What is the difference between AI agents and chatbots?
The whole of AI agents vs chatbots comes down to one word: a chatbot answers, an agent acts. A chatbot receives a message and returns text. An agent receives a goal, decides on a sequence of steps, uses tools to carry them out, checks what happened, and adjusts. One is a conversation. The other is work.
That difference sounds academic until you price it, supervise it or get it wrong. Then it becomes the most important thing on the page.
What does a chatbot actually do?
A chatbot maps an incoming message to a response, and its job ends there.
The good ones are genuinely good. Fed your real services, prices and policies, a modern one will answer most customer questions accurately, in your tone, at three in the morning. If your problem is that people ask you the same forty questions and you cannot get to them all, a chatbot solves your problem completely and you do not need anything more complicated.
Its limit is that nothing changes in your business as a result. The calendar is not updated. The job is not created. The invoice is not chased. Somebody still has to read the transcript and do the thing. If the doing was the bottleneck, a chatbot moves it rather than removing it.
What does an AI agent do that a chatbot cannot?
An agent takes actions in your actual systems, across multiple steps, without a human directing each one.
It books into the real calendar, having first checked real availability. It creates the job in the job management system. It reads a document, extracts what matters, compares it against your history and writes a recommendation. It notices that a step failed and tries the next sensible thing rather than stopping.
The practical test is simple: after the interaction is over, has something in your business changed. If the answer is only "there is a nice transcript", it was a chatbot regardless of what the brochure called it.
Most of our crew are agents in this sense. Keeper does not just discuss a job, it qualifies and books it. Mason does not summarise a quote, it benchmarks the pricing line by line and returns a verdict. That is the distinction that makes them worth building.
Which one does your business need?
You need a chatbot if the pain is answering, and an agent if the pain is doing.
Ask what happens today after the customer's question is answered. If the answer is "nothing, we just needed to reply faster", buy the simpler thing and be pleased about the price. If the answer is a chain of five internal steps that somebody has to remember to perform, that chain is the actual project, and a chatbot bolted onto the front of it will make the backlog arrive faster rather than smaller.
Plenty of businesses genuinely need the simpler thing. We would rather tell you that than sell you an agent for a problem a chatbot fixes.
Why do agents cost more to build?
Because acting in the real world requires permissions, fallbacks and supervision, and all three take engineering that a text reply does not.
A chatbot that gets something wrong produces a bad sentence. An agent that gets something wrong produces a double booking, an email sent to a client, or a payment chased that was already paid. So an agent build carries things a chatbot build does not: boundaries on what it is allowed to touch, defined behaviour when a system is down, logging of every action taken, and human approval gates on anything consequential.
That is also why the honest version of agent pricing is scoped rather than fixed off a price list. The admin disappears is scoped from $4,950 with a $495 monthly run fee for exactly this reason: the shape of the workflow decides the work, and quoting it blind would be guessing with your money.
How much freedom should an agent be given?
Very little at first, then more as it earns it, with the people who do the job deciding when.
We run every agent supervised by default. It proposes, a human approves, and that continues for as long as your team wants it to. After a few hundred approvals, patterns are obvious: this category it always gets right, that category it sometimes misreads. Then you widen autonomy on the first and keep the gate on the second.
The version where an agent is trusted on day one is how the alarming stories happen. Autonomy is granted, gradually and with evidence, not handed over at go-live. That is one of our standing commitments rather than a setting.
What should you ask a supplier before buying either?
Ask what it can change, what it can never change, and what happens when it is wrong.
Those three questions separate a real build from a demo very quickly. A supplier who cannot list the systems it will write to has not designed it yet. A supplier with no answer on the boundaries has not thought about the failure mode. And a supplier who has not defined what happens when it is wrong is planning for you to find out on a live customer.
Add a fourth: whose accounts does it run in. Everything should run in accounts you own, so that if the supplier vanished, the thing keeps working. That is a clause, not a courtesy.
Where to look next
If you want to see the difference rather than read about it, two of the crew are playable on a staged business: one holds a customer conversation and books, the other reads a quote and returns a costed verdict. That contrast makes the point faster than any diagram.
For the governance frame, the Australian Government's AI Ethics Principles set out human oversight and contestability, and the Australian Cyber Security Centre publishes practical guidance worth reading before an agent gets keys to anything.
If you are unsure which one your problem needs, describe the problem to us. We will tell you the cheaper answer when the cheaper answer is right.