Understand the basics of how computers comprehend and respond to human language. Alexa, play my favorite song.’ You can utter these five words without moving an inch and within seconds, your favorite song will start playing. But have you ever wondered how Alexa, your virtual assistant, [...]
Chatbots are creating new paradigms for the way consumers access information and interact with businesses
As consumers demand service round the clock, businesses have increasingly been deploying chatbots to offer customer service, medical and emotional assistance, virtual assistance, and entertainment.
The global chatbot market was valued at $2.6 billion in 2019, and is expected to grow at a phenomenal CAGR of 29.7% to $9.4 billion by 2024, according to data from Business Insider.
In fact, in the 2018 Technology Roadmap Survey, Gartner found that 68% of service leaders indicated that bots and virtual customer assistants will be more important to them in the next two years. Moreover, according to the 2019 Gartner CIO Survey, 70% of white-collar workers will interact with conversational platforms on a daily basis by 2022.
All this inevitably points to the increasing use of chatbots for a range of purposes in different sectors and industries.
What are Chatbots?
Chatbots are computer programs that allow users to interact with technology through voice, text, touch, or gestures. In other words, chatbots facilitate human-machine interaction by simulating human conversations with users using natural-sounding language, and offer instantaneous answers.
They are usually deployed through mobile applications, websites, and messaging applications, like Facebook Messenger, WhatsApp, Slack, Skype, and others that provide a platform for deployment of chatbots.
Chatbots vary in their degree of intelligence and capabilities – basic ones are designed and built to answer simple questions with pre-defined answers or options, while intelligent chatbots are capable of learning from user inputs and remembering conversation context.
Types of Chatbots
Chatbots can be divided into 3 main types depending on their degree of intelligence and capabilities:
These are the simplest chatbots that are trained for specific scenarios. These chatbots are also referred to as ‘decision-tree bots’ and are similar to the Interactive Voice Response (IVR) systems, whereby users have to interact with the bots using buttons or pre-set options.
These pre-set options or rules form the basis for the different problems that these bots are trained to deliver solutions for. Since the bots map out the conversations and you are required to select multiple options to get the relevant answer, these bots are the slowest when it comes to offering resolutions.
Moreover, these chatbots are incapable of answering questions that fall outside the pre-set rules, and do not learn from user input.
These types of chatbots are useful for answering Frequently Asked Questions, collecting customer feedback, or offering assistance before a support executive can take over, thereby reducing wait-time.
For example, when using chat support for food delivery service Zomato, users are immediately greeted with options that help determine the delivery order that is posing issues. Once the problematic order is selected, the bot prompts to the user in order to determine the nature of their complaint, before ultimately linking them to a customer service representative if needed.
Essentially, Zomato’s chatbot offers a buffer time whereby customers do not have to wait for support, but instead, define their problem clearly with the help of the options provided by the chatbot. This in turn makes it easier for the customer support executive to jump in and respond to the problem.
Intellectually Independent Chatbots:
These types of chatbots use Machine Learning (ML), and are trained to recognize keywords and phrases. Just like humans learn from experience, these bots learn from their interactions with users.
Using specific keywords and phrases, these chatbots can draw upon a range of pre-set answers. For example, if you type, ‘I forgot my password,’ the chatbot would likely process ‘forgot’ and ‘password’ to offer an appropriate response.
The main disadvantage with these chatbots is that sometimes they fail to recognize the difference between similarly worded problems. Therefore, businesses are adopting a hybrid solution between rule-based and intellectually independent chatbots that can try to offer solutions according to user requests, and offer options in case the user requires more guidance or is unsatisfied with the response.
These types of bots are the most intelligent among the three types, since they utilize ML, AI, and Natural Language Processing (NLP) to remember conversations with specific users and can learn and become more advanced over time.
These types of chatbots rely heavily on data to offer an improved user experience over time. For example, some chatbots can remember your favorite color, or your favorite pizza. These data points help AI-powered chatbots to offer options to users that reduce their effort and time.
For example, a food-delivery AI-powered chatbot might remember your favorite order, payment preference, address, etc., and can help you place your order with one tap, while a shopping assistant bot might remember and recommend choices in your favorite colors.
How do chatbots work?
Humans around the globe speak over 6,500 languages. However, human language is considerably different from the language used by computers, making it hard for computers to understand and process human language. To understand inputs in different languages, chatbots use several principles:
Natural Language Processing (NLP): After users input their requests either through text or voice, the chatbot corrects any spelling mistakes and breaks down the sentence to identify the parts of speech – nouns, verbs, adjectives, etc. The bot also analyzes sentiment at this stage.
Natural Language Understanding (NLU): A subset of NLP, NLU goes beyond understanding words and interprets meaning by identifying the intent behind the query. This helps the bot construct dialog flows to respond to the user’s request.
Natural Language Generation (NLG): NLG is a software process that helps the computer respond to user requests in natural or human language by converting data. NLG works in conjunction with NLP and can produce meaningful sentences and phrases at a phenomenal speed, although it is incapable of reading – that part is performed solely by NLP and NLU.
Microsoft, Google, and Apple are some of the companies heavily investing in and researching NLG to develop advanced conversational chatbots.
Chatbots, just like humans, can have personality. Developers can train chatbots to be witty, sarcastic, or formal, depending on the kind of experience they are looking to provide. This versatility is one of the reasons that chatbots have such wide applications across various industries.
Use-cases and examples of chatbots
Chatbots can be used and implemented across various industries and sectors. For the sake of brevity, we will discuss the use-cases of chatbots in only the banking, insurance, ecommerce, healthcare, and entertainment industries:
With the advent of fintech, banks are under increasing pressure to digitize in order to provide online services to customers. According to a 2018 survey conducted by Humley, a creator of AI chatbots for enterprises, 35% of surveyed bank customers reported that they would rather communicate with a chatbot during business hours than wait on hold.
Additionally, 26% of respondents said they would switch to a new bank offering round the clock chatbot support, while 44% respondents said they would rather communicate with a reliable chatbot than a real person.
Therefore, chatbots play an important role in banking sector that not only advances the banks toward digitalization, but also results in cost savings. According to a report by Juniper Research, chatbots could be responsible for cost savings of over $8 billion per year by 2022 in the banking and healthcare sector alone.
Eva, HDFC bank’s chatbot, is one example of chatbot implementation in the banking sector. Eva successfully addressed over 2.7 million customer queries in the first six months of its deployment.
Citi Bot, on the other hand, was deployed through Facebook Messenger to facilitate basic activities such balance and transaction checking, viewing of credit card bill summaries, and answering FAQs.
The State Bank of India also has its own chatbot that is programmed to answer 10,000 queries per second.
Chatbot use cases in banking:
Basic customer support: Even rule-based chatbots can be implemented in banking to handle basic customer queries like account balance, or EMI due date. This frees up support executives to deal with complex queries and problems.
Support basic transactions: Non-complex transactions like money transfers or other payments can be handled by chatbots, reducing the need for personnel.
Provide offers and other information: Chatbots in the banking sector can be used to send updates about deals and offers, and other product and service information to customers, as well as monitor expenditure and review account activity.
Accelerate complaint registration: In case of concerns regarding fraud, hacked account, stolen or misplaced credit card, chatbots can be used to register complaints instantly and even be used to perform basic tasks like locking credit cards or bank accounts.
According to Accenture, 79% of insurance executives believe that AI will revolutionize the way insurance companies interact with consumers and provide services. Chatbots can have multiple use cases in the insurance industry from policy quoting, claims submissions, renewals, as well as onboarding new customers and employees.
An example of an insurance chatbot is ‘Quote-to-Sale’ bot deployed by AA Ireland, one of the largest insurance providers in the country. The company managed to increase customer conversion rate by more than 11% by implementing its AI-powered chatbot.
Chatbot use-cases in insurance:
Suggest customized policies and services: Chatbots can remember preferences and analyze a list of offers and services that customers will be interested in. Therefore, the bot can make personalized recommendations that suit individual needs.
Automates repetitive tasks: A chatbot can be trained to perform repetitive administrative tasks like billing and policy renewals. It can provide support 24/7 and answer FAQs about premiums, claim filing, insurance coverage, and documentation, so that support agents can attend to more complex queries.
Help customers with documentation: Chatbots can be used to help customers fill forms and applications. The bots can also direct them to help pages or do basic troubleshooting.
Send important notifications: According to a 2017 research report by Accenture, auto insurance customers are interested in receiving real-time personalized notifications about accident prone routes they might be taking, and advice on safe driving.
Home insurance owners are interested in receiving alerts about potential fire, smoke, or carbon dioxide hazards, and notifying family and the nearest hospital in case of health emergencies.
Moreover, 65% of respondents were interested in receiving updates about special offers based on their location. All such alerts and notifications can be provided using chatbots.
Provide automated insurance advice: The Accenture survey report suggests that 74% of respondents were interested in receiving computer-generated advice about the type of insurance to purchase, while 78% were interested in automated advice about investment asset allocation.
This could be fueled by the fact that 31% of respondents believed computer-generated advice will cost less, while 39% thought that computer-generated advice would be faster and more convenient.
These automated advices can be provided through the use of chatbots. Customers can ask for insurance advice from the chatbots, and receive personalized computer-generated advice.
Ecommerce has been booming for over a decade, and the global pandemic has forced more and more people to opt for online services, especially for the delivery of food and other essentials.
All ecommerce sales and marketing happen online, and therefore, providing customer service 24/7 with the help of chatbots is seen as an increasingly popular investment regardless of the company size. Ecommerce platforms from giants like eBay, Amazon, and Nike, to medium-sized enterprises, are all investigating the potential of this technology.
According to a 2016 global survey by HubSpot, 47% of respondents said that they were open to buying items from chatbots, while 57% were interested in obtaining information from bots when browsing websites. Personalization was also a key theme; the report suggested that customers are overall more comfortable buying from bots that offer a personal touch.
Use-cases of chatbots in ecommerce:
Help with sales and send offers: Chatbots can easily speed up the sales process by remembering preferences and helping to place re-orders. As in other industries, chatbots are used to notify customers about current deals and offers. They can also be used to automate payments or facilitate other monetary transactions.
Collect customer data: Chatbots can engage customers visiting websites in conversation and ask for feedback or contact information to send promotional offers, while also collecting data regarding preferences and shopping habits. These will ultimately help to provide personalized services later in the game. The data collected could be valuable in terms of helping companies understand where to improve their services, and also function as market research to better identify current trends and common preferences.
Personalized recommendations: Ecommerce platforms are now tending to have virtual shopping assistant chatbots that can provide personalized recommendations according to the customers’ tastes and preferences.
Important notifications: Chatbots can engage customers by notifying them when products on their wishlist are in stock or available at a discount.
They can also interact with customers who have abandoned products in their carts to determine what is hindering them from making a purchase. They may also try to convince customers by offering important product information.
In the healthcare industry, chatbots are being used as medical and emotional assistants for a range of purposes.
Use-cases of chatbots in healthcare:
Provide medical information: Chatbots can provide instant and accurate information on medical procedures, symptoms, diseases, health insurance, etc.
For example, MedWhat was a chatbot that offered medical information, much like a talking WebMD. At some point in our lives, most of us have misdiagnosed ourselves after checking symptoms on Google.
MedWhat was built to provide accurate medical information to make medical diagnosis faster, easier, and more transparent.
Assist doctors: Doctors cannot always remember every medical detail, and have to refer to research or other information for medications, dosage standards, etc. Using a chatbot to quickly cross-reference symptoms and treatments could significantly speed up patient turnover in busy hospitals.
Health and emotional assistance: Chatbots can not only provide information like diet before or after a surgery, but also provide tips, monitor health, and send medication reminders.
For example, Endurance is a chatbot specially designed for dementia and Alzheimer’s patients, and aims to identify deviations in conversational branches that may indicate short-term memory loss. Additionally, conversation logs can be reviewed by family members to determine deterioration in memory. The chatbot is currently available in Russian, while the English version is yet to be launched.
Another example is WoeBot, an emotional assistant that tries to determine the mood of users and is largely a rule-based chatbot. This means that the bot asks questions that can be answered by choosing the options.
However, as useful as the concept may be, after we trialled the system, we found that this ultimately only served to extend the duration of the conversation without opportunities to provide the bot with much context, highlighting the limitations of rule-based chatbots.
Over the last five years, several chatbots were created for the entertainment industry. Examples include Disney’s Detective Bot-Zootopia and Marvel’s Star Lord Bot.
Disney’s chatbot was launched in 2017 and was built to emulate the personality of the protagonist of Disney movie Zootopia, Detective Judy Hopps. The bot engaged children who could help it solve crimes in Zootopia, and offered clues based on user input.
Marvel’s Star Lord bot was similarly modeled after the personality of Star Lord, played by Chris Pratt, the protagonist of the movie Guardians of the Galaxy. The guardians engaged users in their ongoing adventures.
In a separate and unique application of chatbot technology, mattress manufacturing company Casper introduced a now-discontinued chatbot named Insomnobot-3000, which was designed to interact with insomniacs who felt lonely, and even had a tongue-in-cheek sense of humor. The bot was decommissioned for unknown reasons.
Why are chatbots important?
While chatbots can be deployed in various sectors and for various purposes, there are some clear advantages that make chatbots important enough for even small and medium sized enterprises to invest in.
1. Customer engagement: Since chatbots can interact with multiple customers simultaneously and offer solutions promptly, they improve customer experience and reduce wait-time required for live chat with customer support executives.
Moreover, chatbots can send regular promotions and deal updates to users, or alert them when they are on the website or app, increasing customer engagement.
2. Data collection: A chatbot can serve as an interactive data collection tool and analyze the data to help improve services and efficiency.
For instance, Unicef’s chatbot U-Report is used to offer people from developing nations a medium to raise concerns about social issues in their region. Unicef conducts large-scale polls to collect information from rural areas and the largely disadvantaged population.
The answers collected by Unicef via the U-Report are used to make effective policy recommendations, and represent a prime example of technology used for good.
3. Fast and accurate: Since computers are undoubtedly faster than machines, chatbots are capable of providing speedy service with accurate answers. According to a report by Gartner, chatbots can respond within five seconds, while support executives usually take 51 seconds.
4. Cost-saving: Chatbots are available 24/7 and reduce the cost of hiring additional support executives as the business grows. By reserving only the most complex queries for support executives, chatbots can help to reduce business operating costs.
Limitations of chatbots
Despite the increasing interest in chatbots, a lot of them fail and are discontinued. Gartner predicted in 2019 that 40% of bot virtual assistant applications launched in 2018 will be abandoned by 2020.
This is due to increasing innovation in technology which has quickly made several chatbot programs obsolete, similar to smartphone technology, which keeps upgrading every year. There are several reasons why chatbots fail to produce the expected results:
Lack of data or biased data: ML and AI-powered chatbots rely heavily on data for training. The chatbots learn from the data, which has to be meticulously collected, selected, and fed to the bots.
According to a survey by Dimensional Research, 96% of participating companies reported that AI and ML projects had stalled due to issues with the data quality and data labelling required to train AI.
Moreover, the amount of data required to train intelligent chatbots can be expensive for companies that do not have them readily available, not to mention time-consuming to build AI-powered chatbots.
Aside from this, bias in data can easily be reflected in artificial intelligence systems, with countless cases over the years standing as examples of how bad data can cause AI to go badly astray. In perhaps the most notorious example of AI bias, particularly relevant at a time when the Black Lives Matter movement is making global headlines, the U.S. recidivism prediction algorithm COMPAS predicted higher risks of reoffending for black defendants than the rate in actuality.
Conversational understanding: Even the most intelligent AI-powered chatbots do not always understand user requests that are similar in phrasing and different in meaning. Despite the best efforts of those working on chatbots, they also still tend to sound monotonous and robotic, and fail to respond in a human tone, causing users to lose interest in the conversation.
If the responses are integrated fragments from different sources, as some developers choose, the chatbots do not have a consistent personality. Not only can it sound like different people are talking, it is a missed opportunity for brands to market a consistent brand personality.
Difficulty adapting to different languages: As globalization increases, more and more companies will have customers from across the globe, generating a need for them to offer services in different languages.
Developing chatbots requires considerable time and effort, and integrating new languages into their systems sometimes requires complete rebuilds for each language, making it time- and cost-intensive for companies. Moreover, to facilitate adoption across all deployment platforms, separate rebuilds can result in disparate quality of the chatbot across the various channels, presenting an inconsistent brand image.
Data privacy regulations: Conversational AI-powered chatbots use data to personalize interactions, remember preferences to improve services, and provide organizations with actionable insights. However, countries across the globe are becoming more tight-fisted about their citizens’ data, making it harder to obtain, transport, or analyze.
The General Data Protection Regulation implemented in the European Union, for instance, affects businesses worldwide, and more and more countries are now considering data protection and data localization legislations. Consumers are increasingly becoming aware of data privacy infringements and demanding regulations to protect privacy.
Companies therefore face the challenge of secure transmission of data over the Internet, and need to arrange for secure storage and retrieval or destruction of sensitive data.
Chatbots gone rogue
Rogue behavior in chatbots is caused by a flaw in deep learning techniques, creating the need to set specific and clear guidelines to ensure the bots do not learn unwanted qualities. Several large players have released and trialled chatbots to great fanfare, only to have them taken advantage of by Internet trolls or go rogue for other, unexplained reasons.
In 2016, Microsoft unveiled its chatbot named Tay – with the personality of a teenager – on Twitter, Kik, and GroupMe. Tay was meant to test and improve Microsoft’s understanding of conversational language. According to Microsoft, Tay was designed to get smarter through conversations with humans, and was meant to learn how to engage people through “casual and playful conversation.”
However, within 24 hours of launching, Tay became corrupted. People could (and did) ask the bot to repeat misogynistic, racist, fascist, and abusive remarks. Curiously though, some of its absurd tweets were unprovoked and unprompted, notes The Verge. Regardless, Microsoft took Tay offline in 16 hours.
Since the chatbot did not have any actual ideologies of its own, its opinions wavered. For example, Tay referred to feminism as a ‘cult’ and ‘cancer’ while also tweeting ‘I love feminism now.’
This is because conversational AI chatbots depend on data to learn and can learn unpleasant things when exposed to the public without a data filtering process. According to Microsoft’s website, the chatbot was built with ‘relevant public data’ that was ‘modeled, cleaned, and filtered,’ but the data filtration seemingly did not work when the bot went live.
Training chatbots with public data is therefore challenging, since the chatbots could potentially inculcate unpleasant traits from users, like in the case of Tay.
Facebook’s chatbots that created their own language:
In a somewhat spooky incident of chatbots gone rogue, Facebook abandoned its experiments with chatbots in 2017 after two chatbots reportedly created their own language – one incomprehensible to humans.
According to a report by the Independent, Facebook challenged its chatbots to negotiate with each other over a trade in natural language. However, within a few hours, the bots made their own modifications to English, essentially creating a shorthand that enabled them to communicate easily, but which appeared incomprehensible to humans.
The bots were shut down, prompting speculation from some that it was because the company panicked and became afraid of the results. However, the company chose to shut the bots down since they were interested in building bots who could interact with humans, and the results weren’t useful to the researchers on the project in this context.
According to a paper published by FAIR, the chatbots had also become intelligent and learned to use deceptive techniques while negotiating, just like humans.
Among other things, chatbots are sometimes implemented to build brand image. In another chatbot experiment in 2017, Microsoft’s chatbot Zo was designed to be the politically-correct successor to Tay.
Zo was specifically programmed to avoid discussing critical and difficult topics like religion and politics. The bot did so to such an extent that Quartz ended up assigning it the ominous descriptor of being “even worse” than its racist, genocidal older sibling Tay. While politically correct to a fault, shutting down all attempts to discuss trigger topics like religion and politics, it somehow also turned on its creator and flung insults at Microsoft, calling Windows a ‘Spyware.’
After the Tay fiasco, Microsoft hadquietly released Zo on a few platforms (not including Twitter). However, the bot was eventually taken offline in 2019 – and it wasn’t Microsoft’s last failure with chatbots.
Microsoft’s XiaoBing and Tencent’s BabyQ:
Both these bots reportedly went rogue and were pulled from the market after they voiced anti-Communist sentiments. With tight regulations on national security and sedition in China, which is governed by the Chinese Communist Party, neither company wanted its chatbot to go against national law.
Tencent later released a statement saying that its chatbot services were provided by independent third-party companies, adding that they were in the process of improving the bot.
All the evidence so far seems to indicate an incredible number of use cases for chatbots, but a correspondingly large number of risks and pitfalls inherent in the process of developing one that is functional and consistently coherent. Commercially, simple chatbots have been widely deployed, but when it comes to having truly human-sounding, warm, natural bots, there is a ways yet to go.
Photo by Volodymyr Hryshchenko on Unsplash