A good AI model is one that recognizes 80%+ of user messages and predicts them to the correct intents. This is best achieved by a sensible intent structure, where intents are well defined and do not overlap, and the number of expressions are representative of the frequency of intents.
When building a model, remember to:
- Trust the data
The intent structure must consider the AI model and your customers’ behavior, and not what makes sense from a reply or process perspective.
- Prioritize the most frequent intents
The aim is to cover the most frequent queries and ensure these are handled by the bot. It is ineffective to build an AI model that recognizes all of customer queries, as certain more complex queries will be better handled by your agents.
- Focus on training
Build a solid foundation by weighting your intents using customers’ real data. Strengthen further after launch with regular training as you learn how customers interact with the virtual agent. Note, it is better for the model to not train than to train poorly.
The Impact Report and Training Center guide you to build the best model based on your customers unique behavior.