Get clear on which shape you're building, and half of good AI product design is already done.