A chatbot and an LLM agent are different tools. How to tell which one your business needs

The word “bot” now means three very different things, and the confusion costs money: some companies buy an autonomous agent where a button menu would have done; others try to solve with buttons a problem that requires understanding the data.

Three levels — three different jobs

A button-menu bot is a menu in Telegram: tap “Stock” and get a file, tap “Request” and fill in a form. There is no model inside; the scenarios are hard-wired. It is cheap and reliable, and for jobs like “take a request” or “send the price list” nothing more is needed.

A bot with an LLM layer understands a question asked in plain words and turns it into a query against live data. “How many of our orders are stuck in assembly?” — and the answer comes back as figures from the CRM, not a 300-line file. That is exactly how the “Production PM” works at a furniture factory: the owner asks from a phone the way you would ask a person, and the system answers. The key difference from buttons: the questions don’t have to be anticipated in advance.

An autonomous LLM agent doesn’t wait for a question. On a schedule, it goes through the data itself, decides for itself what matters, and assigns tasks to people itself: found an overdue payment — a task for the sales rep; spotted a stalled deal — a reminder; pieced the picture together — a report for the manager. The human enters this loop at the end, not at the beginning.

How to choose the level

The working question is: who initiates?

• If the initiator is always a human and the questions are predictable, a button-menu bot is enough.
• If the initiator is a human but the questions change every day, you need an LLM layer on top of the data.
• If the problem is that people don’t even have time to ask, you need an agent that asks itself.

The second question is the cost of a mistake. A button-menu bot can’t make one by definition: it shows what is there. An LLM layer may misread a phrasing — which is why it answers with figures and links that are easy to check. An agent acts on its own — which is why its actions are constrained architecturally: read access to the accounting system, write access only to the CRM and only for approved operation types.

About the money

The levels differ in cost several times over, and that difference is an honest one: a button-menu bot is weeks of work; an agent is months, discovery and pilot included. Not starting at the top level is a perfectly sound strategy: the “Auditor” at a furniture company began as a set of hard-coded daily CRM checks with no model at all — and in that form alone it cleared the “dead” deals out of the pipeline. The LLM layer gets added once hard rules stop being enough.

To see all three levels at work: the “Auditor” — rule-based checks, the “Production PM” — questions in plain words, the AI Sales Employee — an autonomous agent.