# Using AI attributes | Conversation Intelligence

> 

Source: https://help.kustomer.com/en_us/using-ai-attributes-conversation-intelligence-S1Be9adtzx

Last updated: 2026-10-16T17:18:24.922Z

AI attributes turn unstructured information into structured data taken from a conversation's context using Conversation Intelligence. See [Understanding Conversation Intelligence](https://help.kustomer.com/en_us/understanding-conversation-intelligence-r1aBEjuYzl) for more details on how this feature powers insights, reporting, and the actions you take from that data.

AI attributes come in three different types:

*   **Classifier** — Sorts conversations into labels that you define.
*   **Score** — Evaluates a conversation against a predefined, weighted rubric and returns a numeric score with a rationale.
*   **Extractor** — Extracts or summarizes information based on natural-language instructions.

The type you choose depends on what you want to accomplish. For example, you might use a Score AI attribute to measure interaction quality, a Classifier to identify contact reasons, or an Extractor to capture specific information from a conversation.

### Add a Conversation Intelligence attribute

You can create AI attributes from the **Kustomer AI** page.

1.  From the **Kustomer AI** page, click **AI Attributes**.
2.  In the upper-right corner of the AI Attributes page, click **Add AI Attribute**.
3.  Select the type of attribute you want to create (Classifier, Score, or Extractor).
4.  Enter a name and complete the required configuration for that attribute type.
5.  Save the attribute.
6.  Toggle the attribute on when you're ready to begin using it.

> Note: It is possible to add up to 3 triggers to an attribute, and each trigger creates an execution.

The configuration options change according to the chosen attribute type.

#### Add AI instructions

AI instructions provide additional guidance about how the AI should evaluate or extract information for each attribute. Write instructions in natural language and be specific about the behavior you expect. For example, instructions can tell the AI which signals to consider, what information to ignore, or how it should interpret ambiguous situations.

Here are some example instructions for each attribute type.

#### Sample instructions for Classifiers

_Classify the conversation based on the customer's primary reason for contacting support. Consider the customer's intent and the overall context of the conversation, not individual keywords alone._

_Select the label whose description most closely matches the customer's reason for contacting us. When more than one topic appears, prioritize the issue that best represents the customer's main request or the issue that required support action._

#### Sample instructions for Scoring

_Evaluate the overall quality of the conversation using the criteria defined below. Consider the full conversation and use only information that is supported by the interaction._

_Score each criterion according to its individual instructions. Consider whether the customer's request was understood, handled appropriately, and resolved when determining your evaluations. Do not assume an outcome that is not demonstrated by the conversation._

#### Sample instructions for Extractors

_Identify and extract information related to_ **_PRODUCT A_** _from the conversation._

_Return the value only when the conversation provides enough information to identify it. Use the customer's actual meaning rather than relying only on an exact keyword match. Do not infer or invent a value when the information is not present or is ambiguous._

For an Extractor, your instructions describe the information you want the AI to find. For example:

> Identify the primary product the customer is contacting us about. Return only the product name.

Clear, focused instructions generally make the output easier to use consistently in workflows and reporting.

#### Add options to a Classifier

Classifier attributes use **options** to define the categories the AI can choose from.

For each option, provide a clear name and description. The description gives the AI additional context about when that option should be selected.

For example, a **Contact reason** classifier might contain:

*   **Where is my order?** — The customer wants an update on the location or delivery status of an existing order.
*   **Return or exchange** — The customer wants to return or exchange a purchased item.
*   **Cancel order** — The customer wants to cancel an order that has already been placed.
*   **Product question** — The customer is asking for information about a product before or after purchasing it.

Try to make your options distinct from one another. If two options describe very similar situations, the AI may have less information to determine which one is the best match.

Depending on the classifier you create, you can configure the AI to return a single option, multiple options, or a hierarchical category.

#### Add criteria to a Score

Score attributes use **criteria** to define what the AI should evaluate.

For each criterion, provide a name, description, weight, and rating categories with associated values. The AI selects a rating for each applicable criterion based on the available context. These ratings are combined into a weighted score from **0 to 100**, along with a rationale explaining the result.

For example, an **AI-inferred satisfaction** score could include criteria such as:

*   **Resolution** — Was the customer's problem resolved?
*   **Customer sentiment** — Did the customer's sentiment improve, remain neutral, or worsen during the interaction?
*   **Tone** — Was the communication professional, empathetic, and clear?
*   **Efficiency** — Did the interaction progress toward a solution without unnecessary back-and-forth?

Use each criterion's weight to reflect its relative importance in the final score. For example, you might give resolution more weight than tone. More important criteria should contribute more heavily to the overall result.

An AI-inferred satisfaction score estimates satisfaction from the available evidence. It is a _proxy for customer satisfaction_, rather than a rating submitted by the customer.

The rationale explains the selected ratings and provides context for reviewing the score, helping you identify interactions that may need closer attention.

### Edit an AI attribute

You can update an AI attribute when you need to refine how it evaluates your customer interactions.

1.  From the **Kustomer AI** page, click **AI Attributes**.
2.  Find the AI attribute you want to update.
3.  Click the **Edit** icon.
4.  Update the available settings.
5.  Save your changes.

You can refine settings such as the attribute name, instructions, classifier options, score criteria, and when the attribute runs.

Some properties that determine how an AI attribute stores its output can't be changed after the attribute is created. If you need to change the fundamental attribute type or output format, create a new AI attribute instead.

### Turn a Conversation Intelligence attribute on or off

Use the toggle on the **AI Attributes** page to control whether an AI attribute is active.

Turn an attribute off when you want to keep its configuration but temporarily stop it from evaluating new interactions. You can turn it back on when you're ready to use it again.
