Building Predictive Routing Models Using Genesys Cloud AI

How-To-Articles
1

Building Predictive Routing Models Using Genesys Cloud AI

What This Guide Covers

  • You will architect an outcome-driven routing strategy that uses machine learning to match customers with the agents most likely to achieve a specific KPI (e.g., Sales Conversion, CSAT, or First Contact Resolution).
  • You will implement Genesys Cloud’s Predictive Routing engine, shifting from traditional “static” skill-based routing to a dynamic model that adapts based on historical interaction data.
  • When complete, your contact center will operate with a “data-first” mentality, where every routing decision is backed by a statistical probability of success, leading to improved operational efficiency and customer satisfaction.

Prerequisites, Roles & Licensing

  • Licensing: Genesys Cloud CX 1, 2, or 3 with the AI Experience add-on (or included in specific bundles).
  • Historical Data: At least 60 days of interaction data for the target queues with completed outcome data (dispositions, wrap-up codes, or CSAT scores).
  • Permissions:
    • Routing > Predictive Routing > View/Edit/Admin
    • Analytics > Conversation Detail > View
  • OAuth Scopes: routing, analytics.

The Implementation Deep-Dive

1. Defining the Routing Goal (The KPI)

Predictive routing is not a generic “better routing” button. It must be tuned to a specific business outcome.

The Step:

  1. Navigate to Admin > Routing > Predictive Routing.
  2. Create a new Predictive Routing Model.
  3. Select your KPI Goal:
    • Service Level: Optimizes for the fastest answer.
    • Outcome: Optimizes for a specific Wrap-Up code (e.g., “Sale Made”).
    • Duration: Optimizes for the lowest Average Handle Time (AHT).
  4. Architectural Reasoning: Choosing the right KPI is critical. If you optimize for AHT, the AI might route calls to agents who “rush” customers, which could negatively impact CSAT. Always align the KPI with your primary business metric.

2. The Data Training and Back-Testing Phase

Before going live, the AI needs to “learn” which agent attributes correlate with success.

The Step:

  1. Select the Queues you wish to include in the model.
  2. The system will perform a Back-Test against your last 30-90 days of data.
  3. Review the Predicted Lift report. The AI will tell you, for example: “Using Predictive Routing on this queue is estimated to improve Sales Conversion by 4.2%.”
  4. The Trap: If your historical data is noisy-meaning agents use the “Other” wrap-up code 50% of the time-the AI’s predictions will be unreliable. Predictive routing requires high-fidelity “clean” data to be effective.

3. Configuring the “Comparison” (A/B Testing)

Genesys Cloud allows you to run Predictive Routing in a “Test” mode where it only applies to a percentage of traffic.

The Step:

  1. Set the Traffic Split (e.g., 80% Predictive, 20% Standard Skill-Based).
  2. Define the Timeout threshold. If the AI cannot find a “high-probability” match within 15 seconds, the system should fall back to standard routing to preserve Service Level (SLA).
  3. Architectural Reasoning: This “Hybrid” approach ensures that you aren’t sacrificing speed for precision. If no “perfect” agent is available, the call still gets answered in a timely manner by a “qualified” agent.

4. Architecting the “Agent Attribute” Layer

Predictive routing doesn’t just look at skills; it looks at “Attributes” like tenure, language proficiency, and past performance on similar topics.

The Step:

  1. Ensure your Agent Profiles are up to date with custom attributes if needed.
  2. In the Predictive Routing configuration, select which attributes the model should consider.
  3. The Trap: Avoid including “protected” attributes (gender, age) in your model to ensure compliance with labor laws and ethical AI standards. Focus on performance-based attributes like Average_CSAT_Score or Technical_Certification_Level.

Validation, Edge Cases & Troubleshooting

Edge Case 1: The “Cold Start” Problem (New Agents)

  • The Failure: A new agent joins the team, and they never get any calls because the AI has no historical data for them.
  • The Root Cause: The AI cannot predict success for an agent with zero samples.
  • The Solution: Use the Predictive Routing Exploration setting. This allows the AI to “randomly” route a small percentage of calls to new agents to gather the baseline data needed to build their performance profile.

Edge Case 2: Divergent KPIs (Sales vs. Quality)

  • The Failure: Sales are up, but customer complaints about “pushy agents” are also up.
  • The Root Cause: The model is too focused on the “Sale Made” wrap-up code and ignoring CSAT.
  • The Solution: Implement a Balanced Scorecard approach. While the routing model might optimize for sales, use a Data Action in Architect to check the agent’s recent CSAT score. If it’s below a threshold, bypass the predictive model for that specific interaction.

Edge Case 3: “Empty” Queue Deadlocks

  • The Failure: A call sits in queue for 5 minutes even though agents are available, because the AI is “waiting” for a specific high-performing agent who is currently on break.
  • The Root Cause: The Timeout threshold is too high or the “Match Score” requirements are too strict.
  • The Solution: Lower the Timeout to Standard Routing to 30 seconds or less. Predictive routing should be a “preference,” not a “blocker.”

Official References