The difference between a chatbot and an AI agent isn't the technology but the responsibility: a chatbot answers, an agent carries out a multi-step task — it reads, compares, uses tools and prepares a result. The right level for your business depends on how costly a mistake is and how well you can describe the rules of the task. Below are the four levels we use to scope a project.
Level 1: answers from sources (chatbot)
The system receives a question and answers from the company's approved information. It changes nothing. The risk is limited to a wrong answer, which you reduce with good sources and a handover to a human when it can't find the answer. This is the starting level for information and support.
Level 2: assistant with confirmed actions
The assistant proposes an action — a booking, a quote request recorded in the CRM, a prepared email — and carries it out only after the user or a colleague confirms. Each action is limited to a specific tool, with minimal rights. Integrations and the need to identify the person already appear here, but the decision stays with a human.
Level 3: an agent that prepares a result
The agent goes through several steps on its own: it reads a request and the attached documents, consults the catalogue and contract terms, compares, flags what's missing and prepares a complete proposal. A human validates the result before it is sent. Typical examples: preparing quotes from requests received by email, checking a file against a list of requirements, a daily summary of sources to monitor.
At this level, three things matter, and we define them before writing a line of code:
- The permitted sources and what happens when they contradict each other (the agent shows both versions rather than choosing silently).
- The log: every step, every source consulted and every result is recorded, so validation takes a minute instead of redoing the whole job.
- Flagging gaps: a result that says “three pieces of information are missing” is more valuable than a complete one built on assumptions.
Level 4: autonomous execution within strict limits
The agent also carries out the final action without individual approval, but only within explicit limits: amounts, document types, clients, time windows. Anything outside the limits stops and goes to a human. It makes sense only for high-volume tasks with stable rules and a low cost of error — for example, classifying and routing incoming documents, or booking confirmations. It doesn't make sense for commercial, financial or legal decisions.
How to choose the level
Answer three questions for the task you want to automate:
- What does a mistake cost? A wrong answer about opening hours is corrected in one message; a quote with the wrong price sent to a client isn't. The higher the cost, the closer human approval must sit to the final action.
- How well can you write the rules? If a new colleague could do the task from a written procedure, the agent can do it at level 3 or 4. If “it depends on the case” and nobody can say why, stay at level 2 until the rules become clear.
- Do you have the sources and the access? An agent that prepares quotes needs the catalogue, prices and terms in a form it can read. Without them, the first step of the project is putting the data in order, not the agent.
An example of scoping
A distribution company receives quote requests by email every day, in different formats. Level 1 would only answer general product questions. Level 2 would extract products and quantities from the email and put them into a request to be checked. Level 3 would prepare the full quote — contract prices, availability, lead time — and mark the uncertain lines for review. Level 4 would send quotes on its own only for clients and products on an approved list, below a set amount, with the rest going to a human. The same task, four projects that differ in risk, cost and the preparation required.
What must exist in any agent project
- A written list of what it can do alone, what requires approval and what is forbidden.
- A complete log of the steps and sources, accessible to the person responsible.
- A set of representative cases — including difficult and contradictory ones — on which the agent is evaluated before launch and after every change.
- A simple way to stop the agent and return to the manual process.
The context is worth mentioning: according to Eurostat, in 2024 more than 13% of EU enterprises used AI technologies, and Romania had the lowest share, 3%. The gap isn't closed by jumping straight to level 4, but by choosing a well-described task and a level of risk you can control.
See what a project includes in AI agents and, for levels 1–2, in Chatbots & virtual assistants.