Why is Predicted Resolution and Predicted Satisfaction not available for email models in OI?

This article explains what Predicted Resolution and Predicted Satisfaction metrics are and why they are not available for email models.

Predicted Resolution and Predicted Satisfaction are metrics used for voice, chat, messaging, and SMS, where we have access to the full two-way conversation.

Email interactions are fundamentally different from real-time channels due to their asynchronicity. An email “interaction” can involve multiple back-and-forth messages, long gaps in time, or no response from the agent at all (example: forwarded emails, spam/junk, or emails that don’t require a reply). Because of this, the model analyzes the initial email from the cusotmer, not the full thread. This limits what we can reliably predict.
 

What is Predicted Resolution

Predicted Resolution estimates whether any actions remain outstanding for either the customer or the agent.

With email, predicted resolution is not reliable because:

• Only the initial customer email is analyzed. From that perspective, there are almost always actions required.
• Email is asynchronous, there can be days, weeks or even months before an email conversation is concluded, so the model cannot observe how the agent responds or whether the issue is actually resolved.
• Attachments are not analyzed, even though they may contain important context or additional actions.

For these reasons, predicted resolution cannot be determined accurately for email models.
 

What is Predicted Satisfaction

Predicted Satisfaction measures expected customer satisfaction across the entire interaction and is influenced by how the agent handles the conversation.

For email models, this is not possible because:

• We only see the customer’s initial email, not the agent’s response or the customer’s final reaction to the outcome of the interaction.
• The agent influence, which predicted satisfaction relies on, is not visible in a single inbound email.
• The asynchronous nature of email means the model cannot interpret the interaction as a complete, continuous exchange.

Because of this limited visibility, predicted satisfaction is not available for email.
 

💡A suitable alternative for email is Sentiment, which is available for all classified emails.

Sentiment is derived entirely from the customer's stated reason for contact, which in the case of email is the full initial email body (excluding signatures). Sentiment is not reliant on the agent’s side of the conversation and is not something agents have influence over.

Sentiment provides meaningful insight into the customer’s tone and emotional state, even though predicted satisfaction and resolution cannot be applied.

Learn more about Sentiment.