Every AI product now calls itself something. A chatbot, an assistant, an automation platform, an agent, an AI workforce. The words get used loosely, sometimes on purpose, and owners end up comparing quotes for things that do very different jobs.
The differences are simple once you see them. Here's how I explain them to clients before we talk about any tool.
Automation: fixed rules, no judgment
Automation follows steps you wrote in advance. When an invoice email arrives, save the attachment to this folder. When a form is submitted, add a row to this spreadsheet and send this reply. When a job is marked complete, create the invoice.
It's been around for decades, and tools like Zapier and Power Automate made it easy to set up without a programmer. It's fast, cheap and predictable. It does exactly the same thing every time.
That's also its limit. Automation can't read a messy email and work out what the customer wants. If the supplier changes their invoice layout, or a customer writes "can we move Tuesday to the week after", a rule doesn't know what to do. It either breaks or does the wrong thing quietly.
Good for: moving information between systems, sending routine notices, anything where the input is always in the same shape.
Chatbot: answers in conversation, doesn't act
A chatbot talks. Older ones followed a script with buttons ("Press 1 for hours"). Newer ones use an AI model and can answer questions in plain language, drawing on your website, your price list or your policies.
A good chatbot can handle a lot of routine questions well: hours, service areas, what's included, how to book. What it usually can't do is take action in your systems. It can tell a customer how to reschedule. It can't reschedule them.
The risk with AI chatbots is confidence. They'll answer questions you never gave them information about, and some of those answers will be wrong. I wrote about where that goes badly in The Taco Bell lesson.
Good for: answering common questions on your website, after hours, with a clear handoff to a person when it doesn't know.
Agent: reads, decides and acts, within limits
An agent is an AI model that can use tools. It can read an email, look up the customer in your system, check the calendar, decide what to do, and do it: book the slot, draft the reply, flag the exception. It works through a task in steps, checking as it goes, rather than following a fixed script.
That makes agents useful for the work that used to need a person because the inputs were messy: invoices in fifty different layouts, booking requests written in plain English, job notes that need turning into a quote.
It also makes them the option that needs the most care. An agent that can act can act wrongly. The ones worth trusting have limits written in code (it can't issue refunds, it can't book outside opening hours), a check on anything that matters, and a person who reviews exceptions. My invoice system, for example, has three independent reads that must agree to the cent before anything moves. Details in Why we read every invoice three times.
Good for: repeatable work with clear right answers but messy inputs, where a person reviews the edge cases.
A quick way to tell them apart
Ask one question of any product: what happens when the input is unusual?
- Automation breaks or does the wrong thing.
- A chatbot gives an answer, possibly a wrong one.
- A well-built agent recognises it's unsure and hands the case to a person.
If the vendor can't answer that clearly, you know which one you're looking at, whatever it's called.
Most real setups use all three
In practice the best setups mix them. Take a typical missed-call flow for a trades company:
- Agent: answers the call, works out what the caller needs, checks the schedule and books the visit.
- Automation: sends the confirmation text, adds the job to the calendar, creates the customer record.
- Person: reviews the day's bookings and calls back anyone the agent flagged.
Each piece does what it's good at. The automation doesn't need to be smart. The agent doesn't need to send texts itself. The person spends their time on the three calls that needed them, not the thirty that didn't.
Questions to ask before you buy
When a vendor shows you a product, these cut through the labels:
- What can it actually do in my systems? Read only, or can it change things? Which things?
- What does it do when it's unsure? Ask for an example of a handoff.
- Where does it get its information? Your documents, the open internet, or both?
- Who checks its work, and how? "It's very accurate" is not a check.
- What does it cost per task, not per month? That's the number you can compare with staff time.
A product that answers all five plainly is worth a trial, whatever it calls itself. If you'd like a second opinion on a quote you're looking at, send it to me and I'll tell you what it is and whether it fits the job.