Every vendor renamed their chatbot an agent in about eighteen months, so the word now means almost nothing. It is worth pinning down, because the difference is real and it decides whether the thing is useful to you.
The actual distinction
A chatbot completes a turn. You ask, it answers, the exchange ends. Whatever happens next is your job.
An agent completes a task. You give it an outcome, and it does the intermediate steps: looks things up, decides what to do next based on what it found, takes actions in real systems, and comes back with what it finished and what it could not.
The tell is what you are left holding. If you get a draft you still have to send, you used a chatbot with good output. If the thing got sent, logged, and followed up on, that was an agent.
Where it earns its keep
Agents are worth it where the job is multi-step, repetitive, and mostly mechanical.
Chasing. The single highest-return case in a small business, and the least glamorous. Somebody owes you a document, a signature, or an answer. The work is not hard, it is just relentless, and it slips because nobody wants to send the fourth reminder. An agent does not mind sending the fourth reminder.
Enrichment and research. Finding the right contact, pulling the company details, checking whether a lead is worth a call. Bounded work with a clear finish line.
Follow-through. Turning what got agreed into things that actually appear in somebody's week, then noticing when they have not moved.
Where it does not
Anything one-off is not worth the setup. Anything requiring genuine judgment about people should not be delegated. And anything where being wrong is expensive and hard to undo needs a human in front of it, permanently, not just during a trial period.
There is also a category people get wrong: work that looks repetitive but is not. Pricing, hiring calls, and customer escalations all have a shape that repeats and a substance that never does. An agent will produce a confident, consistent answer to each one, and consistency is the wrong goal there.
The part vendors skip
An agent that can act can act wrongly. Three things separate a system you can run in a real business from a demo.
Approval gates. Reversible work runs unattended. Anything that leaves your building, spends money, or changes a customer record asks first. This is a design decision, not a setting, and you should be able to see exactly where the line sits.
An audit log. Every action, with a timestamp and an actor, that you can read after the fact. When something goes wrong at 2am on a Saturday, the log is the difference between a five-minute fix and a bad week.
Reversibility. You need to be able to undo it. If the answer to "what happens when it gets this wrong" is a shrug, the product is not finished.
Ask any vendor those three questions. The answers separate the serious ones quickly.
How we think about it
Every [Goal Boss app](/apps) is built as an agent rather than a feature you have to drive. DocChaser keeps emailing the client who has not sent the form. Pipeline tells you which deal has gone quiet. Notes files the action items before you are back at your desk.
Actions are logged and reversible, and the ones that touch the outside world ask you first. That is not caution for its own sake. It is the only way a small team can hand real work to software without adding a new category of thing to worry about.
The goal is not fewer people. It is that the tenth follow-up email gets sent, which it currently does not.
