Guide
Chatbot or AI agent?
Both appear as a conversation. The difference is not in the interface, but in what happens after the answer.
The same example, handled two ways
Take a simple request: "update this customer record with the new address".
An answer-only assistant, in this example, explains where to click and may cite the documentation. The user then has to do it themselves.
An agent proposes the change, carries it out once authorised, and reports what actually changed. The record is updated, or the agent explains why it was not.
The labels chatbot and agent are not enough to distinguish products. In this example, compare connected features and the outcome.
What acting additionally requires
When an assistant can act, its integration must define permissions, confirmations and outcome verification.
An identity and permissions
On whose behalf is the action performed? A chatbot that gets it wrong gives a bad answer. An agent that gets it wrong modifies the wrong data.
A moment of approval
What proceeds directly, what requires confirmation? That boundary must be explicit and configurable, not decided by the model.
Proof of effect
Did the action succeed? The answer comes from the application, not from the agent's conviction. Without it, every report becomes doubtful.
Handling failure
Refusal, partial failure and uncertain state: the report must distinguish completed work from what still needs attention.
Three levels, not two categories
In practice the boundary is gradual. These three levels are set separately, capability by capability.
- 01
Answer
The agent reads, explains, finds. No effect on data. That is already useful, and it is the lowest-risk level to start with.
- 02
Propose
The agent prepares a change and presents it. The user decides. The intellectual work is done, responsibility stays human.
- 03
Execute
The agent performs the authorised action. This level requires everything above: permissions, proof of effect, failure handling.
How to choose?
Is a chatbot sometimes enough?
Yes. If your need is answering frequent questions without touching data, an agent that can act adds complexity without benefit.
What is the extra cost of an agent?
It does not come from the model but from integration: describing capabilities, respecting permissions, verifying effects. That is application engineering work, not configuration.
Can we start with answering and evolve?
That is what we recommend. The levels are set separately: starting with reading lets you observe real usage before opening execution.