One of the most troubling statistics I have seen in my less-than-short career is as follows: Forrester Research conducted a poll of buyers of generative artificial intelligence solutions, and 70 percent of them said that when given the same prompt a generative AI solution will give the same answer every time. This is just wrong! The beauty, and one of the challenges of genAI, is that it just talks, and it will give different answers all the time. Knowing that buyers of genAI are wrong about such a fundamental capability does not speak well of the decisions these buyers will make on such a powerful software.
Happily, this is not a column about corporate-wide generative AI endeavors. This is a column about using genAI in customer service. It's happening already, and call summarization is the first application. Like the summaries that are becoming part of any video meeting, genAI can take the content of a customer interaction, write up notes from it, note the customer sentiment, and give it a disposition code. The results from a call summarization application are notes that are far better than what human agents tend to create, and agents will not miss the drudgework. This is a cost-saving, experience-enhancing application that is selling at a rapid clip.
I see this as a model for early rollouts of genAI for the contact center, an isolated use case that takes advantage of the strengths of genAI, protects from the vagaries of the technology, and does not require deep technical chops to deploy. There are other isolated use cases for genAI in customer service. Real-time language translation comes to mind; genAI is very good at this and is fast, easy, and relatively safe to deploy. Instead of monstrous, highly technical AI frameworks, I see a sort of piecemeal, one-application-at-a-time approach to deploying genAI for customer service. Be on the lookout for more simple, practical uses for genAI in customer service, coming to a contact center near you.
Enter Infrequently Asked Questions.
We are all familiar with the concept of frequently asked questions, and it's a common use case for chatbots at many companies. Someone at the company figures out the main questions customers or prospects ask and writes up answers to them. Customers access this information via chatbot or directly on a webpage. This is very useful, particularly since there are always questions that are asked repeatedly. Fast answers to those questions are valuable for customers and reduce the load on agents.
Beyond these frequently asked questions, there is always a long tail of questions that are not asked often enough to be worth building out answers. Enter the Infrequently Asked Questions application. GenAI is brilliant at handling these sorts of questions on the fly at runtime without requiring you to anticipate the question or pre-train appropriate answers.
I'm not going to get technical here, but retrieval augmented generation (RAG) at a high level allows you to tell a genAI system that it can only use information from a specific data source to build its answers. RAG complimented with other guardrails can prevent your chatbot from hallucinating, selling someone a car for a dollar, or many of the other things we worry about when it comes to genAI.
This enables a chatbot to point at a specific data set, Ideally, the data would live in a knowledgebase, which allows for scaling of the application through a manageable source of documents with tools to manage version control and the like. But, on a small scale, this can also be a set of PDFs that are fully vetted or possibly a support website. At runtime a customer or prospect can ask the chatbot a question and, if the answer is somewhere within one of the documents, genAI will find the answer on the spot, no pre-training on that specific question required.
Using genAI, the bot can find the answer and return a summary or the specific text of the relevant paragraph in the doc. The bot can also return a link to the full doc for context if desired. If you search on Google, you have used this functionality. The search engine will find an answer within a document, highlight it, and provide a link to the full document. When I talk to conversational AI vendors (the folks who build chatbots and IVAs), this is the application I am seeing go into production; it's out there in the real world today. This application will take off in much the same way call summarization has.
This is part of a larger whole to which I alluded above: In customer service, genAI is going to start as a set of practical, point solutions that solve real problems and provide real value to companies and their customers. First was the call summarization, next will be infrequently asked questions, then we can look at language translations and some other specific point solutions.
I'm an analyst; it is not in my nature to recommend point solutions or piecemeal approaches to anything, but it strikes me that giving raw genAI tools to people who fundamentally don't understand how the technology works is not a good idea either. The buyers are moving forward with frameworks, and the future is coming soon, however...
For today it's Infrequently Asked Questions for the win.
Max Ball is a principal analyst at Forrester Research.