Predictive Routing

Genesys PureCloud Predictive Routing Documentation

Study Notes

Topic	Description
Predictive Routing	Machine learning-powered intelligent routing that matches contacts to optimal agents
AI Engine	White-box models with explainability features showing feature importance scores
Outcome Labels	Custom KPI metrics tracked and used to retrain models continuously
Queue Configuration	Per-queue settings enabling/disabling predictive routing with KPI selection
Model Retraining	Weekly automated retraining with daily data updates ensuring current accuracy
Data Retention	No model data retained beyond 90 days for privacy and compliance

Admin → Architect → Routing → Predictive Routing Configuration OR Admin → Contact Center → Routing → Predictive Routing Settings

Predictive Routing Overview

Genesys Cloud's predictive routing uses machine learning to rank agents for optimal handling of interactions, supporting key performance indicators like average handle time and next contact avoidance. Unlike traditional rule-based routing, predictive routing analyzes hundreds of data points in real-time to identify which agent is most likely to deliver the best customer outcome based on your defined business goals.

Key Capabilities

Predictive routing uses white-box models that allow gaining insights into how the features contribute to a prediction, with each input feature given a percentage/score that represents its importance

Continuously improves scoring accuracy based on outcome data from previous interaction-agent matchups

Real-time agent scoring based on historical performance and current state

Automatically retrains and updates features used for agent scoring with daily data and retrains the data models weekly

Support for custom KPI optimization (AHT, NCA, CSAT, etc.)

Automatic fallback when models are unavailable

How It Works

Contact arrives at system

Contact metadata extracted (type, queue, customer data)

URS Strategy Subroutines submit interaction details to the Core Platform, which scores agents for their historical ability to handle such an interaction

Machine learning model scores all available agents

Contact routed to highest-scoring agent

Interaction outcome captured and logged

Models updated with new outcome data for future predictions

Edition & Module Requirements

Requirement	Details
Minimum Edition	Premium Edition (Genesys Cloud CX 1-4)
Module	Workforce Optimization or Predictive Routing add-on
License Type	Included with qualifying editions
AI Tokens	May consume tokens depending on configuration
Setup	Admin access to Architect and routing configuration

Study Notes - AI Model Features

Feature Category	Examples	Impact
Agent Profile Data	Skills, tenure, department, certifications	High - Core matching criteria
Agent Performance Data	Historic AHT, FCR, quality scores, tenure	High - Predicts future performance
Availability Data	Current status, workload, after-call-work	High - Real-time readiness
Interaction Metadata	Contact type, channel, customer segment	Medium - Context matching
Queue Statistics	Historical patterns, time-of-day trends	Medium - Volume prediction
Customer Data	History, value, previous interactions	Low-Medium - Personalization

AI Model Architecture

White-Box Model Explainability

Genesys helps you deduce the prediction by presenting a global interpretation that describes the average behavior of a model. A high value means that the feature will have a larger effect on the model's predictions and ranking of agents. While a small value is given to unimportant features whose contribution is mostly ignored for the model's predictions.

Model Training & Retraining

To keep up with changing levels of agent proficiency and customer interaction contexts, the data models continually retrain and learn from the latest features. Genesys Cloud updates the features used for agent scoring with daily data and retrains the data models weekly. No data is retained in the models for more than 90 days.

Privacy & Compliance

Genesys does not use PII for the agent scoring process. Genesys Cloud only uses transaction conversation data to train machine learning models.

Data Requirements for Optimal Performance

Recommended Data Volume

Genesys recommends a periodic upload of outcome data to ensure better model training and prediction accuracy. The recommended frequency is once a day, or, at the minimum, once a week. Genesys recommends at least 90 days of data and ideally 180 days of data. If you retain fewer days of data, you can still use and benefit from predictive routing. However, the quality and effectiveness of your AI models and predictions is likely to suffer and the resulting benefit will be reduced compared to standard routing.

Data Sources

Agent profile data such as skills, tenure, department, certificates, employee type

Agent performance data such as historic average handle time for a queue

Custom outcome data via Outcome Attributions API

External data sources for outcome-based KPIs

Minimum Data Thresholds

Data, even if available, is considered for the purpose of routing calculation only if the data volume meets a minimum requirement. If the volume of data does not meet the specified threshold value, the machine learning model does not use the available data. While not mandatory for Genesys predictive routing to operate, for better prediction results, Genesys recommends that you ensure the availability of sufficient volume of data against maximum number of fields.

Outcome Labels & Custom KPIs

Default KPIs

Predictive Routing can optimize for standard Genesys metrics:

Average Handle Time (AHT) - Minimize time to resolve

Next Contact Avoidance (NCA) - Minimize customer repeat contacts

Customer Satisfaction (CSAT) - Maximize customer satisfaction scores

First Contact Resolution (FCR) - Prevent escalations and transfers

Custom Outcome Labels

If you use outcome-based custom KPIs, Genesys predictive routing relies on data from external data sources, which it receives through the Outcome Attributions API.

Uploading Outcome Data

Frequency: Daily recommended, weekly minimum

Format: CSV or API integration via Outcome Attributions API

Data Requirements: At least 90 days, ideally 180 days for optimal models

Processing: Data uploaded to Genesys Cloud for model retraining

Model Updates: Weekly retraining cycle incorporates latest outcomes

Setting KPI Optimization per Queue

During queue configuration, administrators select which KPI the predictive routing model should optimize for that specific queue. The system will then score agents based on their historical performance against that metric.

Queue Configuration Steps

Step 1: Navigate to Routing Settings

Log into Genesys Cloud Admin

Go to Architect → Routing

Select queue to enable predictive routing

Open queue configuration settings

Step 2: Enable Predictive Routing

Locate "Predictive Routing" toggle

Toggle ON to enable feature

System validates that requirements are met

Configuration page appears

Step 3: Select KPI Optimization

Choose primary KPI from dropdown:
- Average Handle Time (AHT)
- Next Contact Avoidance (NCA)
- Customer Satisfaction (CSAT)
- Custom outcome metric

If custom KPI: Verify outcome data is being uploaded via API

Save selection

Step 4: Configure Scoring Parameters

Set minimum agent availability threshold

Define required skills for queue

Configure fallback routing behavior

Set scoring timeout (default: 3 seconds)

Enable/disable explainability logging

Step 5: Data & Outcome Mapping

Confirm queue is sending outcome data

For outcome-based custom KPIs, ensure periodic upload of outcome data via Outcome Attributions API

Map external outcome labels to routing KPI

Verify historical data volume meets minimums

Test data flow

Step 6: Validation & Testing

Enable predictive routing on small percentage of traffic

Monitor model accuracy metrics

Verify fallback routing works correctly

Gather baseline performance data

Review feature importance scores in reports

Step 7: Gradual Rollout

Increase traffic percentage gradually

Monitor agent utilization and contact distribution

Track KPI impact daily

Make refinements as needed

Scale to 100% when confident in performance

Model Scoring Process

Agent Scoring Flow

Incoming Contact
    ↓
Extract Contact Metadata
├── Contact type (voice, chat, email)
├── Queue assignment
├── Customer segment
└── Interaction intent
    ↓
Query AI Model
├── Input: Contact metadata
├── Features: Agent data, performance, availability
├── Process: ML scoring algorithm
└── Output: Agent scores (0-100)
    ↓
Rank Agents
├── Agent A: 87 score
├── Agent B: 72 score
├── Agent C: 91 score (highest)
└── Agent D: 65 score
    ↓
Route to Highest Scorer
├── Select: Agent C (91 score)
├── Fallback: If unavailable → Agent A (87)
└── Escalation: If all unavailable → Queue
    ↓
Log Interaction
├── Selected agent score
├── Feature importance values
├── Outcome when complete
└── Update training data

Feature Importance Explanation

The model provides transparency showing which factors most influenced each routing decision:

High Importance Features (weight >10%)
- Directly impact agent selection
- Should be monitored and optimized

Medium Importance Features (weight 5-10%)
- Contribute meaningfully to decisions
- Worth tracking for insights

Low Importance Features (weight <5%)
- Minimal impact on routing
- May indicate data quality issues

Real-World Implementation Scenario

Banking Contact Center

Queue: Mortgage Support
KPI Optimization: Average Handle Time (AHT)

Agent Profiles:
├─ Agent Sarah
│  ├─ Skills: Mortgage, Refinance, Loan Modification
│  ├─ Avg AHT: 6.2 minutes
│  ├─ Tenure: 5 years
│  └─ Current: Available (0 contacts)
│
├─ Agent James
│  ├─ Skills: Mortgage, Basic Support
│  ├─ Avg AHT: 8.1 minutes
│  ├─ Tenure: 1 year
│  └─ Current: Available (1 contact)
│
└─ Agent Maria
   ├─ Skills: Mortgage, Refinance, Advanced
   ├─ Avg AHT: 5.9 minutes
   ├─ Tenure: 7 years
   └─ Current: Available (2 contacts)

Incoming Contact:
"I'd like information about refinancing my mortgage"

Model Analysis:
├─ Contact Intent: Refinance inquiry
├─ Expected AHT: ~7 minutes
├─ Required Skills: Refinance, Mortgage
└─ Channel: Voice

Agent Scoring (for AHT optimization):
├─ Sarah: 89 score
│  └─ Reasoning: Strong refinance skills, excellent AHT,
│     lower current load
├─ James: 71 score
│  └─ Reasoning: Has required skills but higher baseline AHT
└─ Maria: 85 score
   └─ Reasoning: Best AHT historically but already has
      higher load (2 contacts)

Routing Decision:
Route to Agent Sarah (highest score: 89)

Expected Outcome:
├─ Faster resolution (Sarah's AHT advantage)
├─ Better customer experience
├─ Optimizes against AHT KPI
└─ Load distributed effectively

Model Metrics & Dashboard

Key Performance Indicators Tracked

Metric	Purpose	Good Range
Model Accuracy	How often predictions are correct	>75%
Feature Coverage	% of required data available	>85%
Agent Utilization	Even work distribution	70-90%
KPI Improvement	Impact on selected metric	+5-20%
Fallback Rate	When model can't score	<5%

Monitoring Model Health

Daily: Check model score distribution

Weekly: Review prediction accuracy vs. actual outcomes

Monthly: Analyze feature importance changes

Quarterly: Assess KPI impact and ROI

Best Practices

Data Quality

Validate Outcome Data - Ensure outcome labels are accurate and complete

Consistent Data Format - Maintain standardized data formats in API uploads

Timely Uploads - Daily uploads ensure models have latest information

Complete Records - Capture outcomes for all interactions, not just subset

Model Optimization

Start with One KPI - Master one optimization goal before multiple

Monitor Feature Changes - Track how feature importance shifts over time

A/B Test Approaches - Compare predictive routing vs. traditional on test queues

Iterative Improvement - Refine based on real-world performance data

Queue Configuration

Gradual Rollout - Start with 10% traffic, increase to 100% gradually

Skill Validation - Ensure agent skill assignments are accurate and current

Outcome Mapping - Correctly map external KPIs to routing selection

Fallback Testing - Verify routing works when models unavailable

Change Management

Communicate Changes - Inform agents about predictive routing changes

Monitor Agent Sentiment - Track agent acceptance and satisfaction

Provide Training - Explain how predictive routing affects their work

Set Expectations - Clear goals and targets for improvement

Common Implementation Issues & Solutions

Issue	Cause	Solution
Model not scoring agents	Insufficient data volume	Ensure 90+ days of outcome data available
Uneven agent utilization	Agents with similar skills	Review and update skill assignments
No KPI improvement	Wrong KPI selected for queue	Validate KPI choice aligns with business goals
High fallback rate	Skill mismatches in data	Audit and correct agent skill inventory
Slow model updates	Outcome data not uploading	Verify API integration and data flow
Feature importance unclear	New queue without history	Wait for sufficient data accumulation
Model accuracy low	Poor quality training data	Validate and clean outcome data

Interview Cheat Sheet

Question	Answer
What is Predictive Routing?	ML-powered routing that matches contacts to optimal agents based on historical data
What models does it use?	White-box models with explainability showing feature importance percentages
How often do models retrain?	Weekly with daily data updates; no data retained beyond 90 days
What data is needed?	90+ days recommended (ideally 180); daily or weekly outcome uploads
What KPIs can it optimize?	AHT, NCA, CSAT, or custom metrics via Outcome Attributions API
Where do you configure it?	Admin → Architect → Routing → Queue settings
Does it use PII?	No, only transaction conversation data for training
What's the expected improvement?	5-20% improvement in selected KPI
How do you upload outcomes?	Via Outcome Attributions API or CSV upload to Genesys Cloud
What if model unavailable?	Falls back to traditional routing automatically
How do agents get selected?	Highest-scoring agent routed first; fallback to next highest if unavailable
Can you explain routing decisions?	Yes, white-box models provide feature importance scores for transparency

Key Takeaways

Intelligent Matching - By continuously learning from real-time and historical data, it helps optimize important KPIs like average handle time, first-contact resolution (through next contact avoidance (NCA)) and more

White-Box Transparency - Models explain exactly which factors influenced each routing decision

Continuous Learning - Weekly retraining with daily updates ensures models stay current

Privacy-First - No PII used; data retained only 90 days

Custom Optimization - Select any KPI (standard or custom) to optimize routing around

Data-Driven - Requires quality outcome data for accurate predictions

Graceful Fallback - Traditional routing if model unavailable

Per-Queue Configuration - Each queue can optimize for different KPIs

Explainability - Feature importance scores show transparency in AI decisions

Proven ROI - 5-20% improvement typical on selected KPI

Additional Resources

Official Documentation Links

Predictive Routing Overview: help.genesys.cloud/articles/predictive-routing-overview/

AI Model Scoring: help.genesys.cloud/articles/how-the-ai-model-scores-agents-for-predictive-routing/

Data Requirements: help.genesys.cloud/articles/sources-of-data-for-predictive-routing-decisions/

Use of AI in Routing: help.genesys.cloud/articles/use-of-ai-in-predictive-routing/

Support & Training

Genesys University: genesys.com/training

Community Forums: https://community.genesys.com

Technical Support: https://support.genesys.com

Document Version Info

Last Updated: March 2026
Source: Genesys PureCloud Official Documentation (help.genesys.cloud)
Validated: Current with January-March 2026 releases
Version: 1.0