This article collects detailed information on how predictive routing works in specific scenarios. For more information about the data required to optimize predictive routing, see Data requirements for predictive routing. 

Predictive routing scores the agents who might handle an interaction using a machine learning model. Machine learning is effective at identifying patterns. In this case, patterns identify the agents who deal most effectively with certain types of interactions.

When predictive routing is activated on a queue, Genesys Cloud creates a model using various data sources, including agent profile data, aggregated customer data (such as whether they are a repeat caller), and historic interaction data. When an interaction is offered to the queue, Genesys Cloud assigns the interaction in the following way:

1. When an interaction arrives in that queue, predictive routing creates a list of all agents on the queue. The system retrieves data about the customer and each available agent. Predictive routing does not consider the routing status at this point.
2. Filters the list for required agent language skills and non-language ACD skills (if Skill matching is enabled). 
3. Genesys Cloud uses the model to process agent and customer data in real time and returns a ranking for each available agent. This ranking represents the agents predictive routing expects to have the most positive impact on the target KPI when handling that specific interaction. The highest scored agent ranks first.
   
   Note:  If the number of available agents on a queue is three or less, Genesys Cloud does not score the agents but routes the interactions using the standard routing method.
4. Before predictive routing times out (configured in the Queue details page), agent selection happens differently based on the number of interactions waiting on the queue:
   • Agent and Interaction surplus – When an agent becomes available, on queues that have predictive routing enabled, Genesys Cloud calculates the predictive score of the agent for each of the interactions waiting in addition to the interaction arrival time and priority. The system ranks all interactions by combining their waiting time with the predictive score, and then assigns the agent to the highest-ranked interaction. This behavior means that if an agent is predicted to perform better with one customer than another, then predictive routing will bias the assignment of the interaction toward the customer with whom the agent is predicted to perform better. This method ensures the best utilization of the available agent during times of conversation surplus thereby resulting in optimization of KPIs. 
     For example, if interaction 3 (wait time 35 seconds) ranks above interaction 1 (wait time 42 seconds), the available agent is assigned to interaction 3. However, when the interaction ranking is the same for more than one interaction, Genesys Cloud assigns the agent to the interaction with the highest wait time. 
     
     Note: Interaction surplus method can make estimated wait time calculations less accurate.
   • Agent surplus only – When an interaction arrives, Genesys Cloud calculates the predictive score of the available agents, ranks all agents by combining their time since the last interaction with their predictive score, and then assigns the interaction to the highest-ranked agent. If the highest ranked agents are not available, the system gradually expands the target agent pool, adding lower ranked agents. This process continues until an agent is found, or until predictive routing times out. 
     
5. If Genesys Cloud finds no qualifying agent before predictive routing times out or if the number of the agents on the queue is three or lesser, then the system routes the interaction using standard routing, which is the fallback routing method. 

Related links:

How the AI model scores agents for predictive routing