Short answer: use an automation when the steps are the same every time, a chatbot (or AI assistant) when people need answers in a conversation, and an AI agent when the work needs judgment about which steps to take and in what order. Most businesses should start with the simplest of the three that solves the problem. Automations are the cheapest and most predictable. Agents are the most flexible, but they cost more to build, test and run.
Key points
- Automation follows steps you define. AI can be one step inside it, for example reading an email and pulling out the order number.
- A chatbot talks with people. Older chatbots follow scripts. AI chatbots understand free-form questions and can answer from your own documents.
- An AI agent decides its own next steps and uses tools (your CRM, email, calendar, database) to get a task done.
- The line between them is blurry, and vendors use the words loosely. Ask what the system actually decides on its own.
- Many good solutions combine them, such as a chatbot in front and an automation behind it.
Definitions in plain English
Automation (workflow automation)
A set of steps that runs the same way every time a trigger happens. "When a new invoice email arrives, save the PDF to the right folder, extract the total and add a row to the spreadsheet." Tools like n8n, Make and Zapier are built for this.
Adding AI to an automation usually means one step uses a language model for something rules can't handle well, such as classifying a message, summarizing a document or drafting a reply. The path stays fixed. Anthropic, which makes the Claude models, calls this kind of system a workflow: language models and tools "orchestrated through predefined code paths" (Anthropic, Building effective agents).
Chatbot (and AI assistant)
Software that communicates with people through text or voice. IBM notes that "not all chatbots use AI". Traditional chatbots "follow predefined rules, decision trees and scripted conversation flows," while AI chatbots use large language models to understand intent and respond naturally (IBM, What is a chatbot?).
A common business version is an assistant that answers from your own documents: policies, product manuals, past proposals. It finds the relevant passages and writes an answer based on them. This technique is called retrieval-augmented generation (RAG).
AI agent
A system where the AI model decides what to do next. It picks tools, reads the results and keeps going until the task is done or it needs a human. In Anthropic's definition, agents are "systems where LLMs dynamically direct their own processes and tool usage" (source). IBM describes AI agents as able to "plan, make task-specific decisions and carry out multi-step workflows with limited human involvement" (source).
Example: "Handle this supplier's delivery complaint." The agent looks up the order, checks the delivery status, reads the contract terms, drafts a reply with a proposed credit and asks a manager to approve it before sending.
Side by side
| Automation | Chatbot / AI assistant | AI agent | |
|---|---|---|---|
| What it does | Runs fixed steps when triggered | Answers questions in a conversation | Works out the steps to finish a task |
| Who decides the steps | You, in advance | Mostly scripted, or it only answers | The AI, within limits you set |
| Typical trigger | New email, form, file or schedule | A person asks something | A request or event that needs judgment |
| Predictability | High | Medium (wording varies, scope is narrow) | Lower. Needs testing and guardrails |
| Build effort | Lowest | Medium | Highest |
| Running cost | Low. AI is used for one step or none | Grows with the number of conversations | Highest. Each step is another model call |
| Main risk | Breaks when an input looks unusual | Confident wrong answers if sources are poor | Wrong actions if permissions are too broad |
| Good first example | Sorting and filing invoices | Staff Q&A from your HR and policy documents | Triage of support tickets with drafted replies for approval |
Which one do you need? A quick decision guide
Work through these questions in order. Stop at the first "yes."
- Are the steps the same every time, and can you write them down? Use an automation. If one step needs reading or writing text, add a single AI step.
- Is the main job answering people's questions from information you already have? Use an AI assistant / chatbot that answers from your documents. Keep it read-only at first.
- Does the task need different steps depending on what it finds, across several systems? That's an AI agent. Start with the agent proposing actions and a person approving them.
- Not sure? Start with option 1 or 2 on one process. You can move up later, and most of the work (connections, data cleanup, test cases) carries over.
This matches the guidance from model providers themselves. Anthropic recommends "finding the simplest solution possible, and only increasing complexity when needed", and notes that agentic systems "often trade latency and cost for better task performance" (source).
Common mistakes
- Building an agent for a fixed process. If you can draw the flowchart, an automation will be cheaper, faster and easier to debug.
- Calling a scripted bot an "AI agent." Ask the vendor what the system decides on its own and what it can change in your systems.
- Giving an agent broad permissions on day one. Start read-only or with human approval for anything that sends, pays, deletes or changes customer data.
- Skipping test cases. Collect 20 to 50 real examples, with the right answer for each, before launch. Rerun them after every change.
- Ignoring the data. An assistant is only as good as the documents it answers from. Outdated or conflicting documents produce confident wrong answers.
How the cost compares
Automations are usually the cheapest to build and run. Assistants cost more because of document preparation and testing. Agents cost the most because of the tool connections, guardrails and evaluation work, and because each step is another paid model call. For actual numbers, including a worked monthly example, read how much an AI agent costs to build and run, or see our fixed-price packages on the pricing page.
FAQ
Is ChatGPT a chatbot or an AI agent?
Both, depending on how it's used. A plain conversation is a chatbot. When it browses the web, runs code or works with connected apps to finish a task, it behaves like an agent. The label matters less than what the system is allowed to do.
Can a chatbot become an AI agent later?
Yes. A common path is an assistant that answers questions first, then gains tools step by step, such as looking up an order or creating a ticket, each one tested before the next is added.
Do AI agents replace automation tools like n8n or Make?
Not usually. They work well together. Platforms like n8n and Make now include AI agent features, and an agent often calls fixed automations as its tools. The automation handles the reliable, repeatable part. The agent handles the judgment.
Are AI agents safe for customer-facing work?
They can be, with limits. Keep a human approval step for anything that sends, pays or changes records. Restrict each tool to the minimum access it needs, log every action and test against real examples before going live.
Not sure which one fits your process?
Describe the task in a free 30-minute call. We'll tell you honestly whether it needs an automation, an assistant or an agent, and what the simplest version would cost. See what we build on our AI agents page.