Ask ten people what an "AI agent" is and you'll get ten different answers. Some mean a chatbot. Some mean a piece of software that can use other software on its own. Some are just relabeling automation they built two years ago. The word has been stretched over so many different products that it's stopped meaning anything specific, which makes it hard to know if you actually need one, or if you're being sold a description instead of a tool.
Here's the plain version. An agent is software built around an AI model that's given a goal, not just a question. Instead of you asking it something and it answering once, it decides on its own what steps to take to get there, checking a calendar, sending an email, pulling data from three different places, and it adjusts as it goes if something doesn't go to plan. That loop, decide, act, check, adjust, is the actual difference between an agent and a regular automation or a chatbot that just answers questions.
A simple automation runs the same fixed steps every time, in the same order, and breaks the moment something doesn't match the pattern it was built for. An agent is built to handle the version of the task that doesn't go according to plan, which is also why it costs more, takes more to build correctly, and isn't the right tool for every job.
The kinds that actually show up in a business like yours
Most of what gets called an "agent" this year falls into a handful of categories. Here are the ones that actually matter if you're not a software company.
Workflow agents
The agent reads an incoming email, a form submission, or a document, decides what it's asking for, and takes the next step: updates your CRM, schedules a follow-up, drafts a reply for you to approve. This is the category most relevant to an owner-operated business. It's built to handle the parts of your day that involve judgment, not just repetition.
Customer-facing agents
A chatbot or voice agent that can actually hold a conversation, answer a real question about your business, and book or qualify a lead, instead of just routing people to a contact form. The difference from an old-school chatbot is that it isn't reading from a fixed script.
Research and monitoring agents
The agent pulls information from multiple sources on its own, on a schedule, and surfaces what changed or what needs attention. A dashboard that checks your Google Business Profile, your reviews, and your rankings every week and flags what moved is this category.
Coding agents
These write and edit software. Unless you're building your own product, this category isn't built for you directly, but it's often what's building the other three.
Do you actually need one?
Before the definition, before the category, there's a question worth asking first: does this problem need judgment, or does it need consistency?
If a process is well-defined, same steps, same order, every time, a plain automation does it for a fraction of the cost and it doesn't need supervision. Zapier, Make, a script someone writes once. That's not a lesser version of an agent. For a fixed process, it's the better tool, because it's predictable and it's cheap.
An agent earns its cost when the input varies. Emails that don't follow a template. Requests that need a judgment call. A volume of exceptions that would otherwise land on you or someone on your team every day. That's the actual test: how much of this task is "always do the same thing" versus "figure out what to do."
The cost question is worth being blunt about. Agents run on API calls; you're billed per action or per token, not a flat monthly fee. For a low-volume task, that can be close to nothing. For a high-volume one, it adds up fast, and it needs to be watched, not set and forgotten. A business owner deciding between a $20-a-month tool and an agent that could cost more some months and less others is making a real tradeoff, not just a technology choice.
The mistake isn't buying an agent. It's buying one before you've mapped the workflow it's supposed to replace. Map the steps, count how many of them require judgment, and the answer to "do I need an agent" usually answers itself.
This is the first conversation I have with any client before I recommend building anything: what does the workflow actually look like today, where does it break, and does fixing it need judgment or just consistency. Half the time, the answer is a simpler tool than an agent. The other half, it's exactly where an agent belongs.