AI Agent vs Chatbot vs Workflow Automation: What's the Difference?

Ask five vendors what an "AI agent" is and you will get five answers. Some mean a chat widget on your website. Some mean a Zapier flow with a GPT step in the middle. Some mean software that can plan and carry out multi-step tasks on its own. The labels have blurred, and that makes it hard to know what you are actually buying.

This article separates the three things people usually mean: chatbots, workflow automation and AI agents. For each one we cover how it works, what it is good at, where it breaks, and what it typically costs to run. At the end there is a simple decision guide.

The short version

	Chatbot	Workflow automation	AI agent
Triggered by	A person typing a message	An event (new email, form, order, schedule)	A goal or request, from a person or an event
Path through the work	Conversation, one turn at a time	Fixed steps you design in advance	Steps chosen by the model at run time
Can take actions?	Sometimes, usually limited	Yes, exactly the ones you define	Yes, from a set of tools you allow
Predictability	Medium	High	Lower, needs guardrails
Best for	Answering questions	Repeatable processes	Variable tasks with many possible paths

Chatbots: a conversation interface

A chatbot is software that talks with people through text or voice. Older chatbots were decision trees: "Press 1 for orders, 2 for returns". Modern ones use a large language model (LLM) such as Claude or GPT to understand free-form questions and write natural answers.

The defining feature is the interface, not the intelligence. A chatbot waits for a person to say something, then responds. Most useful business chatbots are connected to a knowledge source, such as your help center, product catalog or policy documents, so they answer from your content rather than the model's general knowledge. That technique is called retrieval-augmented generation, or RAG, and it is the backbone of most knowledge base assistants.

Good at: answering repetitive questions at any hour, guiding people to the right page or product, collecting information before handing off to a human.

Weak at: anything that requires changing data in your systems, unless you deliberately give it that ability. A chatbot that only reads your FAQ cannot issue a refund or update a delivery address.

Typical example: an ecommerce store adds a chat assistant that answers questions about sizing, shipping times and return policy using the store's own help articles, and hands the conversation to a Gorgias agent when the customer asks about a specific order.

Workflow automation: fixed steps, triggered by events

Workflow automation connects your tools so that when something happens in one, defined actions happen in others. A new Shopify order creates a row in Google Sheets and a task in Asana. A signed contract in DocuSign updates the deal in HubSpot and posts in Slack. Tools like Zapier, Make and n8n are built for this, and custom code does the same job when you outgrow them.

The key property is that you, the designer, decide every step in advance. The automation does not choose what to do next. It follows the path you drew, including the branches and error handling you added.

AI fits into workflow automation as individual steps. A model can classify an email, extract fields from an invoice, summarize a call transcript or draft a reply. But the overall flow is still fixed. This is the pattern behind most practical "AI automation" today, and it is what our workflow automation projects usually look like.

Good at: repeatable processes with clear inputs and outputs, high volume work, anything that must behave the same way every time, and anything you need to audit.

Weak at: tasks where the right sequence of steps depends heavily on the situation. If you find yourself drawing a flowchart with forty branches, the work may be too variable for a fixed workflow.

Typical example: a professional services firm gets new client inquiries through a website form. A workflow enriches the company details, uses a model to classify the inquiry type and urgency, creates a deal in Pipedrive, assigns it to the right partner and sends a tailored acknowledgment email.

AI agents: the model chooses the steps

An AI agent is a system where a language model is given a goal, a set of tools, and the freedom to decide which tools to use and in what order. Tools are functions the agent can call: search the CRM, look up an order, check stock, create a ticket, send an email, query a database. The model reads the situation, picks a tool, looks at the result, and decides what to do next, repeating until the task is done or it needs help.

That loop is what makes something an agent. It is also what makes agents both powerful and harder to control. Because the path is not fixed, you need to think carefully about which tools the agent can use, what limits apply, and when it must stop and ask a person.

Good at: tasks that are similar in goal but different in detail. Resolving "where is my order?" when the answer might involve the carrier, the warehouse, a backorder or a split shipment. Researching a new lead across your CRM, their website and past emails before a sales call. Handling a support ticket that could need any of six different actions.

Weak at: high-volume work with a single correct path (a workflow is cheaper and more reliable), anything where an error is costly and hard to reverse, and tasks where the tools or data it needs do not exist yet.

Typical example: a customer support agent that reads incoming tickets in Zendesk, looks up the customer and order, checks the carrier tracking API, and either resolves the ticket with a reply, issues a small goodwill credit within a set limit, or escalates to a human with a summary of what it already checked.

The overlap is real, and that is fine

In practice, many systems combine all three. A chat widget (chatbot interface) talks to an agent (which chooses tools) that kicks off workflows (fixed processes) when it needs something done in a standard way. A sales lead agent might chat with a website visitor, decide whether they are a good fit, and then trigger the same lead routing workflow your marketing forms already use.

So the useful question is not "which label applies?" but "for each part of this job, how much freedom does the software need?" Give it as little as the task allows.

Cost and reliability: the trade-off nobody mentions

Each approach has a different running cost profile:

• Workflow automation without AI costs very little per run. You pay for the platform or hosting, and that is mostly it.
• Workflows with AI steps add a model call or two per run. For classification or extraction this is usually a small cost per item, but it adds up at high volume, so it is worth choosing a smaller model where it performs well enough.
• Chatbots make one or more model calls per message, plus retrieval. Long conversations with lots of context cost more than short ones.
• Agents often make many model calls per task, because each step in the loop is a call. A task that takes eight tool calls is eight or more model requests. Agents also need more logging and monitoring.

Reliability follows the same order. Fixed workflows are the most predictable. Agents are the least, which is why they need clear tool permissions, spending or action limits, and a human review path for anything consequential.

How to choose: a decision guide

Work through these questions in order for the job you have in mind:

1. Is the main job answering questions from people? Start with a chatbot connected to your knowledge base. Add actions later if needed.
2. Can you draw the process as a flowchart with fewer than ten or so branches? Use workflow automation, with AI steps where something needs to be read, classified or written.
3. Does the right sequence of steps change a lot from case to case, and do the tools it needs already exist as APIs? Consider an agent, with tight limits.
4. Is a wrong action expensive or hard to reverse? Whatever you choose, add a human approval step before that action.
5. Is the volume high and the logic simple? Avoid agents. A workflow will be cheaper and more reliable.

A worked example: one problem, three designs

Imagine a B2B parts supplier whose customers email to ask about order status, request copies of invoices and ask for stock availability on specific part numbers.

• Chatbot design. A portal chat assistant answers questions about policies and lead times from documentation. It cannot see individual orders. Cheap and safe, but it only solves part of the problem.
• Workflow design. Incoming emails are classified by a model into three types. Each type runs a fixed flow: look up the order and reply with status, fetch the invoice PDF from QuickBooks and attach it, or check stock in the ERP and reply with availability. Anything that does not fit goes to a person. Predictable and auditable.
• Agent design. An agent reads each email and can call tools for orders, invoices, stock and the customer record. It handles emails that combine several requests ("Where is PO 4471, and can you send last month's invoices?") without needing a branch for every combination.

For most suppliers, the workflow design is the right starting point. It covers the bulk of the volume with predictable behavior. If, after a few weeks, a meaningful share of emails are falling through because they combine requests, that is the evidence you need to move those cases to an agent.

Questions to ask any vendor

When someone pitches you an "AI agent", ask:

• Which steps are fixed and which does the model decide?
• Exactly which actions can it take in our systems, and what limits apply?
• What happens when it is unsure? Who gets notified and how?
• How are conversations and actions logged so we can review them?
• What does it cost to run per conversation or per task, roughly, at our volume?
• Which model does it use, and can that be changed later?

Clear answers to these questions matter far more than the label on the box.

Where to go from here

If you are not sure which approach fits your situation, describe the job rather than the technology. You can read about how we scope projects, or share what you are trying to automate and we will suggest the simplest design that would work.