A chatbot answers questions. Zapier (or Make, or n8n) moves data between apps when a trigger fires. An AI agent reads a situation, decides which of several actions to take, calls the tools to take them, and checks the result. If your workflow is "answer the same questions", buy a chatbot. If it is "when X happens, copy Y to Z", build a Zap. If it is "look at this, work out what to do, do it, and know when to ask a person", that is an agent, and the other two will break on the first message that does not fit the template.
The three get sold with the same words, which is why so many businesses buy the wrong one. This page puts them side by side on the questions that matter: what each can decide, what it costs, what breaks it, and when you should not use any of them.
| Chatbot | Zapier / Make / n8n | AI agent | |
|---|---|---|---|
| What it does | Answers questions from a script or a knowledge base | Moves data between apps when a trigger fires | Reads a situation, decides, acts through tools, checks the result |
| Handles an unexpected message | Falls back to "I didn’t understand" | Never sees it; a Zap has no conversation | Classifies it, acts or escalates with context |
| Multi-step tasks | No | Yes, if every step is fixed in advance | Yes, including deciding the steps |
| Judgment calls | None | None (filters and conditions only) | Within rules you write; approval step for the rest |
| Writes to your systems | Rarely; mostly links out | Yes, mechanically | Yes, through named tools with permissions |
| Typical setup | Hours to days | Hours to a week | Two to three weeks for a pilot; six to twelve for a full workflow |
| Typical cost | $0–$300 a month SaaS | $20–$600 a month SaaS plus setup time | $4,500 pilot; $8,000–$42,000 for one workflow in production |
| Running cost | Subscription | Subscription, per task | Model usage, roughly $30–$600 a month at SMB volume |
| What breaks it | A question off the script | A field renamed in either app; a case the branches did not foresee | Rules nobody wrote down; data it cannot reach |
| Best for | FAQs, hours, order status | Form → CRM → Slack plumbing | Booking, qualifying, chasing, reconciling, triaging |
When the job is answering the same forty questions and sending people to the right page. It is cheap, it is quick, and it is the right tool more often than vendors of agents admit.
When the workflow is fully specified and every branch is known. Automations are the plumbing most businesses are missing, and an agent on top of missing plumbing is a waste.
When the work needs reading and deciding before acting, and the cases vary enough that a template breaks. Agents are also what you build when a chatbot has failed because people kept asking it to do things.
A chatbot decides nothing: it matches a question to an answer. That is a feature when the answers are fixed, and the whole problem the moment someone writes "can I move Thursday to Friday and add my son?"
Zapier decides with conditions: if the form says "Enterprise", route to Priya; else, to the queue. It is fast and reliable as long as reality fits the branches. It cannot read a bank statement and tell you which page is missing, and it cannot notice that two of its own branches contradict each other.
An agent decides within a rulebook you write. It reads the message, works out that this is a reschedule with an added patient, checks what the rules allow, calls the booking tool twice, and confirms. When the message is a symptom, the rulebook says escalate, and it does. The rulebook is the product; the model is the engine that applies it to messy input.
Chatbots and automations are priced as subscriptions, so they look cheap and the cost is the hours your team spends around them: maintaining the script, fixing the Zap when a field is renamed, handling everything that fell through.
An agent is priced as a build plus model usage. The build is the larger number and it buys the rulebook, the integrations, the approval step and the console. Usage is small at SMB volumes and scales with messages. The pricing page has our bands; the cost estimator gives you a range for your case in two minutes.
| Volume | Typical agent running cost per month (USD) |
|---|---|
| Under 500 runs a month | $30–$120 |
| 500–5,000 runs a month | $120–$600 |
| 5,000+ runs a month | $600–$2,500 |
The best setups we see are not one of the three. A Zap moves the form submission into the CRM. The agent reads it, qualifies the lead, books the slot and writes the summary. A chatbot on the website answers the hours question before anyone needs the agent. Each does the part it is good at, and the agent is reserved for the part that needs reading.
That also keeps the agent small, which is what keeps it trustworthy: five tools, a rulebook and an approval step are easy to audit. An agent asked to do everything is not.
Answer them in order. Most people are done by the third.
Does the input have to be read and understood before anything can happen?
Yes: Chatbot or agent. Continue.
No: Zapier, Make or n8n. You are done.
Does something have to be booked, changed, written or sent as a result?
Yes: Agent. Continue.
No: A chatbot with a good knowledge base.
Does the right action depend on the content, not just on which form was filled?
Yes: Agent. Continue.
No: Zapier with conditions, possibly with a chatbot in front.
Can a person approve the actions that matter, at least at first?
Yes: Agent, with an approval step. Start with a pilot.
No: Not yet. Decide who approves, then come back.
Does it happen at least a few dozen times a week?
Yes: Worth building. Scope it.
No: A person with a checklist, until it does.
No. A chatbot has one tool: reply. An agent has several (check availability, book, escalate, update the CRM) and decides which to use. The conversation is one input among others; a scheduled report-assembly agent never talks to anyone.
Zapier can insert an AI step that summarises or classifies, and that is useful. It is still a fixed pipeline: the model produces text, the next step does the one thing it was wired to do. An agent chooses its next step based on what it found, and can loop, retry and ask.
The one with the smallest permission list for the job. A chatbot that can only reply is very safe and very limited. An agent is as safe as its tool list and its approval step, which is why we publish both on every case study.
A chatbot runs $0–$300 a month. An agent pilot is $4,500 fixed for one workflow in two to three weeks, and one workflow in production is $8,000–$42,000, plus model usage. The comparison only makes sense against the hours the workflow currently costs, which is what the savings calculator is for.
Not better, different. A chatbot answers questions from a script or a knowledge base and is the right, cheap tool when that is the whole job. An AI agent reads a message, decides which of several actions applies, does it through your systems and asks a person when the rules say so. If customers only ask things, buy a chatbot; if they ask you to do things, you need an agent.
Who decides the steps. An automation (Zapier, Make, n8n) runs steps you fixed in advance when a trigger fires, and it never reads anything. An agent reads the input, works out which steps apply this time, and can loop, retry and escalate. Automation is the plumbing most businesses are missing; the agent is for the part that needs reading and judgment.
Software that reads a situation, picks an action from a short list it is allowed to take, does it through your tools, checks the result, and hands over to a person when the rules say so. The list of tools is what it can do and the rulebook is what it may do; everything else on this site follows from those two.
Free, no deck. You leave with an approach, a timeline and a price range, or a reason to buy something off the shelf instead.