Fin Categories is currently in closed beta

If you’d like access, please contact Ellie at ellie.short@intercom.io with your workspace ID.

This feature is still evolving. Some features are not yet live - but we’re working on them. This article will be updated as new capabilities are released.

Already using AI Category Detection? Scroll down to learn how Fin Categories compares and how to transition.

Overview

Fin Categories enable Fin to understand and classify what a conversation is about in real time. This allows Fin to apply structured attribute values - like issue type, sentiment, or urgency - that power smarter triage, routing, and reporting.

Quick Start: Set up Fin Categories in 3 steps

Decide what to detect
Choose the structured attributes you want Fin to classify - like issue type, sentiment, urgency, or spam.
Create or convert a category
Go to Train > Categories. Define your attribute values with clear descriptions, or convert an existing list-formatted attribute.
Preview and enable
Use the preview tool to test Fin’s accuracy. Tweak your value descriptions, then enable the category for real conversations.

Once enabled, use the detected attribute values in escalation rules, workflows, and reporting.

How it works

With Fin Categories, Fin listens to every message in real time and assigns attribute values based on live context. For example, if a customer shows frustration or urgency, Fin can automatically detect that and trigger escalation.

Detected values are stored as conversation data attributes, which can be used to power your workflow branching and routing rules, as well as generate reporting insights.

Key benefits

Custom categories that reflect your business: Train Fin to detect attributes like issue type, urgency, sentiment, or spam status.
Real-time, adaptive detection: Fin continuously evaluates context and updates values as the conversation evolves.
Smart routing and escalation: Combine Fin Categories with workflows and escalation rules to route conversations to the right team at the right time.
Reporting-ready structure: All detected category data flows into Fin reports for deep insight—no manual tagging required.
Full transparency and control: View, validate, and override Fin’s categorization logic at any time.

How Fin applies categories

By default, when Fin is involved in a conversation, it will classify enabled categories at key moments:

When handing off to a teammate
When the customer expresses resolution (positive feedback)
When a customer becomes inactive

If you set up rule-based Escalation Guidance, Fin will re-evaluate the conversation after each message, enabling more dynamic, real-time classification.

How to set up Fin Categories

Step 1: Decide what to classify

Think about the types of structured information you want Fin to detect. Common examples include:

Issue type (e.g., Billing, Projects, Account Management)
Sentiment (Positive, Neutral, Negative)
Urgency (Urgent, High, Normal, Low)
Spam detection (Spam, Legitimate)

Step 2: Create a new category (or convert an existing attribute)

To create a new Fin Category:

Go to Train > Categories to get started.
Click New category.
Fill in the Name and Description for your category
Add Category Values (with clear descriptions for each).
Optionally:
1. Toggle Limit visibility if the category is only relevant for certain teams and you’d like to control who sees it in the Inbox.
2. Toggle Required attribute if in the event that Fin cannot detect a category value, you want to ensure that a teammate applies a value before closing the conversation.

To convert an existing attribute:

Go to Settings > Conversation data, click edit on a list-type attribute, and then click Convert to category
Once converted, the attribute will appear under Train > Categories.

Note: Once converted to a Fin Category, an attribute can't be reverted - but you can leave it disabled if needed.

Step 3: Preview before enabling

Before enabling a category, use the built-in preview in Train > Categories to:

Test category values against example customer messages
Check how accurately Fin applies the right value
Iterate on names and descriptions before enabling

[Optional] Step 4: Add subcategories (conditional categorization)

Subcategories let you connect categories together, so Fin can categorize more precisely. Fin will only check a subcategory when its parent category value is detected, giving you more accurate results for cleaner routing and reporting.

You can configure subcategories in the category settings side drawer under the Subcategories tab:

Define the parent category value.
Choose the subcategory that should be triggered when that parent value is detected.
Specify which subcategory values are allowed.

Examples:

If Issue type = Refund request → then detect Refund request reason.
If Sentiment = Negative sentiment → then detect Urgency.

When this logic is in place, Fin will first detect the parent category. If the defined value matches, Fin will then conditionally attempt to detect the linked subcategory.

Note: Subcategory logic is respected in rule-based Escalation Guidance. If a subcategory is referenced in a Rule-based Escalation Guidance, Fin will re-evaluate the conversation after each customer message to see if the parent and subcategory values match.

Using Fin Categories

In Escalation Guidance for intelligent handoff

Rule-based escalation guidance enables you to combine detected Fin categories with other user or company data to build precise rule-based logic that determines when Fin should hand off. Rule-based escalation guidance is evaluated and triggered deterministically, ensuring maximum reliability.

Rule-based escalation guidance work alongside Scenario-based escalation guidance (which works best for soft signals in natural language nuance).

To make it easy for you to configure routing and handover behavior when Fin escalates a conversation to the team, you can create branches in workflows based on which Escalation Guidance has been applied. This gives you full control of how Fin routes to the team for specific scenarios.

In workflows to define handover logic

Use the applied Fin Category values to define branches within your Fin workflow. For example, you can add a branching condition based on the Fin Category detected and route the conversation to a specific team or teammate.

In reports for conversation insights

Automatically filter and segment your Fin reports using the conversation attribute data from your Fin Categories to generate insights and analysis about conversation trends.

In conversations for manual QA and review

Teammates can see when a category has been detected by Fin in Analyze > Conversations when viewing conversation events in the thread or in the sidebar under "Conversation attributes".

Note: Ensure that you have "Show conversation events" enabled to see the conversation events in the thread.

Monitoring and reviewing Fin Categories

Once your categories are enabled, Fin provides real-time stats to help you understand how categories are being applied.

You’ll see:

Conversations → the number of conversations Fin has detected for each category and specific category value.
Resolved → the percentage of those conversations that Fin was able to fully resolve.
Routed → the percentage of conversations successfully routed using that category.

You can also drill down into individual conversations to review how Fin applied a category and validate accuracy. This enables you to monitor categorization accuracy, review miscategorized conversations and update category descriptions based on patterns.

Best practices for high-accuracy detection

To get the best performance from Fin when categorizing conversations, follow these best practices:

1. Create short, descriptive names

Keep category names under ~25 characters.
Use short, descriptive language that immediately signals the category value (e.g., “Login Issues”, “Billing Questions”).
Avoid internal jargon, ticket IDs, or ambiguous terms.

2. Write detailed descriptions

Aim for 100–300 characters.
Describe the types of conversations that belong in the category.
Include:
- Keywords or phrases the customer might use.
- Examples of questions or intents (e.g., “How do I reset my password?”).
- Clarifications on what doesn’t belong if needed.

3. Make categories distinct

Ensure there’s minimal overlap between category values.
Ask: “Would a human find it hard to choose between these?”
Avoid redundant or overly narrow categories.

4. Create an ‘Other’ category to avoid over-eager categorisation

Include a general-purpose ‘Other’ category to capture conversations that don’t clearly fit into your defined categories.
This helps prevent Fin from trying to find a match when none of the options are suitable.
Especially useful for categories that should only apply to a subset of conversations. (e.g., conversations about a specific product or use case).

By following these guidelines, you’ll help Fin interpret prompts more reliably, reduce confusion between categories, and increase your categorization success ratio.

Examples of good category descriptions

Example 1

Account Access 
This category covers conversations where customers cannot log into their account, have forgotten their password, or are locked out for security reasons. Customers often express urgency since they cannot use the product until access is restored. 

Applies if the customer:
- Says they cannot log in to their account. 
- Reports a forgotten or incorrect password.
- Mentions being locked out or account suspended.

Does not apply if the customer:
- Is asking about subscription or billing without mentioning login.
- Is requesting account deletion or GDPR/Privacy help.
- Is reporting a technical bug inside the app after logging in.

Likely Keywords: login, password, locked out, sign in, access denied

Example 2

Refund Requests
Covers conversations where customers explicitly request a refund, mention overcharges, or ask about cancelling payments. This includes disputes over being incorrectly charged or needing money returned.

Applies if the customer:
- Requests a refund directly (“I want my money back”).
- Mentions being charged incorrectly (“I was charged twice”).
- Asks about refund status or refund timelines.

Does not apply if the customer:
- Is reporting a failed payment or declined card.
- Is asking about subscription cancellation but not refunds.
- Is asking about financial hardship or restructuring options.

Likely Keywords: refund, charged, cancel payment, overcharged, money back, refund status

💡 Pro tip

Try passing your category names and descriptions to a writing tool such as Claude AI or ChatGPT to define them clearly.

Example prompt: Write comprehensive descriptions for all of the category values listed - Include all relevant details about what belongs in the category. Think about every type of conversation that should fall under this category and describe them in the description. Providing a detailed description will help our AI Agent categorize support conversations correctly. Include keywords and examples of customer questions.

Example use cases

Below are some examples of ways that customers are using Fin Categories to categorize conversations.

Issue Type category example:

Projects - Projects are a collection of related tasks and activities aimed at achieving a specific objective or deliverable which can involve teammate collaboration, time tracking, milestones or goals, and status.
Billing - Billing encompasses managing subscription plans, invoices, payment methods, discounts, plan features, trials, account restrictions, refunds, and more for a seamless billing experience.
Account Management - Account Management covers discussions related to user accounts, including account creation, deletion, updating personal and payment information, and more.

Sentiment category example:

Positive - A positive sentiment means the user who wrote the message seems generally happy or satisfied and is probably feeling a positive emotion.
Negative - A negative sentiment means the user who wrote the message seems generally unhappy or dissatisfied and is probably feeling a negative emotion.
Neutral - A neutral sentiment means the user who wrote the message seems to be neither happy nor unhappy and it is difficult to guess their emotion.

Spam Detection category example:

Spam - Automated spam that is sent to the customer support agents. This includes auto-responders, newsletters, guest posts, and other general spam messages that could be ignored by the CS analyst.
Legit - Legitimate conversations in which the user has an actual issue that should be handled by a customer support analyst.

FAQs

What does Fin use to detect categories?

Fin uses the category name, its description, and the value names and descriptions when evaluating which category value to apply. Ensure that all of these fields are written in a human-readable way that is easy for Fin to interpret.

What if categorization isn’t accurate?

We recommend checking that the naming and descriptions. Use our best practices and test with real customer messages in the preview tool.

Pro tip: To help review your category value labels and descriptions, you might like to try using an AI writing tool such as Chat GPT or Claude.

Example prompt:

You are an expert in customer-support AI. You are evaluating a taxonomy used by a LLM to classify customer support conversations. For each attribute (e.g., Topic, Sentiment), the LLM chooses the most appropriate category based on a combination of the category name and its description. This taxonomy will directly impact the LLM's ability to classify real customer support conversations. Your task is to assess the quality of this setup using the following best practices: Create clear, concise names - Choose short, descriptive names that immediately convey the category's purpose. Write comprehensive descriptions - Take the time to write detailed descriptions and include all relevant information about what belongs in the category. Think about every type of conversation that should fall under this category and describe them in the description. Providing a detailed description will help Fin categorize conversations correctly. You can include keywords and examples of customer questions. Make categories distinct - Avoid creating categories that overlap too much. Your categories should be clearly different from each other, making it easy to determine which one best fits a given situation. This should be checked within each attribute only. It's fine if a category overlaps with another category from a different attribute. It shouldn't affect the score. Overlap with categories in other attributes is allowed and does not affect this score. Ignore Archived Categories - If a category is marked as archived, do not evaluate or score it. Add 5 Columns to the CSV Clarity/Conciseness (1–5), Description Comprehensiveness (1–5), Category Distinction (1–5), Final Score, Comment Assess each parameter for each category, and write a comment of why you applied this rating. Then calculate the overall score for this setup. After you've done this add one more column: Overlapping Categories. If you think any given category overlaps with other categories - list these categories there.

What if Fin doesn’t detect a value?

Fin will return a null value if none of the categories apply, which will leave the category empty. You can include an "Other" option to catch unclassified conversations if preferred.

I’m using AI Category Detection - should I switch to Fin Categories?

If you’re using the older AI Category Detection [Beta] feature, you can continue to do so. However, Fin Categories is the newer, more advanced version designed to offer better accuracy, smarter routing, and improved reporting.

We recommend:

Requesting access to and using Fin Categories for all new AI-powered categorization use cases.
Planning to gradually move away from AI Category Detection, as this feature will be sunset in the future.

It’s possible to use both features in parallel for now, but all future improvements will focus on Fin Categories.

Want to switch? Reach out to Zoe Sinnott at zoe.sinnott@intercom.io to request access to the Fin Categories beta.

What happens when I convert an existing AI Category Detection attribute to a Fin Category?

When you convert an existing AI Category Detection conversation attribute into a Fin Category, Fin simply uses the same underlying conversation attribute. That means:

The converted attribute will continue to work in your existing workflows and reports without interruption.
When enabled, Fin will start categorizing attribute values at key moments, so you may see these attributes being updated twice at first (once by AI Category Detection, once by Fin).
To simplify, you can eventually remove your AI Category Detection workflow blocks once you’re happy with how Fin is applying categories.
Conversion is a one-way action - attributes cannot be reverted back to AI Category Detection, but they can be disabled if needed.

Do I have to involve Fin in every conversation to use Fin Categories?

Yes. Fin Categories are applied by Fin when it is present in a conversation. This means Fin must be included in your workflows in order to categorize conversations.

If you’d prefer Fin not to answer in certain cases, you can use rule-based Escalation Guidance. This allows Fin to categorise the conversation and then exit immediately, based on conversation or user attributes of your choice.

For example, you could:

Escalate all conversations where Channel = Email
Escalate all conversations where Category = Billing

This way, you get the benefits of consistent categorisation across all conversations, while maintaining full control over when Fin does or doesn’t respond.

Fin Categories [beta]