Based on an internal survey conducted of over 375 Global 500 customers, roughly 80% of concierge and customer service efforts are spent on providing informational services to their end-users, be they, employees or customers. Though there are of course situations where it’s necessary to speak to a human being, oftentimes these are requests for items that are buried in an employee handbook or a customer FAQ, and it’s inarguably quicker to simply call someone and have them find it for you. It makes sense, but it’s also a heavy operational burden on any enterprise, as many times customer and concierge services are managed via a manual workforce and a dedicated team.
Informational services are a necessity but also a time sink. In an ideal world, you would want the team to spend their energies on other tasks that are more judgment-intensive that require customer handholding. This is also compounded by the increasingly more-complex documents that we sign both in business and as consumers — multilayered terms of service and privacy policies, agreements with clients or contractors, things that are commonly dense.
Search Can Be Too Complex And Cumbersome
This problem is compounded in the enterprise. Enterprise content is often buried under various data sources in a variety of data hosting channels, making it extremely difficult for a layperson to figure out by themselves. A reasonable expectation may be to think that there is a central repository to host all the content — and that a search product could fix that problem.
The concept of content search works on the premise that the user is interested in filtering down topics from indexed content based on relevancy to their search query. So, potentially, all that a user can do using search is to filter down topics that they have read through and figure out what they want from the keyword-based content. They’re searching, clicking and searching again based on the information that they have, and then hopefully coming to the “right” answer, based on what they need to do.
Results are just that — results. They’re not computations of an actual answer and consideration of a particular question. They’re not answering the question directly — they’re giving what might be possible — without verification.
So, what does it take to get from here to give the end-user an answer to their question?
How AI Is Replicating Human Cognition To Find Answers
This is an extremely hard problem to solve technically because this involves not just understanding the semantics of a user’s question, but understanding the context of the content personalized to the end-user. Think about answering a question that someone asks you — even something simple, like “Can you meet me at 5 p.m.?”
This requires you to consider not simply whether you’re free that time, but whether you will be able to speak (e.g., location-wise — are you in a crowded room?), logistically (will you have phone signal where you are?), mentally (will you be alert enough?) and so on, at that time. These calculations are done by our brains, but in a massive repository of information, you have to contextualize based on the information available, the question itself and potentially external factors.
Once you respond back to the end-user, you have to be extremely precise — answering the question but also potential other questions, or aid in finding information to further help the user. Now, add the complexities of figuring out content formats, incorporating domain-specific knowledge and giving real-time responses to the end-user, further compounded by doing this at an enterprise scale with a variety of content and loads of data.
At that point, you get a sense of how challenging this problem is to solve.
Using Conversational AI To Find Answers
Using AI, deep learning techniques and natural language can synthesize and understand content available in the siloed information repositories and enable the user to ask and get an answer to a question in plain speak. It’s an incredibly complex problem to solve, one that requires you to combine different techniques to create something akin to the decision-making of a human brain.
Content understanding is a multilayered problem. It involves not just extracting data from highly unstructured content within various formats but also interpreting underlying semantics, identifying entity relationships, structuring and categorizing information that is essential for answering end-user questions in the context of content ingested. It’s not simply a case of ingesting — it’s a case of having the ability to use unsupervised learning to “comprehend” the document and be able to come up with answers to questions a user might ask.
It’s ultimately a reengineering of how we communicate with documents. We’re used to reading them and then asking someone (or searching) for an answer if we can’t find it from the text. The real solution is using new AI algorithms plus machine learning techniques combined with natural language understanding to give people the ability to ask a document a question and get a precise answer, much like a human would respond.
Challenges To Consider
Conversational AI is a rapidly advancing technology that is moving at light speed. As technology suppliers, we do not do ourselves any favors by liberally applying “AI washing” to everything we do. As a prospective technology buyer, consider applying conversational AI initially to a specific use case, one that consumes time and has high user frustration. It’s vital to verify the results from two dimensions — productivity as well as user experience. User experience to a large extent dictates success and adoption. Once you have a successful initial project, broader adoption is easier.
Implementing conversational AI technology at the enterprise level could free up more time for the 21st-century knowledge worker — making your business significantly more efficient.
This article is written by Ram Menon and originally published here