What Is an AI Agent? A Plain-English Definition

An AI agent is software that you hand a goal, not a script. It works out the steps, uses tools to carry them out (your inbox, your CRM, a browser, a spreadsheet), checks what happened, and keeps going until the job is done or it hits a decision that needs you. The simplest way to hold it in your head: a chatbot answers, an agent does.

That definition is short on purpose. Everything else in the agent vocabulary, from MCP to orchestration to RAG, is plumbing that supports those four moves: plan, act, check, repeat. If you can see those four moves in a product demo, you are looking at an agent. If you cannot, you are looking at something else wearing the word.

The four parts every real agent has

Strip away the vendor decks and an agent is a small system with four pieces. You already run the human version of each one.

1. A model. The large language model (from Anthropic, OpenAI, Google, or others) that does the reasoning. Think of it as the rented brain. Nearly every agent on the market runs on one of a handful of these, so "our proprietary AI" usually means proprietary wrapping, not a proprietary brain.
2. Tools. The connections that let the model touch the real world: read a file, send an email, query your database, book a meeting. A model with no tools can only talk. A model with tools can work. This is where most of the business value lives and where most vendor demos get vague.
3. Memory. What the agent carries from one run to the next. Without it, every morning is day one. With it, the agent learns your voice, your clients, your definitions, and gets more useful the longer it runs.
4. The loop. The control logic that says: here is the goal, here is what I have done, here is what happened, here is what I do next. The loop is what separates an agent from a one-shot prompt. It is also where guardrails live.

If you want the layered version of this, with skills and orchestration added, read how the agentic stack fits together. For the purpose of this definition, four parts is enough.

A worked example you can picture

Say you run a $12M services firm and a lead fills out your website form at 9:40 on a Tuesday night. Here is the same event handled three ways.

No automation. Someone sees it in the morning, looks up the company, drafts a reply, and books a call. Thirty minutes of a salaried person's time, twelve hours late.

Traditional automation. A rule fires: new form entry, send template email, create CRM record. Fast, cheap, and dumb. If the lead wrote "we already use your competitor, just comparing" in the notes field, the template goes out anyway and the rep never sees the note.

An agent. The goal is "qualify and respond to inbound leads." The agent reads the form, including the notes. It looks up the company and the person. It checks your CRM for prior contact. It drafts a reply that references the competitor comparison, scores the lead against the criteria you gave it, and either sends the reply (if you have allowed that) or queues it for a human with a one-paragraph summary of why it scored the way it did. In the morning, your rep has a decision to make, not a research project to start.

Notice what made the third version work. Not a smarter model. The agent had tools (web, CRM), memory (your scoring criteria, prior contacts), and a loop that let it decide what to do with an input nobody wrote a rule for.

What an AI agent is not

The word gets stretched to cover a lot of software that does not earn it. Three common impostors:

• A chatbot with a new label. If the product's whole job is to answer questions in a chat window, and a human then goes and does something with the answer, it is not an agent. It may be useful. It is not an agent.
• A workflow with one AI step. A fixed sequence (trigger, step, step, step) where one box says "summarize with AI" is automation that borrowed a language model. The path is still drawn by hand. The moment reality leaves the path, it fails exactly like the old automation did.
• A copilot. Software that sits beside a person and suggests, drafts, or autocompletes. The human is still the unit of work. Powerful, but a different category, and the autonomous agent vs copilot question deserves its own article because the buying decision is different.

None of these are bad. The problem is paying agent prices, or setting agent expectations, for a product that can only do one of them. The longer comparison, with the questions that expose which one a vendor is actually selling, is in agent vs chatbot vs copilot.

Why this matters to a business owner specifically

You are not going to build agents. You are going to buy them, hire people who build them, or direct an agent someone set up for you. In all three cases, the definition is your leverage.

Generative tools (write me an email, summarize this doc) make an existing person faster. Capacity still scales with headcount. An agent takes a whole job: the inbound triage, the weekly report, the reconciliation that nobody wants to do. Capacity scales with how clearly you can define the work. That is the operator-to-architect shift the glossary keeps coming back to. You stop being the one who does it and start being the one who specifies it.

So when you evaluate anything with "agent" on the label, run the four-part check. Which model is under the hood? What tools can it actually use in my systems, today, not on the roadmap? What does it remember between runs? And what happens inside the loop when a step fails? A vendor who can answer all four plainly is selling you an agent. A vendor who pivots to vision slides is selling you something else.

The one-sentence version

An AI agent is a system that takes a goal, uses tools to work toward it in a loop, and knows when to stop and ask. Everything else is detail.

Hold onto that sentence and the next twenty terms you hear on a sales call will sort themselves into "supports that" or "distracts from that." That is what fluency is for.

FAQ

Is ChatGPT an AI agent?

The plain chat window is not. You ask, it answers, the turn ends, and you carry the answer somewhere. When the same product browses the web, runs code, or completes several steps before it reports back, it is behaving like an agent in that moment. Most products now mix the two modes, which is why the label on the box tells you less than watching what it actually does.

Does an AI agent need to be fully autonomous to count?

No. Agent describes the loop: plan, act with tools, check, repeat. It says nothing about how many of those actions happen without a person approving them. A well-run agent stops and asks before the steps you cannot undo, like sending money, emailing a customer, or deleting records. Autonomy is a dial you set, not a requirement of the definition.

What does an AI agent actually need to work in my business?

Four things: a model to reason with, tools that connect it to your real systems, memory so it does not start from zero each time, and a clear definition of done for each job. Most disappointing agent deployments are missing the second or the fourth. The model is rarely the problem.

How is an AI agent different from automation I already have?

Traditional automation follows a fixed path you drew in advance: if this, then that. It breaks the moment reality steps off the path. An agent is handed the goal rather than the path, so it can handle the inputs you did not anticipate and decide what to do next. The trade is flexibility for predictability, which is why you bound an agent with permissions instead of trusting it blindly.