Chat, assistant and agent are three levels of delegation to AI. In a chat, you ask and it forgets. An assistant has learned one of your processes and helps when you call it. An agent decides on its own when to act and where to leave the result. What picks the level is how much you already trust what it delivers and what a mistake costs, not how sophisticated the task is: the most important process in our company, the project quote, runs entirely at the assistant level, and the contract that used to take up to 40 minutes to put together now takes 5.
What an AI chatbot, an AI assistant and an AI agent are#
Chat, assistant and agent are three ways of using the same artificial intelligence, and what changes between them is how much of the work you hand over to it. The tool can be the same in all three cases. ChatGPT, Claude and Gemini work as a chat, can become an assistant and serve as the foundation for an agent.
The easiest way to understand it is to picture someone who just joined the company.
- A chat is a hallway conversation. You ask, the person answers, and tomorrow they do not remember the topic. It is good for research, a one-off question, a piece of text you need today and will not need again.
- An assistant is the intern who was handed the manual. You explained how a task is done at your company, gave them the templates and the examples, and when you need it you call them, they help, and the work comes back to you. You are still the one who decides when to start and when to finish.
- An agent is the employee who was given a role. They have access to the tools (the CRM, the ERP, the task system), know what needs to happen, and decide on their own when to act, what to do and where to leave the result.
| Level | What you hand the AI | Who decides when to act | Needs access to systems | Example from our operation |
|---|---|---|---|---|
| Chat | a question | you, every time | no | quick research on a topic that came up in a meeting |
| Assistant | a written process, with templates and examples | you, when you call it | sometimes, to read documents | putting the contract together from the client's details |
| Agent | a role, with the expected result | the AI itself, on a schedule or a trigger | yes, with defined permissions | the hours report that goes out every Monday without anyone asking |
Anthropic, the company that develops Claude, draws a similar distinction in a technical guide published in December 2024, Building Effective AI Agents: it separates workflows, where the path is defined in advance, from agents, where the model directs its own process and chooses the tools it uses. The guide's advice is the same thing we learned in practice: start with the simplest solution and only add complexity when it is not enough.
What decides the level of a task#
What decides the level is not how sophisticated the task is. It is how much you already trust what the AI delivers on that task, and what a mistake costs.
With a new hire it works like this. In the first month, you review everything they do, because you do not know yet whether you can trust them. By the sixth month, you stop by their desk every now and then and ask how it is going. After two years, nobody asks what they are doing anymore. Trust was built over time and through results.
AI follows the same curve, and the level of use tracks that curve:
- A task you do now and then, with no fixed shape: chat. It is not worth teaching anything, because the task does not repeat the same way.
- A task that repeats but involves a critical part of the operation: assistant. You want help doing it faster, and you want to keep looking at every result before it moves on.
- A task that repeats, where you already trust the result and a mistake is internal and reversible: agent. If it gets something wrong one day, you notice, fix it, and nothing has left the company.
The last one is usually the sticking point. A rule we use at Epicora helps decide: what is internal and can be undone, AI can do on its own; what leaves the company goes through a person first. An email to a client, a message to a supplier, a post, a payment: an agent can prepare any of these, but someone approves it before it goes out.
Why most tasks do not need an agent#
The agent is the level that shows up most in videos and ads, and the one the fewest tasks call for. Most of what a company gains from AI comes from the assistant level.
The clearest example we have is the project quote, the most important process in our company, because it is where the client decides whether to work with us. There is no agent in it. Everything runs on assistants, and a person leads every step:
- The meeting with the client is recorded and transcribed by the video call tool itself. We did not build anything for that.
- The transcript, the WhatsApp messages, the emails and any material the client sent go into a project folder.
- An assistant reads that material following our scoping methodology and helps me understand the company's problem and how other companies in the same industry solve it.
- Other assistants help write and review the scope, lay out the prototype screens and estimate the hours for each module.
The result was a change in the order of the process. Before, the detailed scope and the prototype came after the contract was signed, in an ideation phase that took weeks, and on larger projects took months. The client was already paying and got a 30 to 40 page document to read and approve. Today they get the full scope and the screen-by-screen prototype along with the quote, before they know the price. That whole phase stopped existing.
None of this required AI autonomy. It required a well-written process and a person who keeps making the decisions.
How an assistant changes an everyday task#
The assistant is where most companies should start, and two examples from our routine show why.
The contract. Putting a contract together took up to 40 minutes: gathering the client's details, typing everything into the template and checking it. The person in charge of our finance built an assistant in Claude that does it differently. He picks the contract template and hands it the company registration card (CNPJ, the Brazilian business ID), or the personal documents when the contract is with an individual, along with the scope already attached in the CRM. The assistant extracts the information, confirms each item with him, generates the contract and hands it back for review. Two or three more clicks and the contract goes out for signature. The time dropped from up to 40 minutes to 5, counted from the moment the client's details are in hand.
The development team's tasks. The scope is written in the client's language, and the tasks need to be in the language of whoever writes the code. Turning one into the other took hours on every new project. Two people on the team built an assistant that reads the scope and helps write the tasks in our standard format. It only worked because both sides were already standardized: the scope always follows the same structure, and so do the tasks.
Both examples follow the same recipe, and it works for any company:
- Pick a task that already exists, repeats every week and is not what sets the company apart.
- Write down the steps as if for a new hire, with a template and an example.
- Test it with AI at the assistant level and review everything in the first few weeks.
When an agent is worth it#
An agent is worth it when the task happens on its own, on the calendar or on a trigger, when you already know what a good result looks like, and when a mistake stays in-house.
In our operation, one example is the hours report. Every Monday, an agent opens the task system, reads the hours each person logged that week and compares them with the priorities the company set. If, out of 800 development hours in the week, 300 went to a project that delivers four months from now and only 30 to one that delivers in three weeks, it flags that. Nobody asks for the report. What the agent does is show whether the company's priorities are reaching day-to-day work, and deciding what to adjust stays with management.
Another example, this one at a client, is the agent that handles first contact on Agência OKSE's WhatsApp. It replies, triages and hands off to the team. It was designed with limits that do not depend on good intentions: it only replies when someone writes in, never starts a conversation, and says it is a virtual assistant when asked.
What the two have in common is what makes an agent work: a clear role, access only to what it needs, and a result someone can check.
How to measure whether an agent is working#
Every agent needs a measure of results defined before it starts running, written in terms of your work, not the AI's. Without it, an agent can run for months looking useful.
That is what we did with our first agent. Every day, before the team's morning meeting, it scans the official sources on artificial intelligence (the labs, the companies that build the models, whoever publishes first-hand), cross-references each piece of news with the projects in progress and sends me an email with what might be useful. One morning in September, for example, it read 17 sources, found 12 new items and let 5 through.
Before switching it on, we set the measure: if within 30 days it brought no information I actually applied to a project, it was not worth keeping, and it would be just one more email to read every day. In the first week it had already brought something we applied. With the measure set in advance, whether the agent is worth it is a question you can answer with facts.
Not measuring is a common problem. The third edition of Panorama da IA no Brasil, a Brazilian survey by Zappts published in September 2026 with 167 leaders of large companies, found that 55.1% of Brazilian companies have no structured mapping of the return on AI. The year before, it was 30.6%. Companies are using more AI and measuring less of what it gives back.
That is why, before building any assistant or agent, it is worth answering three questions:
- What process will the AI do, or help do?
- What result do you expect to see change in that process?
- How will you measure whether it worked?
If the company does not measure anything yet, its first AI project can be exactly that: starting to measure.
Frequently asked questions#
What is the difference between an AI chatbot, an AI assistant and an AI agent?#
The difference is how much you delegate. In a chat, you ask, it answers, and in the next conversation it has no idea what was said. An assistant has already been given your process in writing (the steps, the templates, the examples) and helps when you call it, but you are the one who decides when to start and when to stop. An agent is given a role: it has access to the company's tools, such as the CRM or the task management system, and decides on its own when to act, what to do and where to leave the result.
Where should a company start with AI?#
With a task that already exists, repeats every week and is not what sets the company apart. Write down its steps as if for a new hire and test it with AI at the assistant level. Before you start, answer three questions: what the process is, what result you expect and how you will measure it. Without the third one, there is no way to tell whether the AI helped.
When is it worth turning an assistant into an agent?#
When you already trust what the assistant delivers enough to stop reviewing everything, when the task happens often enough to justify running on its own, and when a mistake is internal and can be undone. If the task touches a client, money or anything that leaves the company, an agent can still prepare the work, but someone approves it before it goes out.
Do I need a developer to build an AI assistant?#
For an assistant, usually not. The paid plans of ChatGPT, Claude and Gemini let you build an assistant with your own instructions and documents, and the real work is writing the process down clearly. An agent, which needs to connect to the ERP, the CRM or the task system and act on its own, usually requires integration, permission rules and a way to track what it did, and that is where development work comes in.
How do I measure whether an AI agent is delivering results?#
Define up front what counts as a result, in terms of your work, not the AI's. For our news agent, the measure is simple: if within 30 days it brought no information I actually applied to a project, it was not worth keeping. It delivered in the first week. Without a measure like that, an agent can run for months looking useful.
Sources#
- Anthropic, "Building Effective AI Agents", published on December 19, 2024, with the distinction between predefined-path workflows and agents, and the recommendation to start with the simplest solution — https://www.anthropic.com/engineering/building-effective-agents
- Zappts, "Panorama da IA no Brasil", 3rd edition, a Brazilian survey published in September 2026, with 167 leaders of large companies interviewed between April and July 2026 (margin of error of 7.6 points) — https://zappts.com.br/panorama-da-ia-no-brasil-o-que-esperar-da-ia-nas-empresas-brasileiras-em-2026/
Next step#
If you already know which task you want to hand over to AI and it needs to connect to the company's systems, that is the work we do in AI and automation, from the support agent at Agência OKSE to the technical report in Sistema Cuidar. When the process has consequences and the AI needs to follow the company's method without improvising, the subject continues in how to make AI follow your company's method.