When to use a chatbot, workflow, or agent

Not every AI feature should be a chatbot. Obvious: right up until every prototype starts with a blank text box, because that is the fastest thing to build and the most impressive thing to demo.

The choice between chatbot, workflow, and agent is less a sophistication ladder than a decision about how much the system is allowed to do on its own.

The two common mismatches

A chatbot bolted onto a job that needed a workflow. The user knows exactly what they want; the system makes them describe it in prose, guesses at the parameters, and produces something adjacent. A form with four fields would have been faster, more accurate, and auditable. Conversation is a wonderful interface for exploration and a poor one for a known process.

An agent given autonomy the task never required. The steps were fixed and knowable. Now they are decided at runtime, which means they vary, cost more, take longer, and cannot be traced without effort. You get all of the complexity and none of the benefit.

The blank text box is a decision, not a default

Once a chatbot becomes the front door, narrowing it later is genuinely hard. Users have learned they can ask anything, and every unhandled request now reads as a failure rather than as out of scope.

Starting narrow and widening is easy. Starting wide and narrowing is a product regression, and it will be experienced as one.

Choose against operating cost, not demo quality

Run the use case through each pattern and compare the things you will live with: cost per interaction, latency, how you would debug it, and who owns it when it misbehaves.

The pattern that demos best is frequently the most expensive to operate, and the gap does not appear until you are already committed.

Then write one sentence: what is this system allowed to do on its own? If that sentence is hard to write, the pattern choice has not actually been made yet.

Choose the simplest architecture that reliably solves the problem. The interface is part of the architecture, and the wrong wrapper makes a perfectly good model feel unreliable.

Is your current use case truly a conversation, a workflow, or bounded autonomy?