# AI agent vs chatbot: what's the difference?

> A chatbot answers inside a conversation; an AI agent uses tools and chains steps toward a goal. The difference between a workflow and an agent, examples from social media and customer service, and the controls to ask for before you give it access to your accounts.


**An AI chatbot answers inside a conversation: it gets a message and sends back text. An AI agent works toward a goal: it decides which steps to take, uses tools (your calendar, your inbox, your analytics, an order system), checks the result of each step and keeps going until it's done or until it needs you to decide.** For a business, the difference is what each one can touch. The chatbot leaves you some text; the agent can leave a scheduled post or a sent message. That's why, with an agent, what it asks you to approve first matters as much as what it does.

There's no single definition of "AI agent": vendors, standards bodies and the press use it differently. The distinction above is our own summary. For telling system types apart, a useful reference is ["Building effective agents"](https://www.anthropic.com/engineering/building-effective-agents), by Anthropic, the company that builds Claude (December 19, 2024, read on November 20, 2026).

## An example: the same week with a chatbot and with an agent

*A fictional case.* Paso Firme is a sneaker shop that sells through Instagram. The owner, Carla, runs it with one salesperson.

- **Monday, the week's plan.** Carla types: "Get this week's posts ready." A **chatbot** gives her five captions to copy, paste and schedule herself. An **agent** with access to the account reviews what went out in the last 30 days and how each post did, proposes five posts, saves them as drafts on the calendar and stops: publishing is left to Carla.
- **Tuesday, a DM.** In comes "Do you have the white ones from your last reel in a 42? I need them by Saturday." The **chatbot** answers with what it was given, say the shipping policy and a link to the size chart. If it can't see stock, it doesn't know. The **agent**, connected to inventory, identifies the model from the reel, checks whether a 42 is left, works out whether shipping arrives before Saturday and drafts the reply. Depending on how you set it up, it sends it or holds it for your approval.

Both use a language model and both chat. What changes is that the agent leaves the conversation to look things up and get things done.

## A useful distinction: workflows and agents

Anthropic separates two kinds of systems that combine a model with tools:

- **Workflows:** the model and tools are "orchestrated through predefined code paths." Someone decided the steps in advance.
- **Agents:** systems where LLMs "dynamically direct their own processes and tool usage." The model decides what happens next.

It calls both "agentic systems," and acknowledges that "agent" is used in several ways: some people mean fully autonomous systems that operate independently over extended periods, others mean more prescriptive implementations that follow predefined workflows. The building block of both is an LLM "augmented" with retrieval, tools and memory. And while it works, an agent needs to get the real result of what it did from its environment at each step (what a tool returned, for example) to judge its progress.

**A caveat about the source.**
The same Anthropic page notes that much of the tooling landscape it describes has changed since December 2024. What we take from it is the distinction between workflows and agents, not the list of tools.

Since "agent" means different things depending on who says it, when a vendor offers you one, ask three things: who decides the steps (the model or a fixed script), which tools it can use, and what it does without asking you.

## Chatbot, automated workflow and agent, side by side

| | Chatbot | AI workflow | AI agent |
| --- | --- | --- | --- |
| Who decides what happens | Answers each message according to its instructions | The steps are fixed in advance | The model decides the steps based on the goal |
| What it can touch | The conversation | The tools the workflow is programmed to use | The tools you gave it access to |
| Social media example | Answers opening hours and return policy by DM | Every new DM goes through the AI, which tags it (price, shipping, complaint) and routes it to the matching reply or person, always in that order | Reviews the month, saves drafts on the calendar and works the inbox |
| When it gets something wrong | One wrong message | A design error (a wrong category or destination) repeats on every run until someone fixes the workflow | A mistake can carry into the next steps |

## A chatbot with tools already looks like an agent

The chat window doesn't tell you what's behind it. Anthropic gives customer support as a natural fit for agents, because it "combines familiar chatbot interfaces with enhanced capabilities through tool integration": pulling customer data, order history and knowledge base articles, and handling actions such as issuing refunds or updating tickets. To the customer it's still a chat; for you, it's a system that can move money or change a record.

So the useful question is what it can do outside the conversation. One example of an agent that lives in chats is [Meta Business Agent, Meta's agent for WhatsApp and Messenger](/en/guides/what-is-meta-business-agent/).

## What the risks are and which controls to ask for

Anthropic is blunt about the cost of autonomy: it means "higher costs, and the potential for compounding errors." It recommends extensive testing in sandboxed environments with the appropriate guardrails, letting the agent pause for human feedback at checkpoints or when it hits a blocker, and setting stopping conditions (such as a maximum number of iterations). It also says that to use an agent, you need some level of trust in its decision-making.

Applied to a business's social accounts, here's what we suggest you insist on (our recommendation, not an Anthropic rule):

- **Your approval before anything public.** Publishing, replying in public, deleting a comment and changing an automation should go through you. Whatever appears on your brand's account, someone can see or screenshot before you delete it.
- **Only the access it needs.** Give it access only to the accounts and tools the task requires, and ask up front how that access is revoked.
- **Show where each number came from.** If it says a reel did better, it should name the reel and the period, and separate what's a hypothesis. That way you catch a mistake before it carries into next week's plan.
- **Stop when something is missing.** If it can't find a figure or a tool fails, it should ask you instead of filling the gap with a guess.
- **Start small.** One brand and one task whose result you can check, such as reviewing a draft, before you hand over the whole inbox.

Complaints, legal questions and crises are best answered by a person, agent or not.

## How to tell what you need

Anthropic recommends finding the simplest solution possible and only adding complexity when it's needed. Applied to your case:

1. **If the questions you get have a fixed answer** (hours, address, how to order), a chatbot or a keyword auto-reply is enough.
2. **If a good answer means checking another system** (stock, orders, bookings), you need tools: a workflow if the steps are always the same, an agent if they change from case to case.
3. **If you want to hand off your whole social media workload** (planning, writing, publishing, replying and reporting), that's an agent with access to your accounts. What it does on its own and what you decide, task by task, is in [AI agent for social media](/en/solutions/ai-agent-for-social-media/).
4. **If your problem is automated accounts commenting on your posts**, that's a different topic: we cover it in [social media bots and how to spot them](/en/guides/social-media-bots/).

## How much does an AI agent cost?

There's no single price, for the same reason it differs from a chatbot: it takes more steps. In Anthropic's words, agentic systems "often trade latency and cost for better task performance." The cost depends on:

- **The model and how many times it's called.** Every step (reading analytics, drafting, reviewing) is work for the model.
- **What the messaging platform charges.** If a third-party AI replies through the WhatsApp API, Meta charges per delivered message: its replies are service messages, the first 1,000 each month per phone number are free and from October 1, 2026 the rest are charged, according to its [WhatsApp Business Platform pricing page](https://developers.facebook.com/documentation/business-messaging/whatsapp/pricing). Meta Business Agent, Meta's own agent, is billed separately, by tokens.
- **The software** that packages it, with its plan.
- **Your team's time** to set it up and review what it proposes.

The layers, with their sources and a fictional calculation, are in [how much an AI chatbot costs per month](/en/guides/how-much-does-a-chatbot-cost/). They apply to an agent too: it usually calls the model more often and needs more review.

## Where HeyMark fits

HeyMark is an end-to-end AI agent for social media: it plans, creates, publishes, replies and analyzes. Its agent is called Mark. It lives in HeyMark, writes in your voice, answers your DMs and explains your numbers ([what Mark is and what you can ask it](/en/guides/what-is-mark-ai-heymark/)). If you'd rather work from Claude or ChatGPT, [HeyMark's MCP](/en/mcp/) is free on every plan (AI actions use credits), and there, publishing now, replying in public and changing automations ask for your approval. In ChatGPT, actions such as creating or publishing need developer mode turned on.

## Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.

[Start free](https://app.heymark.ai) · [See how it works](https://heymark.ai/en/#product)

## Developer and agent resources

- [HeyMark MCP documentation](https://heymark.ai/en/mcp/)
- [llms.txt](https://heymark.ai/llms.txt)
- [Full site content for language models](https://heymark.ai/llms-full.txt)
- MCP protocol endpoint: `POST https://mcp.heymark.ai`
- OAuth protected-resource metadata: [/.well-known/oauth-protected-resource](https://mcp.heymark.ai/.well-known/oauth-protected-resource)
- MCP server card: [/server-card](https://mcp.heymark.ai/server-card)
