Imagine a world in which you could talk to your computer as if it were a real person. It could make small talk with you, give you advice, answer questions and even anticipate what you’re going to say next. And it wouldn’t feel like you’re talking to a machine at all.

This is fast becoming a reality thanks to the rapid innovation of conversational AI — a form of artificial intelligence that enables a dialogue between people and computers.

What Is Conversational AI?

Conversational AI is a kind of artificial intelligence that enables people to have a dialogue with computers, whether that be to ask questions, troubleshoot problems or even make small talk. This technology typically comes in the form of virtual assistants or chatbots.

These days, conversational AI is especially top of mind amid the release of ChatGPT — a chatbot created by tech company OpenAI that has captivated investors, tech giants and the general public alike with its human-like dialogue capabilities. Tech platforms like Bing and Microsoft Word and PowerPoint are reportedly looking to incorporate ChatGPT to rival its competitors. And schools are banning its use due to cheating concerns.

Meanwhile, the use of conversational AI across many industries has hit new heights over the last couple of years. And by 2030, the global conversational AI market size is projected to be worth more than $41 billion, according to market research and consulting company Grand View Research.

“This is far and away the most exciting time for conversational AI right now,” Joe Bradley, the chief scientist and a senior VP of data science and machine learning at LivePerson, told Built In. “We’ve wanted to have human-like conversations with these systems for a long time now, and we’ve thought we might be nearing it for the last 10 to 12 years. And I think we really have finally turned that corner.”


What Is Conversational AI?

Put simply, conversational AI is a form of artificial intelligence that enables people to engage in a dialogue with their computers. This is achieved with large volumes of data, machine learning and natural language processing — all of which are used to imitate human communication.

“It’s about having a system that’s able to carry on a conversation with a human user, usually to solve a task or answer the user’s question,” Yves Normandin, a VP of AI technologies and products at customer experience tech company Waterfield Tech, told Built In, “in a way that imitates a human having the same conversation.”

“It’s about having a system that’s able to carry on a conversation with a human user.” 

Conversational AI often comes in the form of AI virtual assistants like Siri or Google Home, which can be used to give directions, manage your calendar, play games and more. It is also typically used as a form of quick communication between a company and their customers, whether that be to ask questions or troubleshoot problems. This is done through a chatbot, or a computer program that simulates human conversation through text or voice commands.

For many years, the public perception of chatbots was that they were generally “kind of lousy,” according to Bradley. “And people weren’t exactly wrong,” Bradley said. “You had to do a lot of painstaking work and thoughtful optimization in order to make them good.”

That’s less the case now, as we’ve seen with pioneering new conversational AI systems like OpenAI’s ChatGPT, Google’s LaMDA and others — some of which sound so convincingly human that they win prizes for it. These days, chatbots are so sophisticated they’ve been used to do everything from treat people with anxiety to defend someone in traffic court.

But it’s important to remember that conversational AI (even in its most impressive forms) remains a form of weak AI — meaning that it has a narrow focus and is only capable of performing very specialized tasks. As of now, there is no form of conversational AI, or AI in general, that has achieved total generalized intelligence, or the ability to learn, behave and perform actions the same way we humans do. And that is because these systems are continuing to be trained on information only, which is a “very two-dimensional way to learn about the universe,” Bradley said.

“You don’t just learn about gravity by reading about gravity. In fact that’s kind of the last way you learn about it as a human. You learn about it first by experiencing it — like falling down,” he said. “Once we teach these things in different modes I think [it] will create a much more robust kind of intelligence. And they’ll probably start to be more and more useful very quickly.”

A Look at the Bigger Picture The Future of AI: How Artificial Intelligence Will Change the World


How Does Conversational AI Work?

Replicating human communication with AI is an immensely complicated thing to do. After all, there’s a lot involved in even a simple conversation between two people, it’s not just the logical processing of words. Rather, it’s an intricate balancing act involving the context of the conversation, the people’s understanding of each other and their backgrounds, their verbal and physical cues — it has just as much to do with the things surrounding words used than the actual words themselves.

Mimicking this kind of interaction with artificial intelligence requires a combination of both machine learning and natural language processing. Machine learning is a process made up of a set of algorithms, features (individual variables) and data sets that teaches a computer how to do certain tasks. As the input grows, the machine gets better at recognizing patterns and making predictions, thus improving its ability to perform that certain task. Meanwhile, natural language processing teaches computers how to understand language, conversations and speech.

Natural language processing consists of four steps: input generation, input analysis, output generation and reinforcement learning. During input generation, users provide a prompt (either voice or text) through a conversational AI website or app. If the prompt is text-based, the AI will use natural language understanding, a subset of natural language processing, to analyze the meaning of the prompt and derive its intention. If the prompt is speech-based, it will use a combination of automated speech recognition and natural language understanding to analyze the input. Then comes dialogue management, which is when natural language generation (a component of natural language processing) formulates a response to the prompt. Finally, through machine learning, the conversational AI will be able to refine and improve its response and performance over time, which is known as reinforcement learning.

The Building Blocks of Conversational AI

  • Machine learning: A subset of artificial intelligence that uses data, statistics and trial and error to learn and optimize a specific task without having to be specifically coded for that task.
  • Natural language processing: A branch of artificial intelligence that teaches computers how to understand, interpret and manipulate human language.
  • Deep learning: A more sophisticated type of machine learning dedicated to training computers to discern information from data. Although deep learning models require massive volumes of data to train them, they require very little human intervention once they have been created.

Today, natural language processing is achieved with the help of machine learning. But it started with computational linguistics and statistical linguistics. This involved techniques like teaching sets of words that tend to be predictive of an outcome, and using statistical analysis to figure out what a given conversation means and what the right thing to do with the information is.

In the late 2000s this technique gave way to the process of creating vectors, or sequences of numbers, out of words, according to Bradley. This allows engineers to take a bunch of data and condense it into numerical form, which can then be used to capture the semantics of a given statement or conversation. Similar words relate to each other mathematically in a way that an AI system can start to parse out what a statement is actually saying.

Then, about a decade ago, the industry saw more advancements in deep learning, a more sophisticated type of machine learning that trains computers to discern information from complex data sources. This further extended the mathematization of words, allowing conversational AI models to learn those mathematical representations much more naturally by way of user intent and slots needed to fulfill that intent. For example, if a person is using a chatbot to book an airline ticket, their intent is to purchase a ticket. The AI system then needs to know what airline they are trying to fly out of, for what day, and so on. 

Bradley said every conversational AI system today relies on things like intent, as well as concepts like entity recognition and dialogue management, which essentially turns what an AI system wants to do into natural language. And in the future, deep learning will advance the natural language processing abilities of conversational AI even further.

Normandin attributes conversational AI’s recent meteoric rise in the public conversation to a number of recent “technological breakthroughs” on various fronts, beginning with deep learning. Everything related to deep neural networks and related aspects of deep learning have led to major improvements on speech recognition accuracy, text-to-speech accuracy and natural language understanding accuracy.

“You can train a natural language understanding system with a lot less examples today than you could earlier, and that’s really spearheaded the chatbot revolution that we’ve seen,” he said. “There is an acceleration of what technologies are able to do in terms of having a conversation with human users. Either open-ended conversations, or what we see most often in enterprise settings is domain-specific conversation to address tasks or questions that users might have.”

Learn the Basics Artificial Intelligence vs. Machine Learning vs. Deep Learning: What’s the Difference?


How Is Conversational AI Used?

Despite its narrow nature, the use of conversational AI is quite widespread across a variety of industries, from fashion to HR to the Internet of Things. According to market research firm Insider Intelligence, the top three uses of chatbots in the United States today are for business hours, product information and customer service requests.

Indeed, a familiar use case is virtual call center agents for customer support, which is what Normandin’s company Waterfield Tech handles. Just as some companies have web designers or UX designers, Waterfield employs a team of conversation designers that are able to craft a dialogue according to a specific task. Usually, this involves automating customer support-related calls, crafting a conversational AI system that can accomplish the same task that a human call agent can.

Eventually, as this technology continues to evolve and grow more sophisticated, Normandin anticipates that virtual call agents will be treated similarly to their human counterparts in terms of their training and oversight. Rather than handcrafting automated conversations like they do right now, these bots will already know what to do. And they’ll have to be continuously supervised in order to catch mistakes, and coached so they don’t make those mistakes again. However, this requires that companies get comfortable with some loss of control.

“When you let an AI go out and talk to customers on your behalf, once in a while they may not always say the right thing. Same way that a human does,” Normandin said. “It’s a total change of perspective. But it’s going to be necessary to make the leaps ahead that we need to do in order to get to really large-scale automation.”

Conversational AI Industry Uses Cases

  • Fashion and retail chatbots
  • Automated HR processes like employee onboarding and training
  • Smart devices and other Internet of Things tech
  • Customer support call centers
  • Automated social media interactions
  • Chatbots to handle administrative tasks at healthcare facilities

Meanwhile, before being acquired by Hootsuite in 2021, Heyday focused mainly on creating conversational AI products in retail, which would handle customer service questions regarding things like store locations and item returns. Now that it operates under Hootsuite, the Heyday product also focuses on facilitating automated interactions between brands and customers on social media specifically. Incidentally, the more public-facing arena of social media has set a higher bar for Heyday.

“We’re moving away from private conversations to being able to do that in public,” Christine Dupuis, Heyday’s senior director of product and AI, told Built In. “If a chatbot is answering publicly there’s a lot more scrutiny. It’s a higher bar to be in the social media space.”

Elsewhere, companies are using conversational AI to streamline their HR processes, automating everything from onboarding to employee training. The healthcare industry has also adopted the use of chatbots in order to handle administrative tasks, giving human employees more time to actually handle the care of patients. Some even go so far as to make medical diagnoses themselves. Many companies are also looking to chatbots as a way to offer more accessible online experiences to people, particularly those who use assistive technology. Commonly used features of conversational AI are text-to-speech dictation and language translation.

Food for Thought Will We Ever See a Real-Life ‘Star Trek’ Universal Translator?


Examples of Conversational AI

Of course, the conversational AI example on everyone’s minds right now is ChatGPT, which has already begun taking over everything from customer service to medical diagnosis. There have been other iterations of ChatGPT in the past, including GPT-3 — all of which made waves when they were first announced. But ChatGPT is a true trailblazer.

“ChatGPT is actually different in the sense that it was actually optimized on conversation data and also, interestingly enough, on a lot of software code,” Normandin said. “This is actually going to be, I think, a game changer in the way conversational AI systems are developed.”

“This is a marvel in terms of technology, it’s almost magical.”

That being said, there are several major issues with ChatGPT. Despite all of its impressive capabilities, this technology is nowhere near perfect, and it regularly generates false information, among other issues. It’s kind of a “black box,” Normandin said. “You have very little control over what it’s going to do.” 

Plus, training a system of ChatGPT’s size and complexity comes with quite a big price tag — Normandin estimated it cost dozens of millions of dollars in CPU time. So it’s really only something large companies like Google, Amazon, IBM and, of course, OpenAI can afford to do. At least for the time being.

“This is a marvel in terms of technology, it’s almost magical. But the real challenge will be how to leverage that technology to build the next generation of conversational AI,” Normandin said. “I think it will require a total rethinking of how these systems are developed in the future.”

And OpenAI is not alone. There are many other companies revolutionizing the conversational AI space. Here are just a handful of them.

Want More OpenAI Content? OpenAI Captures the Magic of AI With ChatGPT and DALL-E


IBM — Watson Assistant

IBM’s Watson computer first made headlines when it played a game of Jeopardy! in 2011. Running software called DeepQA, Watson had been fed an immense amount of data from encyclopedias and open-source projects for a few years before the match — and then managed to win against two top competitors.

Today, Watson has many offerings, including Watson Assistant, a cloud-based customer care chatbot. The bot relies on natural language understanding, natural language processing and machine learning in order to better understand questions, automate the search for the best answers and adequately complete a user’s intended action. It can also be integrated with a company’s CRM and back-end systems, enabling them to easily track a user’s journey and share insights for future improvement.


Google — Dialogflow

Dialogflow helps companies build their own enterprise chatbots for web, social media and voice assistants. The platform’s machine learning system implements natural language understanding in order to recognize a user’s intent and extract important information such as times, dates and numbers. And its state-based data models are advanced enough to reuse intents, intuitively define transitions and data conditions and even handle supplemental questions, allowing customers to deviate from the main topic and return to it again without any confusion.

Once they are built, these chatbots and voice assistants can be implemented anywhere, from contact centers to websites. They can also be seamlessly integrated with agents across platforms along with telephony partners like Genesys, Avaya and Cisco.


Amazon — Alexa

When it comes to virtual assistants, Alexa is a household name. With Alexa smart home devices, users can play games, turn off the lights, find out the weather, shop for groceries and more — all with nothing more than their voice. What’s more, Alexa talks back. It knows your name, can tell jokes and will answer personal questions if you ask it all thanks to its natural language understanding and speech recognition capabilities.

And nearly a decade after its initial launch, Alexa is still learning and improving its conversational AI capabilities. As it continues to gather up data and refine itself, it constantly strives to sound more naturally human.


The Benefits of Conversational AI

Now that conversational AI has gotten more sophisticated, its many benefits have become clear to businesses. 

The primary pro to implementing this technology is its cost efficiency. For instance, when it comes to customer service and call centers, human agents can cost quite a bit of money to employ. Automating some or all of their work can improve a business’s bottom line. It can also help in labor shortages, which have worsened in the wake of the Covid-19 pandemic.

“Finding people and training them and keeping them has become a real nightmare for many companies,” Waterfield Tech’s Normandin said. “We’re actually addressing a very serious staffing issue with these automated agents. And once you’ve trained an automated agent, you can replicate them as many times as you want.”

The Benefits of Conversational AI

  • Increases sales
  • Improves customer engagement
  • Scalable
  • Increases business flexibility (no more language or time barriers) 
  • Efficient and quick

Conversational AI also stands to improve customer engagement in general, particularly in customer service and other consumer-facing industries. With chatbots, questions can be answered virtually instantaneously, no matter the time of day or language spoken.

“As consumers, people expect really fast answers. Nobody wants to wait on the phone line for one hour or two hours so speak to somebody. The chat is really instant,” Heyday’s Dupuis said. This is “really important” for brands, both in terms of customer relations and business efficiency. With chatbots, companies can be available all over the world all at once, speaking any language necessary. “It makes the brand feel more accessible,” she continued. “It’s a scalable way of providing really good customer service.”

Plus, the conversational AI space has come a long way in making its bots and assistants sound more natural and human-like, which can greatly improve a person’s interaction with it. 

“If I’m interacting more naturally with a machine, I’ll get more out of it,” Dupuis said. “Getting closer to full sentences and speaking the way I speak, I think that has value.”

Take it From the Experts Relationship to Revenue: How Conversational AI Can Supercharge Customer Success


The Challenges of Conversational AI

Still, it’s important to remember that a lot of challenges are yet to be overcome in this space. 

For one, conversational AI still doesn’t understand everything, with language input being one of the bigger pain points. With voice inputs, dialects, accents and background noise can all affect an AI’s understanding and output. And with text, slang and unscripted language can do the same. In the end, humans have a certain way of talking that is immensely hard to teach a non-sentient computer. Emotions, tone and sarcasm all make it difficult for conversational AI to interpret intended user meaning and respond appropriately and accurately.

The Challenges of Conversational AI

  • Doesn’t always understand or interpret inputs correctly, which can lead to inaccurate or unhelpful responses
  • Different dialects or accents can affect performance
  • Like all AI, the algorithms that make up conversational AI are made by humans, which means there’s a chance for bias
  • Most have a very specialized or narrow focus, which can fall short of what a user needs
  • Inaccurate responses can lead to a general distrust of the tech

Even if it does manage to understand what a person is trying to ask it, that doesn’t always mean the machine will produce the correct answer — “it’s not 100 percent accurate 100 percent of the time,” as Dupuis put it. And when a chatbot or voice assistant gets something wrong, that inevitably has a bad impact on people’s trust in this technology.

“If I’m talking to a chatbot and that experience doesn’t meet my very high standard for my expectations about communication with a human, then it’s very jarring to us,” Bradley of LivePerson said. “You can have all these things set up, but as soon as that bot makes a stupid mistake and reveals itself as this kind of relatively dumb computing thing, your feet fall out from under you.”

Dupuis said she and her team combat these negative experiences with chatbots by having its conversational AI really lean into its identity as a bot, having it say things like “Sorry, I’m just a robot. Let me transfer you to a human, they’ll be better equipped to answer your questions.” They also work to make it clear to customers how they should be interacting with the chatbot, whether that be pre-set text options or prompting a free-text dialogue with questions like “How can I help you?”

But for conversational AI to really be improved and widely adopted in the future, Dupuis thinks it will need to become more “context-based and proactive,” where a system will be able to anticipate a user’s needs and future questions before they’re even asked.

“Once you can detect and codify a little bit of what the conversation looks like from a machine learning perspective, you can learn on top of that the overall pattern of what people are asking and what people will ask next,” she said. “Understanding the broader context and offering more proactive suggestions, I think we’re going towards that more and more.”

Great Companies Need Great People. That's Where We Come In.

Recruit With Us