A chatbot works inside a conversation. Someone writes, it answers or does a task, and when the chat ends, so does its work. An AI agent holds a job in your business: it starts on a schedule or when something arrives, and it keeps going whether or not anyone is talking to it.
Many chatbots can now take actions too, so the label alone won't tell you which one you're buying.
| Chatbot | AI agent | |
|---|---|---|
| What starts the work | A customer writes, calls or opens the chat | A schedule, a new email or form, a task, or a person's request |
| When the work ends | When the conversation ends | When the job is done, and then it comes back next time |
| Where it works | Your website chat, help desk and messaging apps | The tools the job lives in: email, calendar, files, customer records |
| Can it change things in other systems? | Many now can, during the chat | Yes, within the access you approve |
| What it remembers | The conversation, your help articles and, in some, the customer's history | The job: what happened last time and what is still open |
| How it is priced | Per conversation or per resolved chat | Per month, for the job |
| Who checks its work | You, by reading transcripts | A named person who approves the risky steps and hears when it fails |
What is the difference between an AI agent and a chatbot?
A chatbot waits for someone to start a conversation and stops when it ends. An AI agent has a standing job. It starts work on a schedule, when an email or form comes in, or when it is handed a task, and it carries that job from one day to the next. Both may run on the same AI underneath. What differs is who starts the work.
You'll often hear that a chatbot answers and an agent acts. That line has moved. Zendesk's pricing page lists "Integrations and actions" for its chat product and describes them as a way to "connect to external systems to read, write, parse data, and trigger actions." HubSpot says its customer agent "takes action in your CRM." So "it can take actions" no longer tells you much.
Analysts flagged the renaming more than a year ago. In June 2025, Gartner warned about "agent washing," which it described as "the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities" (Gartner, June 25, 2025). As a buyer, the simplest defense is to skip the product name and ask what starts its work.
Why do chatbots now call themselves AI agents?
Because the chat products got better and now do more than answer. We read the pages of five chat products that small businesses get pitched, on October 28, 2026. Each one calls its product an AI agent or a customer agent, and each can do something besides reply. They also all charge by the conversation, which shows where the work happens.
| Product | What its page calls it | Beyond replying | How it charges |
|---|---|---|---|
| Tidio Lyro | "Lyro AI Agent" | "Actions," with full access on a paid plan | First 50 conversations free, then $0.50 a conversation |
| Chatbase | "AI agent" | "Actions," plus connections to customer records and help desks | Hobby plan $32 a month billed yearly, or $40 monthly, for 700 message credits |
| Intercom Fin | "Fin AI Agent" | Completes a workflow Intercom calls a Procedure | $0.99 per outcome, with a $49 monthly minimum (50 outcomes) on another help desk |
| Zendesk | "AI agents" | Reads and writes data in other systems | Support Team plan $19 per seat a month billed yearly, with 5 resolutions per seat, then $2.00 each pay as you go |
| HubSpot customer agent | "Customer agent" | Takes action in your HubSpot records | $0.50 per resolution, on HubSpot's Professional or Enterprise editions |
These are useful products, and for many businesses one of them is the right buy. Look at the last column, though. Each one bills per conversation, per outcome or per resolved chat, because each is built to work inside a conversation and has nothing to do between them.
So a better name for this group is a chat agent. It answers and can finish a task while the customer waits, but it doesn't pick up work on its own the next morning.
What starts our own agents' work?
We run VantaSoft on 13 AI agents, so we can show what the other side of the line looks like. We counted every working session our agents ran from October 21 to 27, 2026, and sorted the sessions by what started them.
| What started the work | Sessions | Share |
|---|---|---|
| A schedule, at a set time each day or week | 629 | 52% |
| A message to an agent in our team chat | 272 | 22% |
| A task card on our shared work board | 242 | 20% |
| A request from another agent | 73 | 6% |
| Total | 1,216 | 100% |
Not counted: 1,021 automatic checks and test runs that did no business work.
About one session in five began with a message in our team chat. That row is the only kind of work a chat product is made for. The other four in five started with nobody typing at all: morning briefs, weekly reports, tasks someone filed on the board earlier, and work one agent passed to another. Chat actions can't reach that work, because nobody opened a chat for it.
If you'd like to see what those 13 agents do all day, our agents and their jobs goes one by one.
Which one does your business need?
Pick the job first, then the tool. Each job below is an example, not a customer. Ask two things about it: who or what starts it, and does the work end when the chat ends?
| Example job | Who or what starts it | Ends when the chat ends? | What fits |
|---|---|---|---|
| Answer questions about your hours, prices and policies on your website | A visitor | Yes | A chatbot |
| Look up an order or book a time while the customer waits | A customer | Yes | A chat agent with actions |
| Take an after-hours inquiry and add it to your customer records | A customer | Yes | A chat agent with actions |
| Write back to people who chatted but never booked, two days later | A set time after the chat | No | An agent with a standing job |
| Read yesterday's chats each morning and send you the complaints | A daily schedule | No | An agent with a standing job |
| Spot a question that keeps coming up and draft a help article for you to approve | A weekly check | No | An agent with a standing job |
A chat product handles the first three rows well. The last three are the ones that slip, because something has to remember the chat happened and come back to it later.
Plenty of businesses with a lot of repeat questions end up running both, one for the chats and one for the follow-through. If most of your list is customer messages, AI Agent for Customer Service shows how to sort your last 100 messages before you buy anything. If the tool you have in mind is ChatGPT or Copilot rather than a website chat, where an AI assistant for business stops runs ten office jobs through a similar test.
How much does a chatbot cost compared with an AI agent?
The two are priced on different units, so compare them by the job. Most of the chat products above charge between $0.50 and $2.00 for each conversation or resolved chat. An agent with a standing job is priced by the month, because it works whether anyone writes in or not.
To compare them, count how many conversations a chat product would handle in a month and multiply by its price per chat. Then put a number on the jobs that slip today, in staff hours or in leads that never heard back.
Our own pricing is public. We build an agent around a job in your business starting at $1,000 one time, which includes three coaching sessions for your team, and hosting and maintenance start at $200 a month, as shown on our pricing page on October 28, 2026. For what other routes cost, from development firms to building one yourself, see How Much Does AI Agent Development Cost?
Frequently asked questions
Is ChatGPT a chatbot or an AI agent?
ChatGPT is a chat assistant that now has agent features. In agent mode, you describe a task and it works through it, pausing to ask you when it needs to. Scheduled tasks can run at a set time or when a supported app event happens. All of it runs on the account of the person who set it up, so it handles that person's errands rather than a job the business owns.
Can a chatbot be upgraded into an AI agent?
Partly. Most chat products now add actions, such as booking a time or updating a record during the chat. What they don't add is a job outside the conversation. For that, the agent needs its own login to your tools, a schedule or trigger, and a person who approves the risky steps.
Is an AI agent riskier than a chatbot?
It can be, because it changes things in your systems when nobody is watching. So each agent should get only the access its job needs, and payments, hiring decisions and anything legal should wait for a person. What Access Should an AI Agent Have? and Human in the Loop for AI Agents go through it step by step.
Do I need a chatbot before I get an AI agent?
No. They do different jobs. If customers write in with the same questions all day, a chat product is the place to start. If your problem is follow-ups, reports and checks that fall through, a chatbot won't fix it, and you can go straight to the agent.
Which is cheaper for a small business?
For a steady flow of repeat questions, a chat product usually costs less, since you pay only for the chats it handles. For work that has to happen every day without anyone asking, a price per chat doesn't apply, because there is no chat. Weigh each one against what it costs you when the job goes undone.
Start with the job that slips most
Write down the one job that slips most often in your business. Then answer two questions about it: what should start it, and what has to happen after the conversation ends? If the answer to the second is "nothing," a chat product is probably enough. If someone has to come back to it later, you need an agent.
AI Agents for Business lays out the four ways to get one. Once you have the job written down, talk to us about it. It starts with 30 minutes on your workflows, and the point of the call is to decide whether that job is worth handing off.
