AI Agent, Chatbot or Workflow: What You Actually Need, and When

"We need an AI agent" is something I hear in almost every first call. In eight out of ten cases, the person doesn't need one. The question of AI agent vs chatbot vs workflow decides whether you buy the right tool for your problem, or the most expensive one.

All three are currently sold under the label "AI solution". They solve different problems, cost different amounts and break in different ways. Here's the clean separation, with examples from our own brands and a decision table at the end.

Workflow: the fixed track

A workflow is a predefined chain of steps. When X happens, do Y, then Z. Every step is fixed in advance. There's no AI in it, unless you build it in as one of the steps.

That sounds boring, and it's the most reliable of the three forms. A workflow does exactly what was programmed, the same way every time, traceable down to the last step.

Example 1, from our own operations: order processing at nano. An order in Shopify fires a webhook, meaning an automatic message to our system. n8n sends the data to the fulfiller, picks up the tracking number later and informs the customer. It used to be 4 hours of manual work per day, today it's 15 minutes for special cases. The complete setup is in Order Processing Without Staff.

Example 2: returns at mate. Customer fills in a form, a label is created automatically, and after the goods arrive the refund goes out automatically. Not a single decision is needed that wasn't defined as a rule beforehand.

Limit: a workflow can't handle deviations. When something unexpected comes in, it stops or does the wrong thing. So every good workflow needs a branch: special case detected, human notified.

Chatbot: conversation with limits

A chatbot is software that talks to people in natural language, usually in a chat window or by email. Modern chatbots use language models, meaning AI that understands and produces text. A chatbot holds conversations, and it stops there. It doesn't carry out multi-step work.

Example 1: an FAQ bot on your product page. It answers "is this brush suitable for children?" from your knowledge base, meaning a collection of vetted answers and documents.

Example 2, from our own operations: our automatic reply to "where is my parcel?". The AI understands the question, pulls tracking data and answers. That's a chatbot with a workflow attached behind it. How exactly that works is in Automating WISMO Requests.

Limit: a chatbot is only as good as its knowledge base and its connections. Without access to your order data it can only say general things. And in emotional situations, a real complaint for instance, it often makes things worse. A human belongs there.

AI agent: the goal instead of the steps

An AI agent gets a goal instead of a sequence of steps. It decides for itself which tools to use, checks its result and tries a different way when there's a problem. I wrote down the full definition in What Is an AI Agent?. The short version: a workflow gets steps, an agent gets goals.

Example 1: research tasks. "Find out which of our 20 suppliers have announced price increases, and summarize it." The agent reads emails, searches documents, and decides for itself where to look.

Example 2, from our own operations: draft replies in support at nano. The AI reads a complex customer email, pulls order data, checks our policies and drafts a reply. It decides for itself which information it needs, depending on the email. We approve it, and the complete setup is in Pre-Sorting Email Support With AI.

Limit: an agent isn't deterministic, so the same input doesn't guarantee the same result. It makes mistakes, and more creative ones than a workflow makes. So it belongs nowhere that a mistake directly costs money or trust, at least not without human approval.

The decision aid

As a rule of thumb in one sentence: take the workflow when you can write down the steps. Take the chatbot when people ask in natural language. Take the agent only when the steps genuinely differ from case to case.

And usually the answer is a combination. Our WISMO reply is a chatbot in front and a workflow behind. Our support setup is a workflow that calls an agent for drafts.

Three test questions for your specific problem

If the table isn't enough, ask yourself these three questions about your case.

First: can you explain the process to an intern in five sentences? If yes, it's a workflow. "When an order comes in, send it to the fulfiller, wait for tracking, enter it, inform the customer." Fully describable means fully automatable, with no AI at all.

Second: does the problem start with a question in human language? Then you need a chatbot in front, whatever happens behind it. The craft is in the connections: the bot has to reach your real data, otherwise it produces polite non-answers.

Third: would you send the result out unchecked? If no, plan the approval in from the start. That's the right way to build it, and no flaw in the agent. At our place a human reads every draft, and we still save most of the writing time.

And one observation from two years of client projects: the distribution is lopsided. Broadly speaking, most worthwhile automations in e-commerce are plain workflows. Chatbots pay off in a few places, but very visible ones. Genuine agent cases are the exception. The industry's marketing tells you the opposite, because "agent" sells better than "chain of rules".

The most common bad buys

Four patterns I see over and over in conversations.

Bad buy 1: bought an agent, needed a workflow. The classic. Order processing, filing invoices, stock reconciliation: those are fixed tracks. An agent doesn't make them better, only more expensive and less predictable. Anyone selling you an "autonomous agent" for a stable standard process is selling you the buzzword.

Bad buy 2: chatbot with no connections. A bot that doesn't know your order data can't answer the most common customer question and creates more frustration than before. Data connection first, then the chat window.

Bad buy 3: full autonomy without approval. An agent that triggers refunds or sends final replies to customers without human review will eventually get something expensive wrong. Anything touching money or trust needs a human in the loop.

Bad buy 4: AI as a substitute for broken processes. If your returns process is chaotic, you'll automate chaos. Understand and clean up the process first, then pick the tool. The order in which to tackle your processes is in the overview of e-commerce processes.

Conclusion

AI agent, chatbot and workflow are three tools for three different problems. The workflow for fixed tracks, the chatbot for conversation, the agent for tasks with changing steps. The most expensive tool is rarely the right one. Across our own brands, plain workflows do most of the work, and the agents only work where their flexibility is genuinely needed.

If you're unsure what your specific case needs: talk to us at Flowhouse. We'll also tell you when a simple workflow is enough. Especially then.