When to use it
- You want to understand why one of your contents was or wasn't used in a conversation
- You want to know who answers the customer: a flow, an agent or your knowledge base
- You need to understand the effect of the project's settings on routing
- A good piece of content exists in the base, but ClaudIA never reached it
Prerequisites
- Access to the project's Painel da ClaudIA
About this article
Not every customer message reaches your contents. Before that, the conversation goes through a sequence of checks that decides who will handle it. This article explains that entire path.
To write and maintain contents, see the companion article: Complete guide to ClaudIA contents.
The complete path
Order matters. Each check only happens if the previous one didn't divert the conversation. A check that diverts ends the path right there — the following ones are never even evaluated.
How the conversation starts
The first message is a special case: the project can be configured to start in three different ways.
| Start setting | What happens on the first message |
|---|---|
| No start setting | The first message follows the same path as all the others. This is the default behavior, and the most common one. |
| Start with a flow | Every conversation starts with a defined flow — welcome, triage, customer identification. Only after that flow ends does the rest of the path come into play. |
| Start with an agent | The first message goes straight to the questions agent, without passing through the Judge. The entry path is already defined, so there is nothing to route. |
From the second message onward, all of them follow the same sequence of checks.
Is the conversation in the middle of a flow?
A flow is a multi-step script: it asks a question, waits for the customer's answer, then asks the next one. Between one step and the next, the conversation is in the middle of the flow — and that is exactly what this check asks.
Where an ongoing flow comes from
| Source | When it happens |
|---|---|
| Start setting | The project is configured to start every conversation with a flow |
| A controlled flow content | ClaudIA used a content of this type — and each one of them points to the flow you defined when registering it |
| An agent | The agent decided to trigger a flow to solve the case |
The second row is the one that matters most to whoever writes content: it is your controlled flow content that creates the "ongoing flow" of the next message.
While the flow hasn't ended, it is in charge of the conversation. None of the following checks happen — neither the Judge nor the search across contents. This is what ensures that a multi-step process isn't interrupted halfway through by an automatic answer.
Did the customer ask for a human agent?
That request doesn't always lead to the same place — the project defines what happens.
| Setting | What ClaudIA does |
|---|---|
| Direct transfer | Transfers right away, without trying to solve |
| Retention attempt | Before transferring, it replies offering help — if the customer accepts, the conversation continues with ClaudIA |
| Attempt limit | After a defined number of retention attempts, it transfers anyway |
Retention depends on good content. When ClaudIA tries to retain, it needs to show that it can solve the issue — and it is your knowledge base that supports that. A weak base turns the retention attempt into nothing more than a delay before the transfer, which worsens the experience instead of improving it.
During a flow, this request is not detected. If the customer asks for a human agent in the middle of a flow, ClaudIA does not evaluate that request — the flow stays in charge. What can transfer there is the flow itself, through the steps it has. That's why it's worth planning an exit to a human agent inside the longer flows.
Does the project have agents turned on?
If the project does not have agents, the path ends here: every message goes to the search across contents. This is the default behavior.
If it does, the Judge comes into play.
What the Judge is
The Judge reads each message and decides who should handle it: one of the available agents, or the search across contents.
It only chooses among the agents that exist in that project, and the question it asks is whether any of them solves that customer's demand. If none of them does — or if it makes a mistake and points to an agent that doesn't exist — the conversation goes to the contents.
When the Judge isn't even consulted
In two situations the decision has already been made beforehand, and the message goes straight to the agent:
| Situation | Example |
|---|---|
| The Judge is turned off in the project | An operation in which almost everything is a system action — reissuing a bill, order status, updating registration data. Since every demand will end up with an agent, no routing decision is spent on each message. |
| A flow has just handed the conversation back | The project is handled by the agent, but uses an identification flow at the entry point. The customer sends "hi", the flow asks for the ID number, validates it and hands the conversation back — the next message goes straight to the agent, because who handles it was already defined. |
These are uncommon settings. In most projects with agents turned on, every message goes through the Judge.
The three destinations
| Destination | What it is |
|---|---|
| Flow | A predefined script, with fixed steps. While it is running, nothing else takes over the conversation. |
| Agent | A specialist that decides what to do based on the tools it has available and the instructions it received — querying systems, triggering flows, executing actions. The Supervisor chooses which agent handles each case. |
| Content | The search across your knowledge base. It is the default path and also the fallback one — whenever there is no agent that solves the case, the conversation comes back here. |
When in doubt, the system always prefers the contents. If the Judge doesn't find an agent suited to the demand, if it makes a mistake, or if it fails, the conversation goes back to the knowledge base. That's why the quality of your contents remains the foundation of the service even in projects with agents.
Agent and content are different things. The agent solves with tools and instructions — it does not query your knowledge base. It is the content that answers every time the demand is not an agent's case, which in practice is most of the questions.
Once it reaches the content, the type defines the outcome
After the conversation reaches the search, the type of the content found decides what happens next:
| Type | What happens |
|---|---|
| N1 | The answer is sent and the service continues with ClaudIA |
| N2 | The answer is sent and the service is transferred to a human agent |
| Controlled flow | Triggers the flow defined for that content — and the conversation is now "in the middle of a flow" |
N1 does not mean "ends the service". It means ClaudIA remains in charge of the conversation.
Summary of the settings that change the path
| Setting | Effect on routing |
|---|---|
| No start setting | The first message follows the same path as the others. Default. |
| Start with a flow | Every conversation starts with a defined flow |
| Start with an agent | The first message goes straight to the questions agent |
| Agents turned off | Every message goes to the contents. Default. |
| Agents turned on | The Judge starts evaluating each message and decides between agent and content |
| Judge turned off | Every message goes straight to the agents, without prior evaluation |
| Direct transfer | A request for a human agent transfers right away |
| Retention attempt | A request for a human agent gets an offer of help first |
| Attempt limit | After N retention attempts, it transfers anyway |
If you're unsure which setting your project uses, confirm it with your team at Cloud Humans.
Notes
- To write and maintain contents: Complete guide to ClaudIA contents
- About the agents and the Supervisor: What Specialist Agents are and how the Supervisor decides who answers
- About the classification of the service: How a service is classified as N1 or N2
- About the validation of ambiguous content: How the Classifier and Clarifier work