AI agent vs chatbot, and which one you need first

A chatbot answers your customers. An AI agent answers them and then does something about it, such as refunding an order, changing an address or booking a call, inside the other systems your business runs on. That’s where the big support vendors draw the line. Zendesk’s guide says “chatbots simply respond, while AI agents are capable of working toward a resolution”. Salesforce’s primer says chatbots “rely on preset scripts to answer basic queries”, while AI agents “autonomously reason, make decisions, and execute complex workflows”. For a small team the question is simpler. Do your customers mostly need an answer, or do they mostly need something changed?

Deacon is an AI customer support agent for founders who answer support themselves, and it sits on the answering side of that line. It answers your customers in seconds from your own docs, asks one question back when a question is vague, says so when it doesn’t know, and hands the conversation to you. It also shows you what customers are stuck on and what they want next. There’s a free plan, and it needs no card.

We read every vendor quote and price below on the vendor’s own pages on 25 September 2026, and each one links there.

AI agent vs chatbot at a glance

The two words really cover three kinds of tool, because “chatbot” is used both for scripted bots and for AI chatbots that write their answers from your content.

Where the line really sits

Vendors tend to describe a chatbot by its oldest kind, which makes the gap to an agent look wider than it is. Zendesk’s guide defines an AI chatbot as “a tool that follows pre-defined rules to interact with customers”, programmed to recognise keywords and reply with scripted answers. Plenty of AI chatbots sold today don’t work that way, Deacon included, so it helps to look at all three kinds on the vendors’ own pages.

A rules chatbot follows a script

Tidio sells both of the first two kinds in one product, which makes the difference easy to see. Its pricing page says its Flows “follow user-defined rules and scripts to answer customer questions” and calls them “perfect for answering simple, pre-defined questions”. You build every branch yourself. A customer who types something no branch expected gets nowhere.

An AI chatbot answers from your content

On the same page, Tidio says its AI, Lyro, “uses solely your support content to generate personalized responses” and “is not restricted to pre-defined questions”. That’s the middle kind. It reads a question however it’s phrased and writes an answer from what you’ve written down, which is called grounding. There are no flows to build. What it needs is good content, because a careless one fills a gap with something that sounds right, which is why support bots make things up. A good one says when it doesn’t know and gets the customer to a person.

The word “agent” is used loosely here. In support it often means the AI doing a support agent’s job, which is mostly answering, and that’s the sense in which Deacon calls itself an AI customer support agent. Conversational AI for customer service sets the kinds of support bot side by side, from scripted bots to agents that act.

An AI agent takes actions

An agent connects to the systems where your customers’ orders, bills and accounts live, and changes things there. Gorgias puts it plainly on its AI Agent page, “Answers are nice. Actions close the loop.” Its agent can issue a full or partial refund, cancel an order that’s still eligible and update an address before fulfilment, directly in Shopify. Fin, the agent sold with Intercom’s helpdesk, says on its home page that it “updates accounts, processes payments and refunds, troubleshoots technical issues, and much more”. Zendesk’s AI agents promise to “take action across systems”, in the words of its AI agents page.

The line is often a switch inside one product rather than a different product. Tidio’s pricing page says Lyro’s first Action is free and full access comes with a paid Lyro plan. So Lyro is an AI chatbot until you give it an action, and an agent for the jobs you’ve connected.

Every action asks three things of you before it’s safe. The agent needs a connection to the system, such as your billing or your store. It needs to know who it’s talking to, because a refund to the wrong person is worse than no refund. And it needs a rule for when the action is allowed, such as a cancellation only before an order ships.

Where Deacon sits between the two

By the “takes actions” definition, Deacon is closer to an AI chatbot. It answers your customers from your own content, your website, files, text you paste in and the answers you write, in seconds, at any hour and in the language each visitor writes in. It takes no actions in other systems. It won’t refund an order, change a plan or look up an account, and it always declines a question about a visitor’s own account, invoice or data, whatever your docs say.

In how it handles a conversation, it behaves more like an agent. It makes three judgement calls on its own.

• When a question is too vague to answer, it asks one question back. After three it stops asking, sums up what the visitor said in one sentence and hands over.
• When your content doesn’t cover a question, it says it doesn’t know instead of guessing, and offers the visitor a box for their email.
• When a visitor leaves an address or asks for a person, the owners on your team get an email. Your reply from the dashboard appears in the visitor’s chat while it’s open, under your first name and your company’s, and Deacon stops answering in that conversation until you hand it back.

Your site can also pass Deacon details about the visitor, such as their plan, and it takes them into account when it answers. It reads that context. It doesn’t act on it.

It gets better the more you answer. Every question your content didn’t cover waits in Conversations with an Answer this button. Write the answer once and press Save & check, and Deacon shows you the reply it now gives and how many other open questions your answer also covers. Every conversation is saved and searchable by what visitors asked, and Topics group similar questions, count the people who asked and label most of them as a how-to, a bug report, a feature ask and so on. That list is also how you’d spot the one request worth automating later.

Every feature is on every plan, the free one included. Free gives you 50 answers a month with no card and no end date. Starter is $50 a month, or $40 a month billed yearly, for 1,000 answers a month. Nothing is billed past a plan’s answers, and no plan upgrades itself.

Why a founder needs answers before actions

Most of what customers ask a small software company is already written down somewhere. How to set something up, where a setting lives, what a plan includes, why an import failed. None of that needs a system changed. It needs the right page found and explained in seconds, at 2am their time, without you.

An action is where a mistake costs money. A wrong answer can be put right in the next message, or by you an hour later. A refund that went out, an order that was cancelled or a plan that was changed has to be undone, and some can’t be. Giving that to an AI is worth it once you’ve seen the same request arrive often enough to write a firm rule for it.

Actions also take building. Each one needs a connection, a way to check who’s asking and a rule, and each is one more thing to test when your billing or your store changes. Answering from your docs needs your docs, which you’re writing anyway.

And answers show you what to automate. Once every question lands in one place with a count, you can see whether one request, such as “cancel my order”, is common enough to hand to an agent. Until then, you’d be automating a guess. Customer service automation for founders sets out what to automate at each stage.

When you need an agent that acts, and what it costs

You need an agent that acts when the same request arrives many times a week, lives in a system with an API and follows a rule you’d happily write down. Looking up an order, changing an address before shipping, cancelling inside a window and booking a slot are the usual ones, which is why online shops tend to need agents before software companies do. If a person on your team spends hours a week clicking the same button in the same admin screen for customers, that’s the job.

Agents are mostly billed by the conversation, the action or what they settle, rather than a set price for a set number of answers, so a busy month costs more.

• Fin charges $0.99 an outcome in support, at most one a conversation, even if it takes several actions in it.
• Zendesk bills its AI agents by the automated resolution, a request settled without a person, and its pricing page prices each one past a plan’s allowance at $2.00.
• Salesforce sells Agentforce at $2 a conversation, or by the action, where each one uses 20 Flex Credits and 100,000 credits cost $500, so $0.10 an action.
• Gorgias adds its AI Agent to its Starter helpdesk at $30 a month for 30 automated interactions, then charges $1.50 for each one past that. Actions such as cancelling an order start on Basic, $90 a month with the AI Agent.

The vendors that sell to large companies, such as Decagon and Sierra, publish no prices at all. The best AI agents for customer service compares nine, with what each can act on and how it charges.

Questions about AI agents and chatbots

What is the difference between an AI agent and a chatbot?

A chatbot replies to customers, either from a script or, if it’s an AI chatbot, from your own content. An AI agent replies and can also take actions in your other systems, such as issuing a refund or changing an order. Most vendors draw the line at those actions.

Is an AI agent better than a chatbot for customer service?

Only if your customers need things done that you’re happy to automate. If most of their questions are answered by your docs, an AI chatbot that answers well and hands over to a person does the job, with less to build and less that can go wrong. An agent earns its setup once one request, such as cancelling an order, arrives often enough to write a rule for.

Is Deacon an AI agent or a chatbot?

Both words fit, in different senses. Deacon calls itself an AI customer support agent because it does a support agent’s job of answering customers, and it makes its own calls on when to ask a question back, when to say it doesn’t know and when to bring you in. By the definition the big vendors use, where an agent takes actions in other systems, it’s closer to an AI chatbot, because it takes none.

Can Deacon issue refunds or change orders?

No. Deacon answers questions and hands conversations to you, and it takes no actions in other systems. It always declines questions about a visitor’s own account, invoice or data, so a question about a customer’s own bill gets a way to reach you instead of a guess. Your reply from the dashboard appears in the visitor’s chat while it’s open, and human handoff shows how that works.

Is an AI agent the same as live chat?

No. Live chat is a person replying in a chat window, and an AI agent is software that replies and acts by itself. Many tools put both in one window, with the AI answering first and a person taking over. Chatbot vs live chat covers the setup where an AI answers and you step in.

Will AI agents replace customer service staff?

Not the whole job. They can take the questions your docs answer and the requests you’ve written rules for, at any hour, and pass the rest to a person. Judgement calls, exceptions and upset customers still need someone. Will AI replace customer service? looks at what changes and what doesn’t.

Which AI chatbots are good for customer service?

The good ones answer only from your content, say when they don’t know and get the customer to a person. The best AI chatbots for customer service ranks nine for small teams, with prices and the plan where a person can take over.