In the latest edition of the Inner Circle’s Guide to AI, Chatbots & Machine Learning CallMiner’s VP of AI Rick Britt and Senior Data Scientist Yang Liu answered questions from their user community on a variety of topics on artificial intelligence (AI) in the contact center.
Check out our AI experts thoughts on these 6 questions from the Contact Babel user community:
What are the steps we need to take to use AI in our contact center? Where would be the best place to start?
AI can be applied to many places in a call center. One of the first things is probably to start with speech recognition. Without the text transcripts, audio recording in the raw format is much harder to work with. Once we have the text transcript, many natural language processing tools can be used, including search and count of words, scoring of agent behavior, and live monitoring of agents to give suggestions. Prediction models can also be generated based on a specific purpose, such as first call resolution prediction. Speech recognition is essential in this process, there would not be any text analysis without it.
Is there anything that successful AI implementations / projects have in common? Any pitfalls to avoid?
AI comes in variety of forms. Whether it’s the traditional AI that include natural language processing, or the newer forms including statistical learning, people who are implementing the AI must have a good understanding of the data they are working with. Written words and transcripts of speech use the same words differently or may use different words entirely. The same algorithm used on both is going to generate different results. One of the major pitfalls is using readily available tools without understanding how AI works. Because the tools for building model is so easy to use these days, people may not understand the math behind it, and it could generate biased or dangerous models.
How do we measure the ROI of AI? Are there any quick wins we can use to show our senior management?
Start with a simple task. AI is basically a fancy computer program with statistics underneath. Both computer and statistics work well on a large scale. AI is much better doing a specific task, and not good at uncommon tasks that require frequent interpretations. Start with a task that improves the internal process, like faster search function. This will allow you to develop partners in other parts of the company and gather supports.
Solutions such as speech analytics and knowledge bases / case-based reasoning have been around for a long time. What’s so different about AI?
Speech analytics belong to the category of “Good Old-Fashioned AI” or symbolic AI. This kind of AI uses human knowledge to build rules that best replicate results, which are viewed favorable by humans. Statistical learning AI, which is commonly referred to as AI in the current media, uses statistical models to predict results by correlating previous observations with results. In the statistical learning process, AI generalize rules that may or may not make sense to humans. You can use these AI generated rules as features to make predictions. The outcome of these AI generates rule is outperforming symbolic AI in many areas.
How much initial and ongoing effort/resource will AI require? Do we need a dedicated AI professional to keep everything running?
Building models that make accurate predictions without bias requires constant human intervention. In the initial stage, a team consisting of a user experience designer, a data science, and a developer will be needed to frame the solution, build a model, and deploy the model respectively. The data scientist and developer will need to adjust the model from time to time, and make sure it functions as intended. Our world changes constantly, and they have to make sure the world did not change so much, so that the model no longer fit the problem anymore. There will be hardware requirements, but they are much easier to identify and acquire.
Does AI require replacement of any existing technology, or will it work alongside what we already have?
Initially, we would build AI to fit the existing technology and solve a specific problem. To make the AI more effective, some changes to the existing technology will need to be updated or revamped for the long term. A lot of the updates are going to be focused on accessing the data faster.