January 11, 2021

What’s the Difference Between Chatbots and Conversational AI?

We get asked this question at least once a week, and for good reason. The (truthfully warranted) confusion surrounding this topic is as widespread as it is self-perpetuating.

From our perspective, it is mostly the result of years of misinterpretation and misleading semantics that are endemic of any field that gains extreme traction, attention, and popularity in a short period of time. Chatbots and conversational AI are often used interchangeably to describe the same thing, which is valid to a small extent, but on the whole, their differences are glaring and, in a business setting, crucial


To truly distinguish between chatbots and conversational AI, we need to step back from the hype surrounding them and chart out the chronological and technological milestones that have bred this confusion over time. 

So, let's start from the very beginning. 

A Chatbot is Born


Chatbots, as we know them, were introduced to the world by MIT computer scientist Joseph Weizenbaum in 1966 in the form of Eliza, a chatbot based on a limited, pre-determined flow that could simulate a psychotherapist's conversation by using a script. Eliza carried out "conversations" by utilizing pattern matching and substitution methodology that gave users an illusion of understanding on the part of the program but had no built-in framework for contextualizing events.


The irony in this nativity story is that Weizenbaum created Eliza to demonstrate the superficiality of communication between man and machine, and in the process, built a chatbot capable of fooling its Sapien users into believing that it was human. This achievement would later be cemented by Eliza passing a restricted Turing test for machine intelligence. 


As we move down this timeline, keep the following concept in mind: limited, pre-determined flow.

The next notable stepping stone in conversational engineering would take place eight years later at the Stanford Artificial Intelligence Laboratory. Drawing from his past education as a psychoanalyst, developer Kenneth Mark Colby designed a natural language program called "PARRY" that simulated a paranoid individual's thinking. PARRY exceeded expectations and was the first-ever program to pass a full Turing test. 

To generate such impressive results less than a decade after Joseph Weizenbaum's Eliza, Colby employed a radically-different approach to conversational interfaces, conceiving a complex system of assumptions, attributions, and "emotional responses" triggered by shifting weights assigned to verbal inputs. So is this conversational AI?

No. Although PARRY had a better-controlling structure and a mental model that simulated the bot's "emotions," it was still rule-based, meaning it followed a rigid (albeit intricate) if X (condition) then Y (action) formula. 

If X Then Y formula

Add the term rule-based to our list. As a reminder, we've asked you to keep the following concepts in mind: limited pre-determined conversational flow, AND rule-based. Now let's continue.  


The next big name in the field was A.L.I.C.E. (Artificial Linguistic Internet Computer Entity). Composed by Richard Wallace in 1995, ALICE used an Artificial Intelligence Markup Language (AIML), a derivative of XML, that has tags that allow bots to recursively call a pattern matcher so that the language can be simplified. ALICE won the Loebner Prize three times in 2000, 2001, and 2004, an award bestowed on the most human-like systems. 


ALICE was extraordinary in every way, but can it be filed under conversational AI? The answer in this instance is again no. ALICE relied on a vast number of basic "categories" or rules matching input patterns to output templates. Wallace went for size over sophistication; ALICE makes up for its lack of morphological, syntactic, and semantic NLP modules by having an abundance of simple rules. 


In layman’s terms, ALICE had all the trappings of conversational AI, but in actuality, it was just a really big chatbot. Which begs the question: what exactly is conversational AI?


What Is Conversational AI?


To answer this question, we'll start by stating what conversational AI isn't. If chatbots are rule-based and follow a pre-determined conversational flow, conversational AI is (ideally) the opposite. Rather than following a rigid structure, conversational AI relies on Natural Language Processing, Natural Language Understanding, Machine Learning, Deep Learning, and Predictive Analytics to deliver a more dynamic, less constrained user experience than chatbots.


Although it may vary, the standard architecture of conversational AI comprises an automatic speech recognizer (ASR), a spoken language understanding (SLU) module, a dialog manager (DM), a natural language generator (NLG), and a text-to-speech (TTS) synthesizer. The ASR takes raw audio and text signals, transcribes them into word hypotheses, and transmits the hypotheses to the SLU. The SLU's goal is to capture the core semantics of the given sequence of words (the utterance). It identifies the dialog domain and its intent and parses the semantic slots in the user's utterance. The DM's goal is to interact with users and assist them in achieving their goals. It checks if the required semantic representation is filled and decides the system's action. It accesses the knowledge database to acquire the desired information the user is looking for. The DM also includes dialog state tracking and policy selection so that the dialog agent can make more robust decisions. 


Conversational AI standard architecture
Conversational AI standard architecture


This unique composition creates highly flexible and scalable conversational user interfaces that come with clear-cut advantages over chatbots:


Learning at Scale: Conversational AI solutions feed off various sources such as websites, text corpora, databases, and APIs. Whenever the source is revised or updated, these modifications are automatically applied to the conversational AI interface. In comparison, chatbots require continual and costly manual maintenance to their conversational flow to remain useful and effective. Furthermore, having full access to a database and API provides conversational AI solutions with the contextual elasticity needed to carry out fluid interactions with users. If, for example, a user changes their mind mid-conversation or requires a different service than the one initially requested, a conversational AI interface will automatically scrape the information necessary to complete the task. A chatbot is restricted within its pre-defined script and rules and cannot produce any output that was not manually inserted into its flow. 


Understanding: Rather than relying on a pre-written script, conversational AI uses Natural Language Processing (NLP) and Understanding (NLU), both subfields of linguistics, computer science, and artificial intelligence, which allow it to parse and understand inputs in the form of sentences in text or speech format. In stark contrast, chatbots may seem as if they comprehend words and sentences when, in fact, they're just following a rigid set of rules. Conversational AI can interpret, recognize, and grasp the finite nuances of human language, responding to rich context and vernaculars filled with slang, synonyms, homonyms (dual-meaning words), and jargon. 


Omnichannel: Whereas chatbots can only operate through text commands, conversational AI can be communicated with through voice. As such, conversational AI can be deployed as a voice assistant (Siri, Cortana, Google Home), a smart speaker (Amazon Alexa, Google Home), a conversational voice layer on a website, and even a virtual call center agent. This ability to cross over mediums means that companies can install one conversational AI solution across all of their digital channels, all streaming information to a shared analytics hub.

Conversational AI Vs. Chatbots comparison chart

What’s The Catch?


One can easily look at the chart above and wonder how, despite the obvious benefits of conversational AI, chatbots are massively prevalent and utilized on a global scale with more than 300,000 individual chatbots deployed on Facebook, and a market size that is expected to grow from $2.6 billion in 2020 to $9.4 billion by 2024. If conversational AI is better than chatbots, how is it that chatbots are still thriving? 


The answer to this question is rooted in the specific requirements of companies of varying sizes, sectors, and business models. Let's say, for example, that you are the owner of a medium-sized apparel chain. You want to grow your business and ramp up your customer engagement efforts. You conclude that the missing piece of the puzzle is a solution that will assist your burgeoning clientele with tracking the status of their online purchases. This is one case where a chatbot is a perfect tool for the job. You buy a DIY chatbot builder or recruit a freelancer to design a pre-determined conversational flow that will allow users to enter their order number, check it's status, or request exchanges and refunds. Basic chatbot technology can move this kind of conversation forward via bot-prompted keywords or UX features like suggestion buttons. As long as the user doesn't deviate from this exact task, there's no real reason to invest in a generally more expensive conversational AI platform. 


The same logic would apply to a restaurant franchise seeking to streamline its delivery service. Fast food giant Domino's Pizza offers its diners the option of placing their order through a chatbot, track delivery time, or be redirected to a human representative. So long as the user doesn't inquire about out of scope topics, say, opening hours, a chatbot interface is certainly up to the mission. What's more, since Domino's is not in the habit of changing its offerings regularly, updating and revising the chatbot doesn't become a pain point. 


Domino’s Pizza’s Chatbot
Domino’s Pizza’s Chatbot


Chatbots can prove sufficient for some SMBs (small-medium businesses) or large companies that want to fulfill a single task. The same cannot be said about data-heavy enterprises such as healthcare organizations that provide an extensive array of services. These types of institutions are in perpetual flux; physicians move practices and acquire new skill sets. Appointments are bumped or canceled. Hospitals change their policies and open subsidiaries and outpatient clinics. These moving parts demand a conversational AI solution that is in-sync with the ever-fluctuating rhythm of these dynamic organizations. It is also a far superior solution for businesses seeking to showcase their product catalogs within user interactions and empower them to browse with salient ease and efficiency. A patient searching for a specific treatment can use a conversational AI-powered virtual assistant to filter and distill results by a vast variety of attributes such as specialty, gender, language, and even accepted insurance. A hopeful renter or buyer can engage with a conversational AI virtual assistant in a productive dialogue, specifying, examining, and ultimately booking a tour of the desired property all within the same interaction.


Hyro’s conversational AI virtual assistant on Weill Cornell Medicine’s website



A Conversational Future


While on the surface it might seem that way, these two technologies are not at odds. Although conversational AI branched out from chatbots and is unquestionably more advanced, chatbots will go on to fill specific needs and tasks. With continuous innovation in artificial intelligence, machine learning, and natural language understanding, including OpenAI’s groundbreaking GPT-3 release, conversational AI will further evolve to become even more sophisticated. What is clear, however, is that the demand for both solutions will proceed to skyrocket over the coming decades, with 69% of consumers preferring to use chatbots for quick tasks and 70% of consumers intending to replace their visits to their healthcare provider, store, or bank with conversational AI virtual assistants.