AI agent, workflow or chatbot: which one do you actually need?

If the work follows the same steps every time, you need a workflow. If a product already does the job as it ships, buy it and set it up well. Build a custom AI agent only when the work needs judgment, case by case, across your own systems. Most requests we see that start as "we need an agent" turn out to be a workflow with one AI step.

Not sure which one you have? Answer three questions and get an answer on the page.

What is the difference between a workflow, an AI agent and a chatbot?

The difference is who decides the next step. In a workflow, your rules decide, and the steps never change. In an AI agent, the agent decides, within limits you set, and it can take actions in your systems. A chatbot answers questions in a conversation, and usually stops there.

What is an AI agent, in one sentence?

An AI agent is software that uses an AI model to decide its own next step toward a goal, using tools you give it, like looking up an order or sending a draft, and repeating until the job is done or it needs a person. Anthropic's widely cited definition draws the same line: in a workflow, code sets the path; in an agent, the model directs its own process and tool use.

Is agentic AI the same as an AI agent?

Mostly, yes. "Agentic AI" is the broader label for AI that takes actions and works through multi-step tasks; an AI agent is one system that does that. Gartner warns about "agent washing", vendors relabeling chatbots, assistants and RPA as agentic, so ask what the product decides on its own and what it is allowed to change.

What is an agentic workflow?

An agentic workflow is a fixed process where one or more steps use an AI model to do something a rule can't, like reading an email to tell a complaint from an order, or pulling the PO number off a scanned page. The path is still set by you. The AI does the reading, and your rules decide what happens next.

This is the right answer more often than people expect. It is cheaper to build, easier to test and easier to trust than a full agent, and you can turn one step into an agent later if the exceptions pile up.

Is a chatbot an AI agent?

Usually not. A chatbot answers. An agent acts: it updates the order, books the meeting, applies the credit. A chatbot that can look up an order status is somewhere in between. The useful question isn't what it's called, but what it's allowed to do without a person, which is how we set up control for every agent.

When is a workflow enough?

• You can write the steps on one page, and they rarely change.
• When something unusual happens, it's fine for the process to stop and ask a person.
• The inputs are clean: forms, fields, records, not free-text emails or scans.
• If the inputs are messy but the steps after reading them are fixed, add one AI step and keep it a workflow.

When do you need a custom AI agent?

• Most cases need judgment. A person reads each one and decides what to do, and the right move depends on the details.
• The work spans your own systems. It has to look things up and act in your ERP, CRM, helpdesk or software no product connects to.
• The exceptions are the job. The easy cases are already automated, and what's left is the part people spend their day on.
• Actions need rules. Refunds, discounts or order changes that should happen on their own below a limit and go to a person above it.

When should you just buy an off-the-shelf AI tool?

When the job is one most companies do the same way: answering questions from a help center, taking meeting notes, a basic phone menu. Buy the product, and spend the effort on setting it up well, which is where most of the value is. Our AI setup for small teams does exactly that.

The trade-off is control. You get the vendor's way of working, its connections and its limits, and you can't see much of what happens inside. Many teams start with a product and move to a custom agent when they hit a wall: a system it can't connect to, a rule it can't follow, or a subscription that costs more each year than running their own.

What about RPA?

RPA (robotic process automation) clicks through screens the way a person would, following a script. It's a workflow that drives the screen instead of an API. It works for stable, repetitive screen work and breaks when a screen changes. An AI agent reads and decides; it can use RPA-style screen control as one of its tools when a system has no other way in.

What do common requests usually need?

Here's how we'd start with the requests we see most often. The answer can change once we see the real work.

How often is an "agent" request really a workflow?

Often. In our review of 63 recent public job posts from companies hiring for AI agent work, the two most common asks were multi-step workflows (25%) and integrations between existing tools (24%). Many of those are fixed sequences with one AI step, which is a workflow. Calling it an agent isn't wrong, but building it as one usually costs more and is harder to test.

Does it matter if it's one agent or five?

Not to you. Whether it's built as one large agent or a few small ones working together is an engineering choice, and it changes as the tools improve. What matters is which jobs it takes over and how hard each one is: how many systems, how many decisions, how much is at stake when it's wrong. That's also what should drive the cost, which we break down in how much a custom AI agent costs.

How do you decide for your own team?

• Pick one job, not a department.
• Write down the steps your best person takes, and circle every point where they decide something.
• No circles: a workflow. Circles only where they read something: a workflow with an AI step. Circles everywhere: an agent.
• Before building, check whether a product already does it. Then choose the simplest option that does the job.

Common questions

Are agentic workflows the same as AI agents?

No. An agentic workflow follows a path you set, with AI doing some of the steps. An AI agent chooses its own path toward a goal, within limits you set. Many products called agents are agentic workflows, and that is often a good thing.

Is ChatGPT an AI agent?

On its own it's an assistant: it answers and drafts. With tools turned on, it can act for you, which makes it agent-like. A business agent is built around one job, connected to your systems and governed by your rules.

Can a workflow grow into an agent later?

Yes, and it's often the best path. Start with the workflow, watch where it stops and asks a person, and turn the step with the most exceptions into an agent once you know what those exceptions are.

Is an agent more expensive than a workflow?

Usually, to build and to run, because it needs guardrails, testing on real cases and ongoing tuning. That is why we recommend a workflow or an off-the-shelf tool whenever one does the job.

Will you tell us if we don't need a custom agent?

Yes. If a workflow or a product you can buy does the job, we'll say so on the first call.