Talkdesk AI Trainer: Another Brick in the WallTalkdesk AI Trainer: Another Brick in the Wall

Earlier this month, cloud contact center provider Talkdesk announced the first human-in-the-loop (HITL) tool for contact centers. HITL is commonly used in artificial intelligence (AI) to leverage both machine and human learning to create machine learning (ML) models. The process would be to load a bunch of data into a machine-learning algorithm to train the data. The model would run, humans would score the data, then the algorithm runs again. This happens until the AI has been fully trained and can make decisions on its own.

In actuality, almost all AI uses HITL to some degree, but the humans that are in the loop are typically highly specialized and highly paid data scientists because working with the models is sometimes difficult. This new service, Talkdesk AI Trainer, allows agents with domain knowledge to tweak and improve the AI models without requiring a data scientist. The secret is a quality assurance system called “Agreement” that ensures outliers are removed from the training set. This results in more accurate models, which leads to improved accuracy, higher rates of automated self-service, lower costs, and ultimately increased customer satisfaction. Perhaps what’s most interesting is that the service can be “turned on” without the usual data gymnastics required to move data to-and-from a training location. Talkdesk’s Trainer is built into the operational data loop of the contact center.

This service also has some interesting implications for contact center agents as they will gain new skills and potentially shift their careers in the burgeoning field of AI. Agents already have the product and customer knowledge and adding in the ability to work with AI systems will be a real boon to current agents. One of the concerns of AI systems is that they will replace people and eliminate jobs. I’m certainly not going to say that won’t happen but there will also be more jobs created in the field of AI. I’m expecting the net job market to grow because of AI but it will require new skills.

AI in contact centers has been a hot topic of conversation for several years as businesses look to use advanced technology to improve customer experience. AI systems, however, aren’t perfect and can make mistakes, which is fine because bad data can be fed back into the systems and made part of the training set. The key is the ability to tag data as accurately or inaccurately as quickly as possible. That speeds up the overall training process and lets companies move forward with AI initiatives immediately.

While data scientists are required to build complex models that perform tasks such as classification and natural language processing, domain experts must create the data required to train these models. Some domains are simple, such as identifying objects like flowers or bicycles in photographs, but most humans can do this kind of data labeling (every human is a domain expert in this case). However, some domains require specialists to label x-rays for cancer detection or describe types of sediment in underwater photography.

In customer service, domain experts are a necessity because except for trivial problems like password resets, contact center dilemmas require knowledge of the company, industry jargon, the product or service, and pathways to the solution. The more experience someone has solving these problems for customers, the better they will be at creating training data. Talkdesk AI Trainer brings HITL capabilities to the contact center, allowing domain experts (i.e., agents) to use their domain expertise to create high-quality training data. Agents are to contact center issues what radiologists are to x-rays, and they can identify and decode a variety of problems and issues that consumers face on a daily basis.

The release of this service is another brick in the wall for Talkdesk to achieve its goal of automating 80% of customer interactions. This might have seemed a pipe dream just a couple of years ago as the machine learning models for sentiment detection, speech analytics, and conversational AI wasn’t very effective. One of the characteristics of true AI vs. rules-based systems dressed up as AI is that the models get better over time, which requires the models to be constantly retrained by humans. The new Talkdesk AI Trainer should help the company accelerate its path to hitting that 80% number sooner than later.