Automatically triaging tickets based on intent, language, and sentiment

Attention: You must have the Advanced AI add-on to use the features described in this article.

Intelligent triage uses artificial intelligence (AI) to enrich your tickets with actionable information without the help of agents or admins. Specifically, intelligent triage uses AI to predict the intent, language, and customer sentiment for new tickets.

Intelligent triage enables the Zendesk ticket fields for intent, language, and sentiment predictions and confidence, which you can then use for setting up views for teams, automating with triggers or automations, and enhancing Explore reports.

When intelligent triage is enabled, Intelligence in the context panel will automatically be available for users in the Agent Workspace. This panel shows users the ticket's predicted intent, language, and sentiment, along with suggested macros. For more information, see Using Intelligence in the context panel.

This article contains the following topics:

• Requirements for using intelligent triage
• Understanding intelligent triage
• Turning on and configuring intelligent triage
• Allowing agents to see and update intelligent triage fields
• Viewing all possible intent and language values
• Understanding sentiment values

Requirements for using intelligent triage

Currently, intelligent triage is available for organizations who meet the following requirements:

Requirement type	Requirement criteria
Plan	Suite Professional and above (Legacy plans not eligible)
Industry This requirement applies only to intent predictions. You may still use intelligent triage for language and sentiment predictions without being in one of the listed industries.	Retail and e-commerce business-to-consumer (B2C). Does not include manufacturers, wholesale, business-to-business (B2B), or customer-to-customer (C2C). Software platforms or applications. B2C only. Does not include companies that commercialize gadgets or hardware components. Financial services. Traditional and online banking only. B2C only.
Ticket volume This requirement applies only to intent predictions.	More than 1,500 tickets in the supported languages in the last 6 months
Ticket language distribution	At least 90% of tickets in the following languages: English, French, Portuguese, German, Spanish, Italian, Japanese, Russian, Dutch, Korean, Turkish, Chinese, Swedish, Norwegian, Danish, Polish, Arabic, Romanian, Czech
Right model fit This requirement applies only to intent predictions.	To determine how well the current machine learning model applies to an account, we perform a quality check in a sample of tickets from the past 6 months. During the quality check, intelligent triage must label at least 30% of tickets with high confidence and at least 60% with medium confidence.

Limited availability participants

Zendesk previously offered intelligent triage as part of a limited availability offering. To have been part of the limited availability offering, you had to have:

• Met the requirements listed above
• Activated intelligent triage before Dec 12, 2022

Customers who took advantage of the limited availability offering retain free access to the intelligent triage functionality released before December 12, 2022 without needing to purchase a paid plan add-on. Additional functionality won't be included in the free limited availability offering.

Understanding intelligent triage

When you enable intelligent triage, the system adds new fields to your tickets:

• Intent: A prediction of what the ticket is about. To see the possible values, you can go to the Taxonomy tab of the settings page.
• Intent confidence: How likely it is that the intent prediction is correct. Possible values are High, Medium, and Low.
• Language: A prediction of what language the ticket is written in. To see the possible values, you can go to the Taxonomy tab of the settings page.
• Language confidence: How likely it is that the language prediction is correct. Possible values are High, Medium, and Low.
• Sentiment: A prediction of how the customer feels about their request. Possible values are Very Positive, Positive, Neutral, Negative, and Very Negative.
• Sentiment confidence: How likely it is that the sentiment prediction is correct. Possible values are High, Medium, and Low.

These fields are populated via ticket updates by Zendesk’s machine learning model based on the content of the first message of a ticket when it’s submitted. Agents can change the Intent, Language, or Sentiment fields if necessary.

When the fields are populated, tags are also automatically added to the ticket to help you build triggers, automations, and reporting. These tags reflect the values in the Intent, Language, and Sentiment fields, and are structured as follows:

• intent__value of Intent field
• language__value of Language field
• sentiment__value of Sentiment field

Tip: When creating triggers, automations, or reports, you can use the field value or the tag, depending on which one makes your build easier. For example, when creating a trigger with multiple intents, it’s easier to use tags because you can add them faster than adding separate conditions for multiple field values.

To learn how to create a report showing the results of intelligent triage, see Explore recipe: Intelligent triage predictions and confidence.

Tip: You can also use the Tickets API to return the intent and language as custom field values.

Turning on and configuring intelligent triage

Administrators can enable intelligent triage in Admin Center.

To turn on and configure intelligent triage

1. In Admin Center, click Objects and rules in the sidebar, then select Business rules > Intelligent triage.
2. Select Turn on intelligent triage.

   

   You're taken to the following screen, where the status of the Intent, Language, and Sentiment prediction types is initially listed as Pending. After a moment, the status changes to Active, letting you know that intelligent triage is ready.

   

3. To further configure or turn off the individual prediction types, click Settings for the prediction type you want to modify.
4. On the prediction type's settings page, choose from the following options:
   • Turn on <prediction type> triage: Clear the checkbox if you don't want the system to automatically classify tickets by that prediction type.
   • Email channels: The Web form, Email, and Web service (API) channels are enabled by default. If you want, you can also enable the Text, Web Widget, Mobile SDK, and Mobile channels.
     Note: While we generally don’t recommend enabling additional channels, you might consider enabling them if you have a lot of ticket volume on those channels. Channels with more conversational messages, system notifications, or spam will have lower-confidence predictions.
   • Messaging channels: The Web Widget, WhatsApp, and Facebook Messenger channels are enabled by default. If you want, you can also enable the Android SDK, Google RCS, Instagram DM, iOS SDK, LINE, Slack, Telegram, Twilio SMS, Twitter DM, Viber, and WeChat channels.
   • Exclusion conditions: If you want tickets created by agents to be excluded from the intelligent triage process, select the Ignore agent-initiated tickets checkbox.

     

5. Click Save.

Note: If you use any third-party tools to make automated updates to your tickets, make sure those tools are configured to use the safe_update property so that intelligent triage ticket fields aren’t inadvertently overwritten. For more details, see Protecting against ticket update collisions.

Allowing agents to see and update intelligent triage fields

When you enable intelligent triage, the system automatically fills out the Intent, Language, and Sentiment fields, but these fields are visible in tickets by default only if you have a single ticket form.

• If you have multiple ticket forms and you want agents to be able to see and change these fields in tickets, edit your ticket forms and add the new fields by dragging them into the form.
• If you have a single ticket form but you don’t want agents to be able to see and change these fields in tickets, edit your default ticket form and remove the new fields by dragging them off the form.
• If you want to use the intelligent triage field values only for Explore reporting or API use, you don’t need to add them to any ticket forms.

If you opt to let agents see the Intent, Language, and Sentiment fields in tickets, they can update the values in these fields if they feel they’re not correct.

Remember that Zendesk’s machine learning model is intended to detect the language, intent, and sentiment of a ticket’s first message only. Agents should understand that any updates they make to these fields should still be based on the first message.

Viewing all possible intent and language values

For the Intent field, the machine learning model used for intelligent triage is currently trained to identify intents for e-commerce and software organizations specifically. Additionally, intent and sentiment detection works only for the following languages:

English French German Italian Portuguese Spanish Japanese	Russian Dutch Korean Turkish Chinese (Simplified) Swedish	Norwegian Danish Polish Arabic Romanian Czech

For the Language field, intelligent triage can detect approximately 150 different languages.

You can see all possible intent and language values on the Taxonomy tab of the intent or language settings page.

To see all possible intent and language values

1. In Admin Center, click Objects and rules in the sidebar, then select Business rules > Intelligent triage.
2. Click Settings for the prediction type you want to modify.
3. Click the Taxonomy tab.

   

4. Click to expand each category in the list to browse individual intents or languages, depending on which prediction type you're viewing. Alternatively, you can search for a specific intent or language using the search bar.
5. To see the tags associated with each intent or language, click Actions > Show tags. This can be helpful when building triggers or views with intelligent triage conditions. See Creating triggers for automatically triaged tickets and Creating views for automatically triaged tickets.
6. To download a CSV file of all intent or language values, along with their associated tags, click Actions > Download CSV. This can be helpful when building triggers or views with many intelligent triage conditions.

Understanding sentiment values

Sentiment is a prediction of how positive or negative a customer feels about the request they’re submitting. The prediction is made based on the text of the customer’s first message and is grouped into one of the following categories:

• Very positive: The message likely contains strong positive words (like “brilliant” or “perfect”), positive words modified by intensity adverbs, or multiple positive sentences.
• Positive: The message likely contains phrases expressing gratitude, or one to two positive sentences.
• Neutral: The message likely contains factual statements without additional negative quantifiers (like “any” or “always”), or a mix of positive and negative statements.
• Negative: The message likely contains phrases expressing frustration, complaints containing negative words, or repetition of the same dissatisfaction.
• Very negative: The message likely contains strong negative words, capitalized text, multiple exclamation marks, or multiple negative phrases.

Zendesk’s machine learning model is specifically calibrated for customer service. This means that a ticket isn’t assigned a negative sentiment just because a customer has an issue with their order, can’t find the information they need, or some other similar “negative” situation. Instead, the model is tuned to analyze sentiment with the assumption that the customer is contacting customer service because they have an issue that needs to be addressed.