AI agent, chatbot and language model
The three are often treated as synonyms, but they aren't. The language model is the engine; the chatbot is a script, with or without that engine; the agent is the engine plus a goal, tools and rules.
| Criterion | Language model | Flow-based chatbot | AI agent |
|---|---|---|---|
| What it is | Software that generates text from a prompt | A tree of menus and replies designed by someone | A language model with a goal, tools and rules |
| How it decides what to do | It doesn't decide: it responds to the prompt it received | Follows the path laid out for each option | Picks the next action based on the conversation, within what's allowed |
| Acts in other systems | Not on its own | Only through the flow's fixed integrations | Yes: reads from and writes to the CRM, calendar, catalog |
| Where it fails | Makes up an answer when it doesn't know | Gets stuck when the conversation goes off script | Can take the wrong action if the limits aren't written down |
| Sales example | Drafting a proposal from notes | Triage menu: press 1 for a quote | Responding to the lead, qualifying them and booking the meeting in the CRM |
How an agent works under the hood
Every agent, from any vendor, is built from the same parts. When one of them is missing, that's usually where the project breaks:
- Goal: what the agent needs to achieve — qualify a lead, confirm an order, answer a question
- Instructions and rules: what it can and can't do, written by the company
- Tools: the systems it reads from and writes to, such as the CRM, calendar and product catalog
- Context: the contact's history and the knowledge base it can use to answer
- Exit point: the rule that hands the conversation to a person when the case is out of scope
Examples in sales and customer service
In sales, an agent makes sense for high-volume tasks with clear rules — exactly the ones that tend to get pushed aside when the team is busy:
- Responding to the lead who writes at night or on the weekend, instead of leaving them waiting until the next morning
- Asking the qualification questions the team already uses and saving the answers to the opportunity
- Offering open slots on the assigned rep's calendar and confirming the meeting
- Following up with people who stopped responding, taking into account what was already said in the conversation
- Separating quote requests, post-sale questions and complaints, and routing each one to whoever handles it
Limits: what an agent shouldn't do
An agent won't stop on its own when it faces a situation it shouldn't handle. That's why the limits are written before it goes live, not after the first problem:
- Promising a price, timeline or terms outside the written sales policy
- Answering questions about things that aren't in its knowledge base — the language model can make up an answer with complete confidence
- Handling complaints, collections or sensitive situations without a person
- Taking actions that are hard to undo, like canceling an order, without confirmation
- Collecting more personal data than necessary; customers should know they're talking to an automated assistant, and data handling follows the LGPD (Brazil's data protection law)
How to implement it carefully
The approach that works starts small: one task, one written criterion and one rule for handing off to a human. The agent is connected to the system where the data already lives — in sales, the CRM and the channel the customer uses, which in Brazil is almost always WhatsApp — and it's tuned in the first few weeks by reading real conversations, not by the impression from the demo.
The step-by-step applied to sales operations, with what an agent does and where it shouldn't act, is on the AI agents for business page. The detailed comparison with flow-based chatbots, including when a chatbot is still the better choice, is in the comparison between the two.