What “AI agents” actually means in your business.
When someone says they want AI agents, the word is doing five different jobs at once. Here’s what those five are, and which ones fit your business.
A CEO I talked to recently told me he wants ten agents in his business. When I asked him what an agent was, he said, “you know, like AI doing things.” He's not wrong about what he wants. He just hasn't been told the word is doing five different jobs at once right now, and the five things don't all cost the same or do the same work.
This isn't a stupid question. Most people I talk to are confused about it, including people who use AI every day. So let's go through what people actually mean when they say it.
The short version
When someone says “AI agents,” the word is doing five different jobs at once right now, and that's the source of most of the confusion. The five things range from a chat assistant you type into (ChatGPT, Claude), to a tool with AI baked in (Granola transcribing meetings, Notion summarizing docs), to a script that does one task on a schedule (overnight grant scans, daily outbound), to a coding agent on your computer (Claude Code editing files while you watch), to an autonomous agent that takes a goal and figures out the steps on its own. They all count as “AI doing things,” which is what most buyers actually want. They cost different amounts, take different work to set up, and fit different parts of your business. When a vendor pitches you “agents,” your first question is which of the five they mean, because that determines what you're actually buying.
The five flavors people actually mean
- Chat assistants.ChatGPT, Claude, Gemini in their chat windows. You type, they respond. You drive every turn. The most common entry point by far. When someone says “I use AI,” this is almost always what they mean. Cost is around $20 per month per person. Setup is opening the website and logging in.
- AI features baked into tools you already use. Granola transcribes your meetings and drafts the notes. Notion summarizes pages and rewrites text. Cursor autocompletes code. Grammarly fixes your writing. These aren't standalone agents, they're features. The AI is doing work, but inside the boundary of a specific product. You don't have to set them up beyond turning them on. Most useful AI in a small business right now is this category, and most people don't realize that's what they're using.
- Automated workflows that run on a schedule. A script that scans grant databases every night and emails you matches. A workflow that pulls leads from a public source and drafts outbound in your voice. A pipeline that watches your inbox for new vendor invoices and routes them to the bookkeeper. These run without you. They do one job. You set them up once, then check the output. When someone says “AI agents” and they mean a business automation, this is the category. Tools that build them: n8n, Zapier, Make, custom Python or Node scripts.
- Coding agents that operate on your computer. Claude Code, Cursor in agent mode, Devin. They edit files, run commands, make commits, debug their own work. You give them a goal, they execute, you intervene when needed. This is what people mean when they say “agentic coding.” Useful if you're writing software. For a non-tech business owner, this category usually shows up through the developer you hire, not as a tool you install yourself.
- Autonomous agents that take a goal and figure out the steps. A research agent told “find me three suppliers of recycled aluminum in the EU” goes off, queries external sources, calls APIs, comes back with a list. A customer service agent told “answer support emails about returns” reads inbound mail and writes replies. These take a high-level goal, plan steps, execute them, and recover when something breaks, without checking back at every step. Most powerful, most expensive to build right, hardest to trust. This is what gets pitched in startup land when someone says “AI agents.”
What “agent” isn't (yet)
- It isn't always autonomous.A chat assistant isn't doing anything until you type. Most of the “agents” in your business will need you to start them, check their output, or fix them when they get stuck.
- It isn't always one agent per job.A single AI engine (like Claude) can do many different jobs depending on how you instruct it. So when someone says “I want ten agents,” they might actually mean “I want Claude doing ten different jobs.” Whether you build ten separate workflows or one orchestrator running ten tasks depends on the work, not on the word.
- It isn't general AGI yet. No agent right now will fully run your business. Each one does specific work. The marketing makes them sound more capable than they are.
- It isn't a replacement for the person who owns the decision. Even the most autonomous agent should send the work to a human for the calls that need a human. If you build agents that make irreversible decisions without human review, you'll regret it the first time one of them is wrong.
How to know which kind you need
This isn't a hire decision, it's a what-to-build decision. The readiness signals look different from the last note.
- For a one-off question or a creative draft, use a chat assistant. Open ChatGPT or Claude. Ask the question. You don't need to build anything. Most small business AI use starts and stays here, which is fine.
- For ongoing work inside a tool you already use, turn on the AI feature. Granola for meetings. Notion AI for docs. Most of these are included in subscriptions you're already paying for. The setup is a checkbox.
- For repetitive work that needs to happen on a schedule, build an automated workflow. Overnight scans. Daily outbound. Weekly reports. This is where most small business “agent” value lives right now, even if vendors call it something else.
- For software you're shipping, hire someone who uses a coding agent. You don't need to install Claude Code yourself. You need a developer who already does.
- For autonomous multi-step work, scope it before you build it. The autonomous category is real but expensive. Most things pitched as autonomous agents would be cheaper and more reliable as a scheduled workflow with a chat assistant in the middle.
Why the word is so confusing right now
The word “agent” got hot in AI marketing in the last twelve months, and every vendor wanted to ride the wave. So now ChatGPT is an “agent,” Granola is an “agent, ” n8n calls itself an “agent platform,” Claude Code is an “agentic coding tool,” and an autonomous research workflow is an “agentic system.” All of them are using the same word for very different things.
It's the same thing that happened with “cloud” in 2010 and “AI” itself five years ago. The word grew faster than the categories underneath it. Eventually the categories settle. For now, the only way to know what someone means when they say “agents” is to ask: chat, baked-in feature, scheduled workflow, coding tool, or autonomous? Asking that is the question. Don't apologize for asking it.
Here's what I do
When a client says they want agents in their business, I start by asking the question above: which of the five do you mean? Most of the time, what they actually need is some combination of two and three: turn on the AI features in the tools they're already using, and build a few scheduled workflows that absorb the repetitive work. The autonomous category gets the most marketing oxygen but usually isn't what a small business should build first.
The Scoping Engagement is the version of this question shaped for your operation. Two weeks of looking, then a written plan with three projects worth building, each with an investment range and a timeline. If you want to know which of the five your business actually needs, start a conversation.