Below is each term in plain language, what it's genuinely good at, and where it falls over.
What is workflow automation?
Workflow automation runs a chain of steps you defined ahead of time: when this happens, do that. A form submission creates a record. A new deal in the CRM sends a Slack message. A file lands in a folder and gets copied somewhere else. Zapier, Make, and the built-in automations inside your CRM all live here.
Good at: anything predictable, structured, and identical every time. It's fast to set up, cheap to run, and easy to reason about. If a task is truly the same every time, this is the answer, and reaching for anything fancier is a waste of money.
Where it breaks: the moment the input varies. A rule that expects a phone number in a form field has nothing to say about a phone number buried in the third sentence of an email. Traditional automation cannot read, weigh, or interpret. It matches.
What is a chatbot?
A chatbot is a conversation surface. Someone types a question, it answers. Older ones followed scripted decision trees; newer ones use a language model and can answer from your documents and policies.
Good at: answering the same questions over and over — hours, service area, policy details, how-do-I questions. A good one, pointed at your real policies rather than generic text, deflects a meaningful share of routine messages.
Where it breaks: most chatbots answer but don't do. Asked to actually reschedule the appointment, cancel the order, or issue the credit, a plain chatbot hands the person back to a queue — which is exactly the moment the customer wanted something to happen. A chatbot that can take action is really an agent wearing a chat window.
What is an AI agent?
An AI agent reads unstructured input — an email, a voicemail transcript, a photo of a delivery note — works out what it means, decides what to do based on rules you set, and then does it using the tools it's been connected to. It can look up a record, draft a reply, create the order, and stop to ask a person when something is unclear.
Three things separate an agent from the two above:
- It handles variation. Fifty customers describing the same problem fifty different ways is a normal Tuesday, not an error condition.
- It uses tools. It can query your CRM, write to your accounting system, check the calendar. Reading and writing, not just chatting.
- It makes small judgments inside a boundary you draw. Which category this is, which person should handle it, whether this is routine or needs escalating.
Good at: the work that sits between two systems and is currently done by a person copying, sorting, and deciding. Inbox triage. Order entry from email. Drafting quotes from a request. Chasing invoices with the right tone for each customer. Three of those four top the seven workflows most worth automating first.
Where it breaks: anywhere the rules genuinely can't be written down. If the correct action depends on knowing that this particular customer's brother-in-law is on your board, no system is going to work that out. Agents also need guardrails that plain automations don't — approvals, escalation rules, logs — because a system that can act can act wrongly. We've written separately about the guardrails that make AI safe to run in your inbox.
What is RPA?
Robotic process automation drives software the way a person would: moving a cursor, clicking buttons, typing into fields on a screen. It exists because a lot of important business software has no other way in.
Good at: getting data in and out of old systems that offer no API and aren't going anywhere.
Where it breaks: it's brittle by design. A vendor moves a button in an update and the robot fails. In a small business it's usually a last resort, worth it only when a system you can't replace is holding real hours hostage.
AI agent vs. chatbot vs. automation vs. RPA, side by side
| Category | What it does | Good at | Where it breaks |
|---|---|---|---|
| Workflow automation | Runs a chain of steps you defined ahead of time | Predictable, structured work that is identical every time | The moment the input varies — it matches, it cannot read |
| Chatbot | Holds a conversation and answers questions | The same questions over and over, answered from your real policies | It answers but does not act, so the customer still waits in a queue |
| AI agent | Reads messy input, decides what to do, and acts in your systems | Work between two systems that a person does by hand today | Rules that genuinely cannot be written down; it needs guardrails |
| RPA | Drives other software by clicking through its screens | Old systems with no API that you cannot replace | Brittle by design — a vendor moves a button and it fails |
Which one does my business need?
Work through it in this order. The first "yes" is usually your answer.
| Question | If yes |
|---|---|
| Is the input structured and the rule identical every time? | Workflow automation |
| Do people mostly need answers, not actions? | Chatbot, pointed at your real policies |
| Does something have to read messy input, decide, and then act in your systems? | AI agent |
| Is the only way into a critical system through its screen? | RPA, reluctantly |
Two things worth saying plainly about that table.
Most real workflows use more than one. A well-built system might use a rule to catch the message, an agent to read and route it, and a plain automation to file the result. Nobody has to pick a side; the labels matter far less than what the thing actually does on a Tuesday.
Start at the top of the table, not the bottom. The most expensive mistake in this whole category is building an agent for work a five-dollar-a-month automation already handled. If your input is a web form with six fixed fields, you do not have an AI problem.
Do AI agents need more supervision than automations?
Yes, and this is the part sales decks skip.
A rule-based automation does exactly what you told it, forever. When it's wrong, it's wrong the same way every time, which makes it easy to spot and fix.
An agent handles variation, which is the whole point — and the same flexibility means it can be confidently wrong in a new way. That's not a reason to avoid agents. It's the reason well-built ones queue drafts for approval before anything reaches a customer, escalate rather than guess when input is unclear, and log every action with the reasoning attached. Supervision is a design feature, not an admission of weakness.
The practical version: the more the system can do, the more the build is about boundaries rather than capability.
Does the label change what you should buy?
Not much. Vendors relabel products constantly — the chatbot from two years ago is an "agentic platform" now, with the same feature list. Ignore the noun and ask four questions:
- What does it read, and in what format?
- What can it change in my systems, and what can't it touch?
- When it's unsure, what does it do?
- What can I see afterward about what it did and why?
Any product worth buying answers those in a sentence each. If the answers arrive as adjectives, you're being sold a category, not a system. Those four questions are also the ones a diagnosis answers about your own work — see what a fixed-fee AI workflow audit actually looks like.
That question — buy a product or build something fitted to your process — deserves its own answer, and we've written an honest build-versus-buy framework for custom AI agents versus off-the-shelf AI tools that covers when off-the-shelf wins outright.