AI agent vs chatbot vs workflow automation: what you actually need
A chatbot answers, a workflow follows fixed steps, and an agent chooses its own steps and uses tools to finish a task. Most companies need a workflow with one or two model steps more often than they need an agent. Pick an agent only when the path through the task cannot be written down in advance.
The three in one table
| Chatbot | Workflow automation | AI agent | |
|---|---|---|---|
| What it does | Replies to questions in conversation | Runs the same steps every time a trigger fires | Plans steps, calls tools and checks its own progress |
| Who decides the path | The user, turn by turn | The person who designed the flow | The model, within limits you set |
| Can it act in other systems | Rarely, mostly read-only | Yes, through fixed connections | Yes, through tools it selects |
| Predictability | High for covered questions | Very high | Lower, so it needs evaluation and limits |
| Typical example | Answering policy questions from a knowledge base | New order creates an invoice and a shipping label | Investigating a billing dispute across three systems and drafting a resolution |
| Main running risk | Wrong or outdated answers | Breaks when an input format changes | Takes a wrong action or loops, with higher usage cost |
When a chatbot is enough
Use a chatbot when the job is to find and explain information: product questions, internal policies, onboarding material. With retrieval over your own documents and clear source links, this covers a large share of support and internal help requests.
It stops being enough when users expect it to do something, such as change a booking or update an address. At that point you are adding tools, and the design questions become those of an agent.
When a workflow is the better choice
If you can draw the process as a flowchart and the branches are known, build a workflow. Model steps fit inside it well: classify an email, extract fields from a PDF, draft a reply for a person to approve. The flow stays fixed and only the language-heavy step uses a model.
Workflows are cheaper to run, easier to test and easier to explain to an auditor. Many projects sold as agents are better delivered this way.
When you need an agent
An agent earns its extra cost when the task has too many paths to enumerate and each case needs a different sequence of look-ups and actions.
- The inputs are open-ended, such as free-text requests that span several topics.
- The number of steps varies per case and depends on what earlier steps return.
- The work means gathering evidence from several systems before deciding.
- A person doing the task today relies on judgement more than on a checklist.
- Mistakes can be caught by a review step or undone before they cause harm.
A five-question test
- Can you write the steps as a flowchart that covers nine cases in ten? If yes, build a workflow.
- Does the task end with an answer and no action? If yes, a retrieval chatbot is likely enough.
- Does each case need a different order of tool calls? If yes, consider an agent.
- Can a wrong action be reversed or reviewed before it takes effect? If no, keep a person in the approval step.
- Do you have real past cases to test against? If no, collect them first, whichever option you pick.
What this means for choosing a partner
Bring the test results to your first agency call. A good partner will challenge the answer in both directions and may propose something smaller than you asked for. Be wary of any firm that recommends an agent before it has seen examples of the task.
If your result is mostly workflows, look for a team that builds both and can keep them in one architecture, so the agent you add later reuses the same connections, permissions and logs.
Common questions
Is a multi-agent system better than a single agent?
Not by default. Splitting work across several agents helps when sub-tasks need different tools, permissions or context. It also adds cost, latency and more ways to fail. Start with one agent and a clear tool set, and split only when evaluation results show a specific bottleneck.
Can a workflow tool run an agent?
Yes. Tools such as n8n and custom orchestration code can call an agent as one step in a larger flow. This keeps triggers, retries and logging in the workflow and limits the agent to the part that needs judgement.