Understand

Thankful has two methods to understand users which both involve trade-offs.

Deep Learning AI

Thankful offers a shared, Deep Learning AI trained by hand across millions of conversations that agents and customers have experienced on Thankful’s platform.

Thankful builds these training sets and your AI models for you and updates them every day as new data arrives. Thankful understands your customers even after processes change, new slang becomes widespread, or major events like COVID-19 occur and dramatically alter the language used in service interactions and the problems customers encounter.

Using Deep Learning AI, Thankful achieves an astounding 99.9% accuracy across all tickets and channels. Thankful has nuanced understanding of why customers are writing in, how to help them, and even knows when it can’t handle a ticket and is best to leave it to an agent.

Nuance is an incredibly important part of customer service that’s frequently overlooked. Understanding the request only at a high level, such as that a customer wants to track their order, cannot yield good service. For instance, if a customer writes in, “Where is my order? I checked the tracking link and it hasn’t moved in 6 days. This was for a birthday but now won’t arrive in time,” simply providing the tracking link is a terrible experience. Is their package stuck in transit? Should we reship the order? What was their promised delivery date and are we outside the SLA? Should we discount future orders? How can we make this right?

We need to speak to and address the nuance, and only Thankful supports this. There’s no shortcut, except investing heavily in AI over many years with an in-house team of experts in offering incredible service. That’s how we achieve 99.3% CSAT.

Custom Models

Some businesses need to understand nuance outside of Thankful’s shared training models, and there may not be enough data to support a Deep Learning approach. For these situations, Thankful supports Custom Models which are unique to just your business and the training data is separate.

These models perform well on smaller datasets but their accuracy is less than Deep Learning and will not improve dramatically as the training set grows.

The accuracy of the model will depend on the data provided to it and the classes selected. To achieve the best accuracy:

• Choose dissimilar classes. Dissimilar classes are easier to distinguish than highly similar classes, and thus dissimilar classes require fewer training examples.
• Use fewer classes, as it is easier to classify accurately into a small number of classes than a larger number.
• Ensure each class the same number of training examples.
• Provide 30-250 training rows per class taken from real conversations with users.