Episode 2

The three kinds of AI: chatbot, assistant, agent

This is the companion guide to the video. It has the comparison, a way to pick the right kind for a task, the which-kind prompt, and a list of current examples that gets updated as the tools change. If you have not watched yet, start there.

Episode 2Watch on YouTube

The one idea

The three kinds differ in one thing. That thing is how much of the work you hand over, which is also how much you have to trust the tool before you use it.

Think of it like working with a capable temp. You can text the temp a question. You can sit the temp next to you and work through something together. Or you can hand the temp a whole project and check the work on Friday. Same person, three levels of handover.

Most managers only ever need the first two, and the chatbot alone covers a surprising amount of the value in the first year.

The three kinds, compared

The chatbot

You ask, it answers.

What it does: You type a question or a request and it responds. It can be one exchange or a long back and forth, but every turn is you asking and it answering, and then you take the output and do something with it yourself. It does not see your files or your email unless you paste them in.
Use it for: A quick explanation, a definition, a first draft of a paragraph, rewording something clumsy, a list of options, a sanity check on your own thinking. Low stakes, self-contained, and you can judge the answer on sight.
What it needs from you: The question, plus enough pasted-in context for that one exchange.
The limit: Context. It only knows what you paste in and what it picked up in training. Each new session usually starts cold.

The assistant

It works alongside you.

What it does: It works in the same space you do. It can see the document you are both looking at, the files you upload, and sometimes your email and calendar if you connect them. It does several steps in a row while you supervise and correct as the work goes.
Use it for: Anything where the AI needs your real context to be useful: your meeting transcript, your numbers, your team members’ names, your rough draft. This is the right kind for your one weekly task.
What it needs from you: Context and guardrails every time, and your review of the result. You are still the person accountable for what goes out.
The limit: It will confidently do the wrong thing if you brief it badly. It is a supervised tool, not an autonomous one.

The agent

You hand off the whole job.

What it does: You describe a whole job and it runs on its own. It takes many steps, decides what to do next, uses other tools along the way, and comes back when the job is done or when it is stuck. You are not watching each step as it happens.
Use it for: Two things. A task you have done by hand enough times to specify it exactly and can check cheaply: research, monitoring something on a schedule, working through a batch of similar items. Or a throwaway prototype: describe something you could never build yourself and let a coding or design agent rough it out for a meeting.
What it needs from you: For real recurring work: tight guardrails, a clear definition of done, and a standard you can check the result against. For a prototype: a clear description, and the understanding that it is a prop for a conversation, not something you ship.
The limit: You are giving up direct supervision. It fits only when you can judge the result without being able to judge the method. Most managers are not there yet for real work, and that is completely fine.

How to pick

Match the kind to two questions: how much does the task need your context, and how much does a wrong answer cost.

A quick question, no real stakes, an answer you can check yourself. Use the chatbot.
A recurring task that needs your context and produces something you review. Use the assistant. This is your one weekly task.
A job you have done twenty times, can specify exactly, and can verify cheaply. Consider an agent as the reliable delegate, but not yet.
Something you could never build yourself, where the result is a prop for a meeting. A demo page, a mock, a quick prototype. A coding or design agent can do this now, and you judge it by looking at it.

Rule of thumb: start one level lower than you think you need for real recurring work. Move a task up only when the level below is working and has become boring. The prototype is the exception, because nothing depends on the first version.

The which-kind prompt

Describe a task and have the AI tell you which of the three kinds fits and what each version needs from you. Copy this, replace the bracketed line, and paste it into whatever AI tool you use.

There are three ways I can work with AI: as a chatbot (I ask, it answers, I use the output), as an assistant (it works with my files and context while I supervise), or as an agent (I hand off a whole job and it runs on its own).

Here is a task I do regularly:
[Describe the task in two or three sentences, including what a finished result looks like.]

Tell me three things:
1. Which of the three kinds fits this task best right now, and why.
2. What I would need to prepare for that version to work well.
3. What would have to be true before it is safe to move this task up to the next kind.

Current examples

Updated September 2026

The names move around. The three kinds do not. This list gets kept current so the video does not have to. For the full company-by-company breakdown, see the AI Tools & Platforms Tracker, updated as the landscape changes.

Chatbot

Claude, ChatGPT, and Gemini in their standard chat interface, on the web or in the app. The free tier or a plan around $20 a month.

Assistant

The Claude desktop app working with files on your computer. ChatGPT with attachments, Projects, or a custom GPT. Gemini in the side panel of Google Docs and Gmail. Microsoft Copilot inside Word, Excel, and Outlook.

Agent

Claude’s agent and computer-use features, ChatGPT’s agent mode, Google’s Antigravity, and custom or in-house agents built on top of these. Note that “agent” is also the current marketing word, so read it with some suspicion. Plenty of things labeled agents are assistants with a nicer wrapper.

The coding branch, for engineers

Claude Code and Google’s Antigravity are assistants and agents built for writing software. “Vibe coding” means describing what you want in plain English and letting the AI write the code. You do not need any of this unless you write code, with one exception below.

The one agent move worth trying now

The safe early use of an agent is the throwaway prototype. Hand a coding agent a plain-English description of a small demo, a web page that shows an animation or an interaction you want to pitch, or hand a design agent a page of instructions and get back a rough mock. You judge it by looking at it. It is a prop for a conversation, not something you ship. Nothing depends on the first version, which is exactly why it is safe to try.

The caution: the demo page does not go live and the mock does not reach a client as finished work until a professional has taken it the rest of the way.

Do not have an AI tool yet?

Any of Claude, ChatGPT, or Gemini works, and the picking above works on paper even if you do not. The Episode 1 guide has the short version of which to pick and when the paid plan is worth it.

Want a straight answer on which kinds your team should be using?

We assess where AI actually fits in your organization, design the systems, and train your people to run them.

Start a Conversation