Time to Shift the Paradigm
In the late 90s the field of A.I. fell into disrepute. The promise was evident but the delivery fell well short. The period came to be known as the (first?) A.I. winter. Thanks to rapid improvements in computing capability, abundance of data and a major paradigm shift in A.I. methodology the long winter is now behind us and the impact of A.I. on our daily lives is rapidly becoming manifest.
A number of factors contributed to A.I.’s fall from grace, foremost was the mismatch between expectations and reality of A.I. solutions. While popular media were predicting the rise of robot household assistants and human-like capabilities the reality was that A.I. solutions at that time were often merely single-purpose models & glorified decision trees. The same worrying discrepancy seems now to be taking hold in the land of virtual assistants.
Ambitions vs. realistic expectations
We have recently witnessed such marvels as IBM Watson’s triumph over human opposition in Jeopardy! and Google’s Duplex successfully booking an appointment for a haircut, while many of the chat bots that front our customer-service organizations are currently little more than elaborate if-then-else scripts. Here we see the same risk: customers expect sentience while many chat bot implementations barely qualify as automation. Interactions between customers and chat bots are stilted, formulaic and so lacking in flexibility that any deviation from the script yields the ominous moniker “Sorry, I don’t understand what you’re saying”.
Case in point, this piece of swag from a 2016 chat bot conference:
Unfortunately the issues don’t stop there. Large trees (or even forests) of scripted dialogue and long dictionaries of relevant entities require constant, labour intensive maintenance. Small changes in business logic need to be implemented in many leaves of the tree or entire branches might need restructuring.
So, why the discrepancy between the flashy chat bot demos and the annoying simplistic agents we encounter in the wild? Limiting the scope of the bot’s challenge (question answering for Watson, appointment booking for Duplex) certainly helps. These situations are highly predictable and it’s easy to prepare the bot for most stages of dialogue that it is likely to encounter. Obviously one must also not underestimate the sheer manpower that big tech brought to bear on these bots to realize these undeniably impressive results. This level of investment is, however, somewhat beyond the budgets currently available for chat bot implementations.
So are practical, cost-effective, low-maintenance, human-like chat bots doomed to be as elusive as artificial general intelligence has proven to be? The answer is a resounding “No!”, but businesses and chat bot-developers do need to take up the challenge to shift the current paradigm. If we do not, there is a real danger of permanently damaging customer perceptions of chat bots to the extent that negative sentiment will lead to avoiding smart and dynamic chat bots in favor of more traditional communication channels (i.e., human agents) and ultimately defeating the purpose of employing chat bots in the first place.
The good news is: the tools and frameworks for developing unscripted chat bots are already available. Chat bots can be equipped with knowledge models to allow rudimentary reasoning about business logic. Discourse models allow bots to retain context and more easily resolve ambiguity and be dynamic in the dialogue. This will lead to far more naturalistic human-chat bot interactions and, thereby, a radically improved customer experience.
Statistical learning methods exist to allow bots to actively update their vocabulary based on closed-loop learning and customer feedback. Certainly this still requires some human-in-the-loop moderation, but in a more direct and intuitive manner and in far less time than maintaining dialogue trees would require. This making the models more effective and more efficient.
Investing in more versatile language processing capabilities of bots will deliver value. Success lies in a thorough assessment of processes, a focussed service design and skilled execution. The emphasis should be on the desired output and ambitious process design instead of upgrading a poor existing process with some A.I. stardust.