Understanding Customer Sentiment in JSM: What Data Does Atlassian AI Use?

[Thais Sales Monteiro de Sousa]

Hello Community,

I'm trying to better understand how the Customer Sentiment feature in Jira Service Management works behind the scenes.

Does anyone know what types of information Atlassian Intelligence analyzes to determine the sentiment of a ticket?

For example:

• Does it only evaluate the customer's comments and requests?
• Are agent responses also considered?
• Does it analyze the entire conversation history or only the most recent interactions?
• Are factors such as ticket status, SLAs, or CSAT results included in the sentiment calculation?

I'm particularly interested in understanding how the AI arrives at a sentiment classification (positive, neutral, or negative) and whether Atlassian has shared any technical details or best practices regarding the model's inputs.

I did not find any official information on this.

Any insights, documentation links, or experiences you've had with the feature would be greatly appreciated.

[Martin Runge]

Hi @Thais Sales Monteiro de Sousa

I dug into this, and I didn't find a published breakdown of the model inputs, so here is what's confirmed and what I would infer from testing.

The only official statement on inputs comes from the Atlassian beta announcement, which says Atlassian Intelligence uses "ticket context like the title, description, and comments" and updates the value in real time as new comments come in. That is the extent of what Atlassian has documented publicly.

From what I have observed in practice:

• It reads the text of the work item, not the metadata
• Status, priority, SLA breaches, and CSAT scores do not feed into it
• CSAT is a separate, customer-submitted rating and is not correlated with the sentiment field
• The value shifts as the conversation grows, so it reflects the latest state rather than a one-time scoring at creation

I found this page with some details: https://support.atlassian.com/jira-service-management-cloud/docs/about-customer-sentiment-analysis/

Cheers, Martin

[Martin Runge]

One thing that is easy to miss is that Sentiment is a normal field. An agent (user) can overwrite the AI value manually, and the field is available in JQL. You can use JQL to build a queue and validate the AI against your own reading of the tickets for a few weeks before using it operationally.