Loading

New Implementation of LLM-based Deflection and Abandonment Metrics within Agent Analytics (Update on July 16, 2026)

Julkaisupäivä: Jul 15, 2026
Kuvaus

Note:

The metrics will be updated on July 16, 2026. 

You may find that the current Agent Analytics metrics for deflection and abandonment, based on runtime events (like "end conversation" clicks or session timeouts), are often inaccurate and fail to reflect true issue resolution or user frustration.

The new Agentforce feature utilizes Large Language Model (LLM) judges to provide a more precise, qualitative, and holistic evaluation of entire agent-customer sessions, moving beyond rudimentary metrics to deliver actionable insights on agent performance and customer satisfaction.

Ratkaisu
New Metrics Definition that starts on July 16, 2026:
 
  • Deflection: A rank on a scale of 1 to 5 (1: No resolution, 5: Fully achieved) evaluating how completely the user's primary goals were achieved by the end of the session.

  • Abandonment: A TRUE/FALSE/UNSURE evaluation, assessing the likelihood of a user disengaging due to failure or dissatisfaction.

  • Key Technical & Business Details: The metrics are calculated and asserted upon session closure.
    Each LLM-derived metric includes an evaluation reasoner to provide users with a snippet-level summary of the LLM's logic.

  • The deflection and abandonment LLM calls are intended to be excluded from billing for out-of-the-box usage.

  • Explain the session outcome logic:
    Escalated: If yes, count as escalated.
    Abandoned: If no escalation, check abandonment=TRUE or likelihood of deflection (below 3). If yes, it counts as abandoned. 
    Deflected: If neither, determines high likelihood of resolution (scores 4 and 5).
    Ambiguous: Some sessions may ‌get a score of resolution=3 or Abandonment=FALSE/UNSURE, which indicates that there's no concrete determination for them. In these cases, we classify them into a new Ambiguous category. 

  • Rollout/Enablement:
    Note:
    • The new metrics are auto-enabled for generic agent types (Employee, Service) upon STDM enablement. For existing enabled customers, the transition will happen seamlessly with no additional steps required from their end.
    • The baseline data used to calculate the existing deflection and abandonment metrics will remain fully accessible for querying in the data objects and will not be deleted.
    • Data will show up only from the start of May. Sessions that were generated earlier than than are not scored.  This will happen since the new implementation will begin curating data only from the moment of enablement.
    • The metrics in the dashboard are still presented in percentage. The data to compute these metrics behind the scene will point to a different source.
    • This change does not apply to GovCloud.
Knowledge-artikkelin numero

005305353

 
Ladataan
Salesforce Help | Article