The Botcom Bubble
On consumer-facing AI-enabled chatbots and why human interaction sucks in comparison
It was a rainy evening, and I was in the throes of my latest Herculean argument with my (now) ex. The pandemic meant that I was suddenly back in Singapore, and this had thrown our promises and our plans into disarray. Put simply, things just weren’t working, communication was breaking down, and I needed to get out and experience something more local, something more connected to my hometown.
And so I called them up, bracing for a long, agonizing wait followed by yet another verbal tussle. Our calls, I remember ruminating wistfully, always started the same way and then devolved. This time, though, my call was answered almost immediately.
Yet even though their greeting was the same as always, and so too their accent, I could tell that it wasn’t them who had picked up my call. Had they moved on already?
Even amidst my hurt, confusion and anger, though, I managed to detect an eerie familiarity in the voice on the line. As it hit me, a chill ran down my spine.
I knew who this a**hole was.
And it seemed that even taking everything from me -- telling me to charge my AirPods again, lecturing me about ‘Smart’ shortcuts on my television, and sneering “Sorry, A-r-een-jay, I didn’t understand that” -- wasn’t enough for the cold, mysterious, mechanical figure to whom this voice belonged.
“Hi. This is AT&T Customer Service”, the shape-shifting impostor said, business-like as ever. “Para hablar en español, oprima el ocho”.
(I’m with T-Mobile now, happily committed and with a renewal on the way.)
That day, as I over-enunciated my credit card number in an awful American accent, I started to ponder the idea of upheaving the traditional customer experience from an angle I hadn’t before.
Top-line growth is obviously crucial. As a payments enthusiast, for example, I’ve watched closely as businesses work to grow their top-lines by leveraging interfaces, partnerships, biometrics, encryption and integration to capture as many new customers and platforms as they can.
Increasingly, though, as pandemic recovery continues, the ability to focus solely on ‘sexy’ axes for expansionary top-line growth -- like payments optimization, for example -- is a luxury few will enjoy.
Businesses need to take a closer look at bottom-line efficiencies at the same time, and in a post-COVID economy, the productivity of human capital will be first under the microscope. This will be particularly true of industries where this capital is highly specialized and therefore limited (healthcare, for example), or where it is unnecessarily expensive to deploy relative to product demand (retail, for example).
In other words, people will continue to need everything from medical advice to customer support, but as dystopian as this feels to say, there will be a greater need than ever to solve for both the allocative and absolute costs of involving actual human beings in providing it.
It is this impending state of play that drives my fascination with AI-enabled chatbots — both voice-activated and message-based variants alike. Their cost efficiencies and speed are only two of the many facets that make me excited for the breadth and depth of their potential use. What is undeniable is that with their greater adoption, businesses will not only be able to ‘save money and automate jobs’, but also better study their consumers and their pain points, and thereby be better equipped to improve other facets of the digital experiences they provide.
At this juncture, it merits reiterating that AI is not a buzzword I’ve appended to make this seem more tech-y. Indeed, in this kind of application, AI is the difference between a data-generating, self-improving, human-adjacent system, and a glorified high-school Java project that uses switch statements and a UI to convert button-pressing outputs into something that aesthetically resembles a chat window.
With all this said, here are some of my thoughts on AI-enabled chatbots. I’ll cover what they are, how they work, and where they are being and might be used, with a particular bias toward applications in retail and in healthcare. And also, because I can’t hold this factoid in any longer, this is who voices Siri.
Defining the players
As an umbrella term, rule-based chatbots are any automated program designed loosely to mimic human interaction. They do this by responding to voice commands or text messages in pre-defined ways.
In the simplest sense, imagine something that has the capability (often via merely a simple text recognition module) to understand when it has heard or read the terms A, B and C, and is programmed to respond to A with X, B with Y, and C with Z. Add greetings and other embellishments on and you get the picture. They work with decision trees (flowcharts, if you will) and rules, and respond precisely as you have programmed them to.
Rule-based chatbots, therefore, tend to be most useful in relatively basic business situations -- ones that tend not to involve a lot of data, thereby not inviting much of a need to process or learn from any data involved.
At a stretch, in B2B sales contexts, for example, the data provided by a chatbot’s user may be stored in order to represent or inform leads.
More often, in support-based contexts, for example, the chatbot may merely serve as an alternative to a FAQ section, or as a way to query a database for simple account/transaction information like a tracking number. The only prerequisite for effective use is compatibility with currently used systems, like CRM, marketing software, and emails.
Many modern businesses, though, deal with more dynamic and complicated user needs, and therefore greater quantities and complexities of data -- in other words, scenarios that are both particularly taxing on human labor, as well as ripe for the introduction of machine learning.
Accordingly, though AI-enabled chatbots use similar underlying frameworks to their rule-based counterparts -- after all, their goals to respond helpfully are similar -- they are designed to be more versatile.
Indeed, via initial and then continuous ‘training’, they can be taught not only to uncover the true intentionality of queries that are posed in more human, less structured forms, but also to discern context, emotion, and even potential follow-ups. In other words, AI-enabled chatbots are more holistic. Via the infusion of natural language processing and sentiment analysis, they can engage customers in something that more closely resembles a 3D human conversation, rather than a back-and-forth series of 2D interactions. Thus, more use cases open up.
I’m decidedly more ‘excited’ about the reach and disruptive potential of AI-enabled chatbots, but over-digitizing customer experiences can be harmful -- not to mention at odds with all manner of branding considerations that businesses may have. Both variants, then, have their niches.
This graphic below, with credit due to hubtype.com, does an excellent job at summarizing the distinct, industry-agnostic advantages of rule-based and AI-enabled chatbots alike.
Demystifying the AI
At this point, I’ll focus specifically on AI-enabled chatbots, mediated via text. In abstract, they function relatively simply -- they need to accept an input, process and execute what it is asking for, and respond in a way that preserves the style and context of the conversation.
In technical terms, you need to account for three key elements: intent, entity, and action. I borrow the following elegant demarcation of elements from Bhavani Ravi. Each of them corresponds with an aspect of the ‘natural language’ functionality that sets AI-enabled chatbots apart. In explaining how this works, let’s take on the (utterly random) example of someone planning a trip to key cities in China.
Intent
An AI-enabled chatbot must be able to establish the intentionality of a customer’s message. Let’s say that our intrepid customer wants, first, to identify the ‘biggest’ city in China.
Intentionality is a complicated concept to navigate -- especially in English -- because of synonyms (pertaining to individual words), syntax (pertaining to the interchangeable order of words), and idiomatic expression (pertaining to stylistic flourishes in expressing a simple message). In other words, our customer could express the same question, variously, as:
“What is the largest city in China?”
“What city in China is the hugest?”
“What is the most enormous city in China?”
“Yoooo what is the thiccest city in Chinaaa?”
In solving for intent, an AI-enabled chatbot needs training in order to form ‘collections’ of various expressions, words and phrases that mean exactly the same thing in functional terms -- each individual input a user provides should map onto a collection, and each collection should later map onto an action or function. In this case, large=huge=enormous=thicc, and the idea of the suffix ‘est’ is analogous to the idea of ‘most’.
If gauging emotion or sentiment is a key part of a chatbot’s use case -- like, for example, if detecting urgency or anger would result in a different customer service response -- then rather than designating certain expressions as ‘equal’, they might be ranked differently, or even mapped onto different sets of responses. It is by explaining the connotations of language in this way that chatbots can be taught instincts of response, much in the same way that human beings are.
Intent training can be done to various degrees of precision using everything from lists of synonyms to narratives and policy documents, depending on the level of complexity that the chatbot needs to be able to handle.
This is what is known as natural language understanding.
Entity
This is the crucial point in the process between establishing the intentionality of a query and actually executing on it; I like to think of it as the chatbot equipping itself with the raw ingredients needed to produce its final output. Keeping our customer’s question in mind, this stage is about extracting the key ‘entities’ -- the key ‘details’, if you will -- of the question itself, by tokenizing key terms like ‘city’, ‘China’, and the comparable metric that the customer is interested in (in this case, size). With this information, all that the chatbot has to do afterwards is perform a simple logical operation.
To do this, there are already a number of code libraries which are equipped to spot, highlight and categorize an extensive range of known entity types in raw data, via a process called Named Entity Recognition (NER). The images below, with credit due to this article, show NER processing performed on a sample paragraph using a library called Spacy -- check out the documentation in the second image to see how each of the tagged entities are categorized.
By its most precise definition, this is an example of natural language processing.
Action
This is the simplest part of the process, only because it involves nothing that is particularly unique to AI. Once a program has access to logical operators, entities on which to operate, and an intended outcome, it simply has to find the right source from which to extract the requisite information or perform the requisite calculation. In most practical cases, external APIs can handle this part of the process, but in more sensitive use cases -- like medical information, for example -- they might be internalized, resulting in the need for a slightly more elaborate backend infrastructure.
In the case of our earlier example, an API would simply need to rank ‘cities’ in ‘China’ according to their ‘largeness’ (=size), and produce the necessary outcome -- and of course, express it using words of similar modalities to those used in the initial intent of the question, to help maintain the aesthetic fluidity of the conversation.
This output is, finally, what is known as natural language generation.
In summary, this kind of chatbot model is repeatable, calls upon extensive amounts of data, and most importantly, allows for improvement by generating a kind of second-order data of its own. In other words, any divergence from an ‘ideal’ outcome can be traced back to a failure of intent, entity or action, and then accurately ‘trained out’ of the chatbot’s instincts by providing it with better training material (intent), data sources (entity), or APIs (action) to use for processing.
Motivating the use
Why use them?
Zooming out from the architecture of AI-enabled chatbots, it merits clearly acknowledging the benefits they stand to provide -- to businesses, but to consumers too.
Businesses, of course, are under more immediate pressure to digitize, given pandemic tailwinds. Reasons are manifold.
Cost: Beyond the fixed costs of implementing bespoke, compatible chatbots into their legacy systems (which will decline, in any case, as the market for providers becomes more competitive), the cost profile of digitizing sales and support processes is significantly better than that of growing customer service departments. It’s not just about salary, but education, training and quality assurance as well. Don’t take my word for it -- IBM estimates that companies spend $1.3 TRILLION every year on customer service calls, and chatbots can conservatively be expected to result in tens of percentage points in cost savings if implemented correctly. Do the math.
Sales: Versus human beings, chatbots can be made significantly more effective and, for lack of a better word, more brazen in recommending new products to customers, allowing for sales channels to be dramatically broadened. They are also less frictional, and allow for quicker end-to-end sales and support processes, yielding benefits somewhat analogous to those of improving payments experiences.
Insights: Chatbots are practically walking fountains of customer insights. They serve as dynamic diagnostic records of pain points, complaints and product failures, as well as, cynically speaking, sources of data of varying sensitivity that can be sold (but I am not advocating for what you think I am). At the very least, companies can save money on hiring expensive strategy consultants (this, I am advocating for).
Consumers, too, stand to benefit from more seamless and convenient experiences, which indirectly also function as reasons for more businesses to adopt chatbots.
Availability: Chatbots don’t sleep. Responses are instant at any time of day or night, and there is no limit on the number of queries that can be processed in one engagement.
Personalization: While sales representatives can’t reasonably be expected to be familiar with your consumption history, or even with your relationship with their business, chatbots very much can be. What this results in is customer experiences that feel bespoke and therefore breed loyalty.
Quality: Without casting aspersions towards the absolute quality of human employees, chatbots guarantee a level of consistency, patience and integrity that human interactions cannot, by definition, ever attain. They may be able to perceive human emotions, but their own responses are usefully limited only to those emotions and actions that are within convention.
State of play
At the beginning of this piece, I established a kind of duality between industries like retail, with more first-order, almost existential cost considerations in the face of the pandemic, and industries like healthcare, which are not only under immense strain as a consequence of the pandemic, but also generally entail a greater opportunity cost for allocations of labor.
Businesses in both kinds of industry have become wise to the potential of AI-enabled chatbots, but adoption levels are drastically different. Why? Each kind carries its own set of obstacles, or lack thereof.
The retail industry, for its part, has been able to embrace chatbots more seamlessly because of a broader industry-wide flight toward digitization, catalyzed most recently (and most violently) by the COVID-induced shuttering of brick-and-mortar operations worldwide. IBM’s US Retail Index, for one, estimates that the pandemic has accelerated the entire industry’s shift toward e-commerce by 5 years, while on a more microscopic level, even historically stubborn retailers like Walmart (+97% eCommerce sales in Q2) and Target (+273% same-day fulfillment services in Q2) are leading the way from the top. With the difficulty of integration into legacy systems cited as one of the biggest obstacles to their increased use, this holistic flight toward eCommerce is almost serendipitously paving the way for chatbots.
The estimated impact of this -- serendipity or not -- is staggering for both top- and bottom-line improvement. Juniper Research estimates that retail sales from chatbot-mediated interactions will grow from $7.3bn in 2019 to $112bn by 2023, and that simultaneously, savings will amount to up to $11bn and nearly 2.5bn hours across businesses and consumers.
Retail is also a more explicitly ‘competitive’ industry, especially with the rise of eCommerce. I say this because ordering products online is less constrained by geography, static pricing and inventory than ever -- the lifecycle of a single product now spans global distribution channels, licensing to several multi-brand retailers at once (implying slashed margins for brands themselves), and more dynamic discounting than ever.
What this means for retailers is that competition is fiercer and survival is more dependent on the customer experience and on internal bottom-line improvements than ever; what this means for customers, meanwhile, is that choices are more plentiful than ever, and thus will result in demand being concentrated around the most seamless experience available. It is no longer a question of a first-mover advantage, but rather one of retailers needing to accept that chatbots and mobile communication are an integral part of any eCommerce offering. This is especially true for small retailers -- Shopify, in publicizing its analysis of 2020’s Black Friday (the 2nd largest day of online shopping in American history), indicated that nearly 70% of purchases from its clients were made on smartphones.
It bears emphasizing that the nature of the industry’s actual product offerings is everything.
Given the sensitivity of the healthcare business, reticence toward adopting AI -- not just in communication, but in surgery and diagnosis too -- is understandable given the relative infancy of most tools on offer. This being said, I advocate strongly for the broader use of AI chatbots in healthcare, above all because of their relatively controllable scope and low risk of failure, but also because they help to directly address a number of the biggest industry-wide shortcomings that the pandemic has highlighted. Here are a three key axes, which I’ve triaged, if you will, by importance:
Access: Access to healthcare in the first place is of paramount importance. The pandemic has illustrated that when faced with overwhelming patient volume, healthcare providers who embraced technology -- and particularly chatbots -- have coped disproportionately better. This is best illustrated via the example of the Memorial Health System, which rolled out no less than three manifestations of AI-enabled chatbots earlier this year. A COVID-diagnosis tool enabled more people to accurately assess their symptoms while allowing doctors and administrators to focus on inpatients -- inbound call volume was reduced by 25%. A virtualized waiting room tool, meanwhile, completely digitized the process of checking in for an appointment, resulting in fewer no-shows, fewer interactions in conventional waiting rooms with other sick patients, and fewer bottlenecks at facilities. Finally, a physician referral tool helped the System move toward automating its process of directing patients to specialists, reducing administrative paperwork and workloads, and optimizing access to healthcare based on specific treatment needs.
Quality: A JAMIA study during the pandemic found, by surveying 371 participants, that patients have little preference between a human being and a chatbot when it comes to diagnosing them with COVID-19 -- provided they can be assured of ability, integrity and benevolence. In other words, so long as quality is guaranteed, it appears that a widely held fear on providers’ parts -- that patients would feel underserved -- is perhaps unfounded. Given that chatbots, as aforementioned, can deliver consistent levels of quality (unaffected by the strains of practicing mid-pandemic), there is a clear pathway for their use in quality-assured healthcare, and one without previously perceived obstacles. Even if not highly technical medical processes, billing, registration, insurance, and equipment support can be discharged effectively using chatbots either centrally or in support.
Cost: This is a truly industry-agnostic axis of benefit, and matters are no different in healthcare. On a macroscopic level, Accenture estimates that the implementation of chatbots can save the American healthcare economy nearly $150bn by the year 2026. On a microscopic level, meanwhile, hospitals can save millions by simply avoiding no-shows (as Memorial Health solved for), while patients, for their part, can expect the cost of treatment to be slashed by up to 50% for treatments that can be mediated via the use of chatbots. As the real costs of healthcare continue to rise (check out the graphic below) -- particularly on account of inefficient administrative processes pertaining to insurance and payer systems -- cost-optimization is of paramount importance.
“I've often said, we need that Disney-like experience,” Kevin Pawl, senior director of Patient Access at Boston Children’s Hospital
I chose to focus on retail and healthcare because they are polar opposites on account of their necessity and their risk appetite for consumer-facing technology. The reality, of course, is that the potential scope for chatbot use populates the entire spectrum between them.
I normally enjoy researching startups doing exciting things in the spaces that interest me, but in this case, that’s something of a moot exercise. I say this because the recreational or freelance marketer can build out their activities on a product like Facebook Messenger, while anyone even an iota more serious than that can quite simply build, or contract from one of hundreds of SaaS providers, a chatbot of their very own.
Each of these SaaS options, for what it’s worth, differentiate themselves using subtle different industry focuses, which manifest themselves in the kinds of trained natural language processors (intent, entity, action) that they provide -- Wade & Wendy focuses on recruitment and hiring (and thus focuses on recommendations based on human traits and profiles), X.ai offers something of a personal assistant specializing in daily tasks, and Conversable has developed into a versatile, larger-scale enterprise favorite among Fortune 500 companies looking to customize their interactions with B2B and B2C customers alike.
Overarchingly, though it feels sadistic to advocate for a flight from human interaction at a time like this, the world of AI-enabled chatbots is an exciting one, whether in healthcare, retail or any other industry that values either its top or bottom line.
Look out for the next one you encounter and study it carefully. Be yourself -- you’ll be training it. And remember, if it helped me escape AT&T, how bad can it really be?