AI agents vs AI bots vs AI chatbots: what's the difference?
Vendors use all three words for the same products, so the labels tell you very little. What tells you something is how the software behaves: who decides the next step, whether it can act in other systems, and whether it needs a person in the conversation. Here is a plain test for each, and when each one is the right tool.
A chatbot answers people in a conversation, from a script or a language model, and generally doesn't act in other systems. A bot runs a fixed sequence of steps someone defined in advance, such as an RPA bot, a Slack command bot, or a workflow automation, started by a trigger and identical on every run. An AI agent is a language model that works toward a goal by deciding its own next step and taking actions through tools, over several steps, and adapting to what it finds. To tell them apart, ask who decides the next step, whether it can take actions, and whether it runs without a person in the chat.
Three words that vendors use loosely
Customer support tools call their chat widgets agents. Automation tools call their workflows bots. AI assistants that can browse and send email still get called chatbots. None of this is necessarily wrong, because the terms have no official definitions and the products keep gaining each other's features.
That makes the label a poor buying signal. Two products both sold as an "AI agent" can behave in completely different ways: one answers questions on your website, the other updates records in your billing system at 2am with nobody watching. The risks, the costs, and the controls you need are different for each.
So this guide defines the three by behavior rather than by marketing. The definitions overlap at the edges, and the overlap is covered too.
A chatbot answers in a conversation
Software that holds a conversation with a person and answers them. It starts when someone writes to it, and its output is a reply in the chat.
Chatbots come in two broad kinds. Scripted chatbots follow a decision tree: pick an option, get the next prepared message. Language-model chatbots generate their answers, often grounded in a help center or product documentation, so they can handle questions nobody wrote a script for.
Either way, the shape is the same. A person asks, the chatbot answers, and the work ends when the conversation does. A chatbot is good at answering common questions at any hour, collecting details before a hand-off to a human, and pointing people to the right page.
What a plain chatbot doesn't do is act. It doesn't refund the order, update the CRM, or chase the invoice. Some chatbots are given a few narrow actions, such as looking up an order's status. Once a chatbot can choose among actions and take several in a row, it is behaving like an agent with a chat window in front of it, and it needs the controls an agent needs.
A bot runs the steps someone fixed in advance
Software that carries out a predefined sequence of steps when a trigger fires. The person who built it decided every step; the bot only executes them.
"Bot" covers a lot of ground, but the common thread is that the logic is fixed. Three kinds show up most often at work:
- RPA bots, from robotic process automation, which repeat recorded clicks and keystrokes in an application's interface, usually to move data between systems that have no API.
- Chat-command bots, such as a Slack bot that posts the on-call rota when someone types a command, or posts an alert when a deploy finishes.
- Workflow bots, the trigger-action automations built in flow builders: when a form is submitted, create a row, then send a message.
The strengths come from being deterministic. The same input takes the same path every time, so a bot is predictable, cheap to run at volume, and easy to audit: you can read the steps and know exactly what it will do.
The weakness comes from the same place. A bot only handles the cases its builder anticipated. A renamed field, a moved button, or an input that doesn't fit the branches stops it or, worse, sends it down the wrong path. Someone has to maintain the steps as the systems around them change. Adding an AI step, for example to summarize text or classify an email, doesn't change this as long as a person still fixed the order of the steps.
An AI agent decides the next step and takes it
A language model given a goal, a set of tools, and constraints, which decides its own next step, takes actions through those tools, and keeps going over several steps until the goal is met or it needs a person.
The difference from a bot is who writes the procedure. With a bot, a person decides the steps at build time. With an agent, you describe the outcome and the rules, and the agent works out the steps at run time, every run. Anthropic draws the line the same way in its engineering guide, "Building effective agents": workflows are systems where language models and tools are "orchestrated through predefined code paths," while agents "dynamically direct their own processes and tool usage."
That is what lets an agent handle the input nobody planned for. It can read an unusual email, look up the customer in one system, check their balance in another, notice the open dispute, and decide not to send the reminder. Two runs of the same agent can reasonably take different paths.
It is also why agents need more care than the other two. Their output is less predictable by construction, each run costs model time, and when they are wrong the result often looks plausible. An agent that can act needs limits on what it can touch, a person on the consequential steps, and a record of what it did.
Three questions that tell them apart
Ignore what the product is called and ask these three questions about what it actually does.
- 01
Who decides the next step?
If a person decided every step when they built it, it is a bot, however much AI sits inside individual steps. If the software chooses which step comes next based on what it finds, it is an agent. A chatbot only decides what to say next.
- 02
Can it take actions in other systems?
If its only output is a reply to the person in front of it, it is a chatbot. If it creates, updates, sends, or deletes things in other systems, it is a bot or an agent, and the first question tells you which.
- 03
Does it run without a person in the chat?
A chatbot needs someone to talk to it; when they leave, it stops. Bots and agents start from triggers such as a schedule, an incoming email, a webhook, or an event in another app, and they do their work whether or not anyone is watching.
A support assistant on a website that can look up the order, decide a refund is due, and issue it is an agent behind a chat interface. A Slack bot that answers questions with a language model is a chatbot that lives in Slack. Judge the behavior, then apply the controls that behavior needs.
Chatbots, bots, and AI agents side by side
| Chatbot | Bot | AI agent | |
|---|---|---|---|
| What starts it | A person writing to it | A trigger: a schedule, an event, a command | A trigger, or a person handing it a task |
| Who decides the next step | A script, or the model choosing a reply | The person who built it, in advance | The agent, at run time |
| Takes actions in other systems | Usually not, or a few narrow ones | Yes, the ones it was built to take | Yes, through the tools it was given |
| Needs a person present | Yes, someone has to be in the conversation | No | No, though it may pause to ask one |
| Unplanned input | Answers as best it can, or hands off | Fails, or takes the wrong branch | Works out a path, or escalates |
| Predictability | Medium for language-model chatbots, high for scripted ones | High: same input, same path | Lower by design: runs can differ |
| Typical cost | Per conversation or per message | Per run or per step; cheap at volume | Model usage per run, varying with the work |
| Main risk | A confident wrong answer | Silent breakage when a system changes | A plausible wrong action |
| Good for | Answering common questions at any hour | Fixed, high-volume plumbing | Recurring work that needs judgment across systems |
Which one the job needs
None of the three is the advanced version of another. Each is the right tool for a different shape of work, and plenty of teams use all three.
- Use a chatbot when the job is answering people: customer questions on your website, an internal help desk, a guided intake that collects details before a person takes over. If the answer is in your documentation, a chatbot can deliver it at 3am.
- Use a bot when the steps never change: copy every new form submission into a spreadsheet, post a message when a deal closes, move a file nightly. A flow builder or a script does this cheaper and more predictably than an agent, and you should keep it that way.
- Use an AI agent when the work needs judgment across several systems and the right path depends on what it finds: chasing overdue invoices except the disputed ones, triaging inbound requests, preparing a weekly report from three tools, onboarding a new employee.
- Combine them when it helps: a chatbot can collect a request and hand it to an agent, and an agent can call a fixed workflow for the part that never changes.
- Keep it manual when nobody can tell whether the output was right, or when the rules live only in one person's head and haven't been written down yet.
What to check before an agent can act
Because an agent decides its own steps, the questions to ask a vendor are different from the ones you'd ask about a chatbot or a bot. These apply to any agent product, ours included.
- What can it touch? Look for permissions down to the specific action, not access to a whole app, and check whether the limit is enforced by the system or only written in the prompt.
- Who approves the consequential steps? Find out whether a person can be required before a send, a payment, or a deletion, who that person is, and whether they can correct the action rather than only refuse it.
- What starts it? An agent that only runs when someone types to it is a chatbot with tools. Work that arrives as an email, a form, or an event in another system needs a trigger for it.
- What happens when a run fails halfway? Ask whether it resumes from where it stopped and how it avoids repeating an action that may already have happened.
- What record is left? Someone should be able to open a run later and see what the agent decided, which tools it called, with which arguments, and who approved what.
Where Boring AI fits
We make Boring AI, a console for supervised automation, so read this section as the vendor's view. It runs agents in the sense this guide defines: they start from business events, decide their steps, and act through the tools you grant, with a person on the steps that matter.
- Started by business events, not a chat: schedules, webhooks, inbound email, public forms, app events, the Chrome extension, the workspace API, MCP clients, and other agents.
- Limited to exact actions: you grant the specific read and write actions an agent needs across 1,500+ connected tools, and the runtime enforces the list, so a withheld action cannot run.
- Approval on by default: new agents start with every connected action waiting for approval. The reviewer you assign sees the exact call and can approve it, modify the arguments, or reject it. An agent can also pause mid-run to ask a person for missing context.
- A procedure when you need one: numbered instructions run as an enforced plan, one step at a time, which keeps an agent close to a bot's predictability where the order matters.
- A record of every run: steps, tool calls, and approvals stay reviewable, and an eligible failed run resumes from the last completed step instead of starting over.
Boring doesn't put a chat widget on your website or answer your customers' questions in a conversation, and you don't chat with its agents in Slack, Teams, or WhatsApp. The console's operations assistant answers your team's questions about your agents and changes nothing until you confirm; it isn't a customer-facing chatbot. If you need to deflect support questions, use a chatbot built for that. If your steps never change, a flow builder or a script is cheaper and simpler.
Where the two meet, Boring can sit behind the chatbot. A support request forwarded to an agent's email address can start a run that looks up the account, checks the billing history, and drafts a reply that waits for a teammate's approval. If that's the kind of work you have, white glove means our team can build and operate it with you.
Questions about this
What is the difference between an AI agent and a chatbot?
A chatbot answers a person in a conversation and generally doesn't act in other systems; when the person leaves, it stops. An AI agent works toward a goal, decides its own next steps, and takes actions through tools such as a CRM, a mailbox, or a billing system, and it can run from a trigger with nobody in a chat. The same language model can power either one. The difference is the tools, the triggers, and who decides the steps.
What is the difference between a bot and an AI agent?
A bot follows steps a person fixed in advance, so the same input always takes the same path. An AI agent decides its steps at run time based on what it finds. Bots are more predictable and cheaper at volume; agents cope better with input nobody planned for. Adding an AI step to a bot doesn't make it an agent if a person still fixed the order of the steps.
Is an RPA bot an AI agent?
Not by itself. A robotic process automation bot repeats recorded steps in an application's interface, and it does what it was recorded to do. Some RPA products now add AI steps or agent features alongside their bots. Apply the same test: if a person fixed the steps, it is a bot; if the software chooses its next step, it is behaving as an agent.
Can a chatbot become an AI agent?
Yes. Give a language-model chatbot tools and let it choose among them over several steps, and it is an agent with a chat interface. At that point it needs agent controls: limits on which actions it can take, a person on the consequential steps, and a record of what it did.
Are AI agents better than bots?
No, they are suited to different work. For a fixed, high-volume task, a bot is cheaper, faster, and more predictable, and replacing it with an agent adds cost and uncertainty. Agents earn their place on recurring work that needs judgment across several systems, where drawing every branch in advance doesn't hold up.
Does Boring AI build chatbots?
No. Boring AI is our product, a console for supervised automation that runs AI agents started by business events, limited to the actions you grant, with approval on by default and a record of every run. It doesn't provide a customer-facing chat widget. If you need one, use a chatbot product, and Boring can handle the work behind it, such as a support email that needs an account looked up and a reply approved.
Stop reading. Describe the workflow.
Request access and describe it in a sentence — or ask about white glove and our team will build and run it with you.