To train, or not to train? That is the business question.

Viktoria Krane

As recently as five years ago, many business leaders were skeptical of artificial intelligence(AI) and natural language processing(NLP), wondering what role these technologies could play within their businesses. No one is asking that question today. Instead, companies are feeling left behind, scrambling to choose from the array of AI-based solutions featuring trending technologies: Natural language understanding (NLU), large language foundation models, machine learning (ML), contextual understanding, conversational AI, chatbots, etc., the list goes on and on. 

In addition to the recent advancements in AI and ML which have surely propelled the proliferation of AI-enabled solutions, the pandemic has also caused a massive shift in the way consumers interact with businesses.  Most interactions are taking place through digital channels, and customers prefer self-service over speaking to a live support person to resolve their issues. In fact, 69% of consumers try to solve their issues on their own today. This new status quo has created ample opportunities for AI-driven automation. But how can a prudent entrepreneur sift through all the options and choose the ones that can have an immediate impact on the business?

Let’s start with machine learning. While the idea of a self-improving machine learning algorithm is very appealing, the cost in both time and effort may not be justified. ML projects are expensive and time-consuming. A typical AI project can take anywhere from 3 to 36 months to complete depending on the scope and complexity of the use case. To properly train an AI model, you need tens of thousands of labeled data samples to get started. Data professionals need to curate and label company- and domain-specific content to build solutions tailored to their specific use case. Unfortunately, data labelers are hard to find and are not cheap. Even if you are lucky enough to find them, the sheer volume of content that needs to be labeled could be extremely high. After spending months going through repeated cycles of training, testing, and optimization, you finally get the model working and launched into production. Congratulations, you now have to pay even more qualified and expensive humans to maintain and update it. To add another layer, many ML engineers and data scientists don't know how to build secure, scalable enterprise applications. According to a recent McKinsey study, the challenges of the hand-off between the AI team and IT to incorporate the ML model into the existing business applications are underlined by the abysmal 11% success rate of this transition. 

Some large organizations have already built sprawling ML departments, competing for a small pool of data scientists and NLP experts. For some well-funded use cases that benefit from plentiful data, if you can afford the time needed to train the model, building the models yourself is worth the effort. It can yield high accuracy and provide custom solutions. But for most use cases, ML proves to be too expensive to build and too rigid to use given the constantly-changing data. So, what’s the alternative? Is it possible to leverage AI to create business value without drawing on your already overstretched staff? 

In the title, we asked whether to train or not to train. At Kyndi, we don’t believe that training large ML models is efficient for most use cases. Instead, we offer a way for companies to get a head start with our generalized English language model and make it their own by teaching it about the company specifics. What is the difference between training and teaching? Training requires a lot of labeled data; teaching does not. We have very simple no-code tools for you to teach the language model to understand words unique to your company and industry such as product names, company-specific acronyms, and industry terms. As those things evolve or you have new things to add, you can continue to teach Kyndi to expand on its knowledge. It takes days, not months. Through tuning and optimization, you teach your personalized model to understand the specific industry vernacular. Also, if you know the questions your customers typically ask, you can create a smart frequently asked questions (FAQ) section. If you don’t know what those questions are, you can use Kyndi’s analytics tools to gain insights into your customers’ most popular questions and improve your content for a better self-service outcome. Unlike large ML models, Kyndi’s model can be quickly updated at any time since our no-code tools were designed to be used by domain experts and business users, rather than just technical staff. 

Imagine trying to troubleshoot a television that is not connecting to the internet after you accidentally changed the connectivity settings. You go on to the manufacturer’s website and scroll through their FAQs, and even download a 105-page operations manual only to find the answer 3 hours later on page 102. This is a frustrating experience that we have undoubtedly all had. What if you could offer a better search experience by delivering the right answers to your customers instantly, whenever they have questions? That’s what we are doing at Kyndi. Using a completely novel approach to understanding language, we’ve created a generalized English language model that is already pre-trained for natural language understanding. You can ask a wide range of questions any way you want and we will bring you back consistently relevant answers. Whether the answer is in the FAQs, your product website, a manual, or other unstructured data, we help you improve the self-service search experience for customers, as well as your employees. A delightful search experience like this will quickly result in increased customer satisfaction and loyalty. 

Curious about how Kyndi can help you improve your customer experience?

Book a  demo with us today!