Predicted CSAT (P-CSAT)

Predicted CSAT (P-CSAT) is a score generated by a machine learning model that estimates the customer satisfaction rating a customer would give to a support interaction, without requiring the customer to complete a survey. The model analyses signals from the interaction itself, the customer's history, and the resolution outcome to produce a predicted score on the same scale as the traditional CSAT survey (typically 1 to 5).

Traditional CSAT has a fundamental limitation: survey response rates typically range from 5% to 25%. This means that 75% to 95% of support interactions have no satisfaction data at all. The responses that do come in are subject to selection bias: customers who had extreme experiences (very good or very bad) are more likely to respond, while those in the middle are underrepresented. P-CSAT addresses both problems by generating a score for every interaction.

The model is trained on historical interactions where both the interaction data and the actual CSAT survey response are available. It learns which interaction patterns, such as long wait times, multiple transfers, negative sentiment in customer messages, or unresolved outcomes, correlate with low satisfaction, and which patterns, such as fast resolution, positive language, and first-contact resolution, correlate with high satisfaction.

Once deployed, P-CSAT enables real-time quality management. Supervisors can identify interactions that are likely to result in low satisfaction while they are still in progress, enabling proactive intervention. Quality assurance teams can prioritise review of low-P-CSAT interactions. And leadership can track satisfaction trends across 100% of volume rather than a biased sample.

P-CSAT models are typically supervised learning classifiers or regressors trained on historical ticket data paired with actual CSAT responses. The model ingests a range of features from each interaction and learns the relationship between those features and the survey outcome. The most impactful input features vary by organisation but generally fall into consistent categories.

P-CSAT decomposes into the same drivers as traditional CSAT, but with the advantage of being available for every interaction. This makes the metric tree actionable at a granularity that survey-based CSAT cannot achieve.

The tree shows that P-CSAT is influenced by factors both within and outside the support team's control. Responsiveness and interaction quality are directly manageable through staffing, training, and tooling. Resolution quality depends on agent capability but also on product stability and engineering responsiveness. Customer context reflects the broader customer relationship.

Because P-CSAT is available for every interaction, you can decompose it by agent, team, channel, issue type, and customer segment. This granularity enables targeted interventions: if one channel consistently shows lower P-CSAT scores, it points to channel-specific issues in staffing, tooling, or process.

1. 1

   Use P-CSAT for real-time supervisor intervention

   Configure alerts when an in-progress interaction's predicted satisfaction drops below a threshold. This allows supervisors to offer assistance, suggest approaches, or take over interactions before they result in a negative outcome, rather than discovering the problem days later in a survey.

2. 2

   Prioritise quality assurance reviews by P-CSAT

   Instead of reviewing a random sample of interactions, focus QA efforts on the lowest P-CSAT interactions. This concentrates coaching on the interactions that most need improvement and provides agents with specific, actionable feedback on their most challenging cases.

3. 3

   Identify and address systemic drivers of low P-CSAT

   Segment low-P-CSAT interactions by issue type, channel, and time of day. If certain issue types consistently generate low scores, the problem may be in the product or documentation rather than agent performance. If scores drop during specific shifts, staffing levels may be insufficient.

4. 4

   Improve model accuracy with richer feature engineering

   The quality of P-CSAT predictions depends on the quality of input features. Invest in sentiment analysis, conversation summarisation, and customer health scoring to provide the model with richer signals. Better inputs produce more accurate and more actionable predictions.

5. 5

   Close the loop between P-CSAT insights and agent training

   Use interaction patterns that correlate with high P-CSAT scores to build training materials. Show agents what effective empathy, clear communication, and efficient resolution look like in practice. Turn the model's learned patterns into coaching opportunities.