State of Conversational AI in 2021

The modern world started with conversation

by Nate Joens, Co-Founder, and Head of Innovation, Structurely

Societies, kingdoms, cultures, and simply humans, at their core, advanced through conversation. We wrote messages in scrolls, passed knowledge from generation to generation, and shared valuable information amongst one another. All in order to advance society as a whole. 

There are nearly infinite ways to communicate nearly infinite things. Hundreds of languages, shorthand, images, videos, GIFs, and emojis are, at their core, how we communicate now, how we’ve communicated in the past, and how we will continue to communicate.

Conversations are and always will be at the center of society, whether you’re a consumer or company a conversation must ensue to advance in any way.

So let’s advance.

Conversations are an Ancient Technology

Our earliest artifacts of human civilization come from analyzing conversations. Hieroglyphics and before, our prehistoric ancestors were communicating with one another. Fast forward to today, and civilization is still communicating. In some ways, no different than our prehistoric ancestors (emojis are just hieroglyphics, right? 😀 🤔)

So what’s changed since then?

Of course, in today’s digital age, the channels and ways in which we communicate are different. There’s Facebook Messenger, Text messaging, live chat, and hundreds of more channels we have conversations on. But there’s more to it than that. Since the first computer was invented humans began having conversations with computers. The way in which we communicated was through code, inputs and outputs, functions, executions, and more.

Fast forward to today, and while we still have professionals who write code, and interact with computers on a low level – most of society today interacts with computers through…that’s right, conversation.


Conversations can look like many things – code, sales conversations, and pitches.

But, just as we’ve evolved and advanced from inputting executions into a terminal on our computers in order to perform simple actions, we’ve also advanced in how we have conversations.

Enter the era of artificial intelligence (AI).

As a society, we’ve long dreamed of being able to communicate with our computers in natural language, to perform basic to advanced tasks. 

And within the last 10 or so years, that dream has become a reality thanks to AI, specifically, conversational AI.

But it hasn’t all been pretty.


The chatbot hypecycle has been real since 2012.

We were promised the world by the press, “AI” companies, marketers, and salespeople, but the products we were promised rarely delivered. (picture source Gartner)

Today, we’re well beyond the Peak of Inflated Expectations, we’re through the Trough of Disillusionment and quickly gaining speed through the Slope of Enlightenment accelerating towards the Plateau of Productivity.

That’s where this State of Conversational AI in 2021, and Impactful AI come in.


There have been many technological advancements since the dawn of time, and no time more than now where we’ve seen so many breakthroughs in AI.

However, much of what we saw during our recent period in the Trough of Disillusionment caused the industry great pause.

Our mission with Impactful AI is to accelerate the Slope of Enlightenment in this AI revolution.

This mission starts now, in the year 2021, with this guide, the State of Conversational AI in 2021.


by Dr. Jason Mars, Professor of Computer Science

Since circa 2013 one of the most important scientific disruptions began in computer science.

The AI application space of natural language processing (NLP), a space that lived and breathed technology based on computational linguistics, was turned on its head with the invasion of deep learning (computers mimicking how our brain cells learn). Instead of identifying the nouns, adjectives, and verbs of a sentence, these models just listen and understand, much as our brains do.

Though this not-so-hostile deep learning takeover in NLP was not as abrupt or dramatic as the same invasion into the application space of computer vision, which felt more like an overnight change of the world, its slower insurgence has proven just as complete and permanent.

Understanding this change in the technology space is important, as not only is it interesting, but it’s also descriptive of why we are where we are, and also prescriptive of where we’re heading in 2021 and beyond. 

First, let’s demystify some terminology. Terms like Artificial IntelligenceMachine Learning, and Deep Learning have been thrown around in marketing campaigns and abused by many. In layman’s terms, AI is simply having the computer do something you’d expect an intelligent being to do, regardless of how. Machine learning is the art of using models that are trained with examples, typically these models use human expert-defined “features,” in other words they are designed to look for certain attributes in datasets. Deep Learning is a special case of deep learning where the model learns only from datasets and features emerge automatically in the neurons. 

The text on the right was written by a deep learning AI, meticulously constructed word by word from a trained language model. 

This progress in the development of new theoretical underpinnings for NLP tasks has resulted in a gold rush style momentum that feels like a whiplash-inducing burst of innovation. It all started in 2013 with the Word2Vec model presented in the academic paper “Distributed Representations of Words and Phrases and their Compositionality”.

This paper was the first to demonstrate a neural network model trained to encode meaning in words by simply scanning through ‘reading’ millions of real-world sentences. This innovation was quickly followed up by another heavyweight contribution in 2014 with the Glove model presented in “GloVe: Global Vectors for Word Representation”.

That work demonstrated a new way to ‘read’ through large amounts of human-written text while capturing global context and co-occurrences of words. With these two juggernaut contributions in hand, the academic and commercial realms chewed on these inventions generating many related works, technologies, and innovations. 

Then in 2018, we observe another groundbreaking burst of innovations with the inventions of the GPT, Elmo, and Bert models.

All three of these contributions presented the value of reading through data while acknowledging the order of words in their sequence while ‘reading.’ Additionally, with the BERT model, we see the fruition of transformer models and attention layers. Without getting too techy these types of models can capture context and sequence in a parallelizable way, meaning the computer can analyze multiple parts of the data simultaneously while still ‘seeing’ the sequence relationships.

Acknowledging sequences in these models may seem like a small tweak but it cascades into massive capabilities in NLP intelligence. These works are the works that lead up to the GPT3 model for which you’ve just heard the model speak for itself above. 

From a practical standpoint, this means practical applications and use-cases that have been pondered for a long time can finally be realized at a level reliable enough to be commercialized and deployed in production.

Thus far we’d discussed where we’ve been.

But where are we going? 

Well, our trajectory is prescriptive of what’s to come. What we’ve seen up to now is the tip of the iceberg of the revolution coming for which we are now only feeling the pre-quakes.

For example, and in particular, there is a doozy upon us. There is a new insight that I believe serves to drive the next revolution in NLP. It is inspired by generative adversarial networks (GANs). These AI techniques have already boggled the mind in the visual sphere, and we are only starting to envisage what the implications are in NLP and other spaces.  

In a nutshell, GANs represent a reinforcement learning-inspired methodology for deep neural networks. The key principle is to have a neural network that is challenged and brutally co-trained with an adversary neural network automatically. AI training AI.

Imagine a sparring partner while training a boxer. You may be able to get a decent boxer just using a training bag, but put in a sparring partner and you can get the best boxer. This technique has resulted in AI that can invent people’s faces and generate pictures of these fake people that humans can not tell were produced by an AI. 

Here’s a challenge, which of these pictures are real? 

None of them are real! An AI constructed these people from thin air. These are faces generated by an AI built with GANs. Imagine what these models can do in the realm of NLP.

Well, it turns out there are a few scientific inventions that are just starting to ponder these questions. For example in 2020, the Electra model presented in “ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators” GAN-inspired ideas are leveraged and are demonstrated to generate State of the art Language models.

I can only ponder when we get to the point that such a GAN-inspired technique can generate textbooks that are as indistinguishable from human-written textbooks as the images above are similarly indistinguishable. It’s breathtaking to ponder the text above written by GPT3 is an AI that has only had a punching bag, imagine what happens when giving that AI a sparring partner. 

In 2021 and beyond, we will be having artificial intelligence writing whole email drafts for us, writing computer software code for us, and rewriting our poorly written articles to be better.

If I were to start 3 companies / new products today it would be first a technology that automatically generates, scores, and rewrites marketing emails and account-based marketing assets for other business, a virtual assistant that serves as a mock interviewer that allows you to practice your job interview and will tell you how to improve, and toy robot that servers as an AI that learns from children continuously and chit chats with them ever-evolving its knowledge of the world, growing with your child. The tech is ready. 


by Dan J. Levy, Editorial Strategy, Zendesk

Messaging is having a moment. Even before the pandemic, people around the world were relying on the messaging apps on their phones to keep in touch with friends and family, order food, and get around. But in the last year messaging has truly become a lifeline — and more businesses have joined the conversation.

Among companies using Zendesk, customer support tickets over messaging have surged more than 50 percent in the past year, with customers reaching out for help over social messaging apps and via companies’ homegrown properties. Between Facebook’s trio of increasingly unified messaging channels (WhatsApp, Messenger, and Instagram), and growing business messaging platforms from Apple and Google — not to mention good old SMS — there’s no shortage of ways for businesses to chat with customers.

But while the consumer messaging apps get most of the attention, the vast majority of customer conversations are still happening on businesses’ own websites and mobile apps. In fact, according to our latest CX Trends report, 45 percent of customers prefer to use embedded messaging to talk to businesses.

For customers, the benefits of messaging a business are obvious:

But the other thing that makes messaging so powerful for business is that it’s built for automation.

Conversational AI allows companies to be available to customers 24/7, and to intelligently route customers to the right people, at the right time, with the right context. Our research has shown that customers are happy to talk to a chatbot as long as they answer their questions, and allow them to speak to a real human if need be.

We believe that in order for brands to deliver truly personalized customer experiences, conversations need to flow freely within an organization — across channels and departments. Sometimes businesses need “out-of-the-box” solutions and sometimes they need to build and customize to their needs. That’s why we made an open messaging platform and marketplace that allows companies to connect all their business systems to Zendesk, including specialized chatbots for everything from sales to marketing to eCommerce.

An open platform allows businesses to gain a truly unified view of the customer and for us all to build a more conversational future together.


by Pete Jones, Director of Demand Generation, Structurely

As this guide has detailed, the science behind conversational AI has changed the landscape for traditional AI. The combination of conversation and AI will forever change how you talk to your customers and has changed the playing field for business.

As Dan said in the previous section:

Your customers want to message you on the platforms they feel comfortable with, and they expect you to be there when they are. Further, they expect the conversation to be consistent across these platforms, one of the leading challenges for business.

Thankfully, technology has evolved to produce chatbots that are indistinguishable from humans. That can be customized to a brand or an organization and specific to a unique function in the customer journey to provide your company’s best experience.

Your sales team can follow up with a lead without even hitting a keystroke or placing a call. Your marketing department can get more out of their ad spend, and your customer support team can message with your leads at 2 am. 

But, you don’t have to hire more staff to meet these needs. With AI Assistants help, you can accomplish these tasks without a significant increase to your bottom line. 



Marketers spend an excessive amount of time strategizing, writing, and planning to build demand within their ideal customer profiles. But, with conversion rates between 1 and 3 percent on average, most of the leads generated end up in a database, not as a conversion.

How many contacts do you have?

A thousand, five thousand, ten thousand?

You’ve spent countless hours and dollars getting those leads into your database. It is in your best interest to try to squeeze in any opportunities you can get. Just because they failed to open your email or show you the buying signal does not mean the lead is dead. It just may not have been the right time.

You’re still sitting on a gold mine of contacts ready to be mined and not with another canned email from your marketing automation solution. Instead, the data is ready to be mined with a nurturing strategy that uses two-way messaging, text or email, for an entire 12 months.

Keeping your brand in front of your leads increases your brand awareness and increases the likelihood they’ll think of you when they want to consider your product. Leveraging conversation rather than relying on stale canned messages is the future and the best way to increase your conversation rates and your overall return on investment.

Conversational AI can help with this predicament. With an automated nurturing strategy and a solution that uses two-way communication, no lead is lost forever. In fact, the nurturing strategy is a great way to keep your sales team happy on an otherwise slow month.


Sales organizations tend to navigate to the newest technology to do more with less. Their motivation to spend more time closing and less time cold-calling has provided innovative software companies opportunities during the past several decades.

Just think about Salesforce for a moment. Those innovations have helped sales organizations improve their efficiencies and data best practices, but it hasn’t solved the age-old sales problem.

Lead follow-up.

Consumer behavior is changing; they want answers now, and they want a company that can be responsive to their requests 24/7/365. They don’t want to wait for a canned email message or schedule a demo in three days. There is a gap; salespeople are more motivated by closing business than answering questions for unqualified leads or even cold-calling. As a result, this Drift study proved:

Failing to satisfy consumers’ expectations is akin to just shooting yourself in the foot and a colossal waste of the lead’s dollars. It also leaves a sour taste in the consumers’ mouths that will likely lead them straight to the competition. 


According to PwC, customer support teams have a pretty low margin for error with their customers; as 32% of all customers say they will switch to the competition after just one bad experience.

The larger problem is “bad experiences” can vary across the board from low question resolution to long wait times or endless transfers. While it appears the deck is stacked against customer support teams, there is hope.

Solutions like Zendesk provide teams a communication channel across all messaging platforms like Facebook, WhatsApp, WeChat, and many more.

That means you can reach your customers where they are when they want resolution.

Layering in conversational AI within a platform like Zendesk’s Sunshine Conversations allows your customer support team superpowers. They’ll now have the option to engage your customers across messaging platforms like Facebook Messenger immediately, Twitter, WeChat, WhatsApp, text message, and email, all from one central platform.

A Salesforce survey shows that

Your team needs to be consistent, present, and engage your customers immediately with answers to their questions. It’s a tall task, but you can do all of that with an AI Assistant.

As 2020 showed us, we can never be too prepared to respond to crises. Your AI Assistant can scale to accommodate a spike in messages when there is a time of need.

Lowering your CSAT no longer requires the overhead of hiring and training new staff to accommodate customer expectations.

Adding an AI Assistant to your customer journey will forever change the way you engage with your customers. Peaks and valleys in business will no longer cause irreversible harm to your operations or make you lose face with your market. They’ll also not make your staff go crazy, working a ton of overtime trying to keep up and drive morale down.

It is time to advance. Your businesses, society, and consumers are ready they just don’t yet have the trust built with conversational AI technology. But, that is where we come in.