What is conversational AI?

‍Let’s start by breaking it down. According to the Oxford Learner’s Dictionary, conversational is an adjective that means “not formal; as used in conversation.” AI is short for artificial intelligence,  “the study and development of computer systems that can copy intelligent human behavior.” When we put the two together, we get: A computer system that participates in a conversation by responding to and asking questions in a human-like manner.

‍Can we make it even simpler? Of course– It’s basically AI for language.  ‍

First, let’s agree on some vocabulary

‍Before jumping into the details of conversational AI, there’s some vocabulary that we need to cover to make sure we’re speaking the same language.

‍Machine learning (ML):
The process of adding information to a model that a computer learns from by recognizing patterns within the data
‍
‍Intent:
‍The intention of the user; The intention of the user, the reason the user is getting in touch

‍Expression:
‍A way to express the intent; linguistic variations of the intent

‍Entity:
‍A “keyword” that helps the NLP identify the most important elements of an intent

‍NLP model:
The AI model that we use for language, that’s trained using intents and expressions

‍Voice/chatbot:
An automated conversation partner; the user interface that lies between the user and the NLP model

‍Confidence scores:
From 0-100%, how sure the NLP model is that it correctly recognizes an intent

‍Natural language processing (NLP):
A branch of AI that uses rule-based logic and machine learning to read and recognize language, and give or create a response; it encompasses NLU and NLG

‍Natural language understanding (NLU):
A subset of NLP, it’s a computer’s ability to understand natural human language and contextual nuances and assign confidence scores

‍Natural language generation (NLG):
A subset of NLP, it’s a computer’s ability to generate a natural “human” response by following a set of rules and templates based on the user’s input

‍Step-by-step: How conversational AI works

To work properly, conversational AI requires a conversation. This is where natural language processing comes into play– because the process allows computers to participate in the conversation. According to IBM, we can break down the NLP process into 4 steps:

‍

Step 1: Input generation

This is what the user sends (or inputs) to the bot (NLP model). In conversational AI, the input is always written (text) or spoken (speech) language.

‍Step 2: Input analysis

‍The NLU analyzes the user’s input and assigns confidence scores to different intents.

‍Step 3: Dialogue management

Based on the confidence scores, the NLP determines which predefined response to send or uses NLG to create and send an appropriate response.

‍Step 4: Reinforcement learning

Use machine learning to improve the performance of the bot’s NLP model over time.
‍

‍3 different approaches to conversational AI

‍Intent-based AI

An approach to AI where you define which questions (intents) the bot SHOULD answer. A chatbot with an intent-based approach uses AI to recognize what the question is and complete a pre-defined action response.
👍 Control the entire scope
👍 Define the exact responses

👎 High barrier to entry; requires an initial investment (expertise, time)
👎 Limited scope
👎 Difficult to scale‍

Generative AI

An approach to AI where you define which questions (intents) the bot SHOULDN’T answer. Generative AI automatically will try to answer every question and provide a dyanmic response, so your role is to identify the questions that you want to block the AI from answering.
👍 Low barrier to entry; can start immediately
👍 Personalized, unique conversations
👍 Broad scope
👍 Easy to scale

👎 Lose control over what the bot says
👎 Known to go off-script and fabricate incorrect responses

‍Hybrid AI

The best of both worlds! You combine intent-based and generative AI to leverage the strengths and reduce the weaknesses of each approach.
👍 Control the topics that you want/need to control
👍 Easier to scale
👍 Broad scope
👍 Create personalized responses when appropriate‍

When should I use conversational AI?

‍To figure out if conversational AI is the right choice for you, use the trifecta method. In short, this means identifying intents (flows or questions) within your use case (situation) that are easy to recognize, easy to answer, and repetitive. If they check all three of these boxes, then they can be automated with AI.  For the fine details  on scoping your next chatbot project, check out this article.
‍

Our favorite examples of conversational AI

In the Best Bots of Belgium 2022, we compiled a list of the conversational AI community’s favorite Belgian bots. Then, we tested each on the bot's ability to quickly and easily answer users’ questions.

The most important characteristics of a good conversational AI chatbot are natural language quality, conversational design, structure and flows, and visualization. Based on these characteristics, our favorite chatbots are BRUce from Brussels Airport, Thalys, and NMBS/SNCB.‍

BRUce

BRUce is the perfect travel buddy when navigating Brussels Airport. He’s easy to connect with online, via WhatsApp, or on Facebook Messenger. If you’re a traveler who would rather grab breakfast and read a book, instead of anxiously standing in front of the gate info screens, BRUce can help you find your flight and subscribe you to automatic updates via WhatsApp. He’ll let you know as soon as you’re assigned a check-in desk, a departure gate or when boarding starts.  

Not only can he answer any basic questions you have roaming around the airport, but he offers features that set him apart from the competition. We’re talking about sleek visuals, conversational copy (multiple bubbles with short responses), multilingual options, and integration with real-time information.

NMBS/SNCB

Stuck on a train platform with no idea where to find your connection? NMBS/SNCB’s Mobi has you covered. Connect with the virtual chatbot through WhatsApp, Facebook, and the SNCB website to answer any and all of your urgent matters from “where is my train?” to “plan a journey.”

And don’t worry about speaking English– Mobi is multilingual in English, French, and Dutch. Type anything in one of these three languages and the conversation will flow from there. If Mobi’s initial prompt and button options aren’t enough, ask to speak to an agent for an immediate connection.

Eurostar bot (formerly ThalysBot)

Not only does the Eurostar bot (formerly ThalysBot) look nice, but it’s one of the most expansive chatbots out there. No need to use your keyboard– instead refer to one of Eurostar bot’s many button options to answer exactly which question you have. With the click of a button, you will either have a perfect response or a link to a webpage that can provide whatever information you’re looking for.

The Eurostar bot is personalized– meaning after logging in, it caters to your Thalys (Eurostar) account information. This integration means that you don’t have to wait around and confirm your reservation status or claim with a human! We appreciate ThalysBot for introducing itself as a bot, too, preparing for any upcoming EU chatbot laws.