While many chatbots didn't live up to the hype, industries like fintech, healthcare, and retail are quietly adopting the technology to free up busy professionals’ time and offer guided, personalized experiences to consumers.
In 2016, chatbots were all the rage.
That year, Facebook made the Messenger bot platform the centerpiece of its F8 developer conference. Microsoft’s Satya Nadella referred to chat as the “third run-time” — an indispensable piece of operating a platform, second only to the operating system and the web browser.
Mentions of chatbots in earnings calls and press releases skyrocketed, and for many, it seemed that chatbots might be the next big disruptive technology. Thousands of companies commissioned their own chatbots in anticipation.
In the end, though, the expected paradigm shift didn’t happen.
There are many reasons why chat didn’t take off in 2016. For one, consumers found that many of the tasks the first chatbots were built to perform — like relaying the news or finding a recipe — took more time when a bot was involved.
Another problem was that bots regularly needed human assistance to understand commands. Even Facebook’s much-hyped personal assistant, M, closed down shortly after it was revealed that human handlers were responsible for some 70% of the bot’s responses.
But while many chatbots didn’t meet users’ high expectations, they haven’t entirely fallen short.
Today, the bots are still being used across industries like fintech, healthcare, sales and CRM, retail, and even law — and they’re having important, though quiet, effects. Harnessing advances in natural language processing and artificial intelligence, they’re making healthcare more navigable, shopping more personalized, and lawyers more efficient.
The important chatbots of 2019 aren’t all-knowing virtual butlers; they’re highly targeted applications of conversational technology. While they may seem less flashy, these bots are advancing their technology and making a demonstrable impact on their industries.
In this report, we round up the industries where the real “chatbot revolution” is taking place — and highlight the most important lessons for companies hoping to leverage the technology.
Table of Contents
- Personal finance: Using chatbots to help people manage their spending
- Commercial Banking: How JP Morgan’s chatbot saves the company 360,000 hours a year
- Diagnostics: Using chatbots to triage care and reduce the burden on GPs
- Patient engagement: Reducing intervention without sacrificing patient satisfaction
- Mental health: How CBT chatbots help patients reframe negative thoughts
- Sales: Automatic lead qualification makes reps more efficient
- Customer Service: Making complex websites easier to navigate
- Retail & e-commerce: Convenient shopping & scheduling
- Gifts: Increasing conversions by helping customers find gifts faster
- Personalized shopping: How an on-page concierge helps Levi’s sell more jeans
- Research: How robot lawyer ROSS reduces research time by 30%
- Small Claims: Eliminating headaches and improving citizen access to the law
- Tasks: Chatbots are best at repetitive tasks, not novel ones
- Staffing: Most chatbots work best alongside and supplementing a GUI, not fully replacing one
- Data: Chatbots are best at dealing with structured but limited corpora of information
- The road ahead
The rise and fall of chatbot hype
What is a chatbot?
When a company creates a chatbot, what exactly is it trying to do?
Simply, a chatbot is a text- or voice-based interface that lets users execute certain actions and retrieve information using language.
At the most rudimentary end of the spectrum, there are the bots that banks use to prompt callers through a phone tree, telling them to say “yes,” “no,” “check my balance,” etc. A bot like this merely lays a chat interface on top of a simple service; there is no AI involved.
At the other end of the spectrum are chatbots like Facebook’s M, which was billed as a fully capable AI virtual assistant. The bot worked through a combination of artificial intelligence and human workers.
Once the chatbot was in the hands of users, however, it became clear that there was a gap between what the service promised and what it actually offered. While M was capable of handling simple questions like “Where can I get a good burger in Chicago next week?” and performing tasks like reserving a table, it couldn’t handle more complex requests.
The goal had been to fully automate virtually all of M’s tasks; however, the platform never surpassed 30% automation, according to a post-mortem of the product in Wired. Meanwhile, the resource-intensive, human-supported model was draining Facebook’s cash, leading to the project’s eventual shuttering.
Between the rote model that can respond “yes” and “no” to a series of pre-programmed questions, and the comprehensive “virtual assistant” ready to respond to any request is a middle ground: chatbots that cannot process and respond to all generalized requests, but do use some degree of artificial intelligence or machine learning to perform their tasks.
In the best applications, those tasks are (1) more time-consuming without a bot, and (2) simple and specific.
For example, CRM platform Intercom’s “Answer Bot” can interpret questions and retrieve appropriate answers from the site’s FAQ and help documentation — something that would take users a long time to search for themselves.
Users can also mark responses as useful or not, allowing the bot to learn (in a very basic form of machine learning) what kinds of answers it should give in the future. The “Answer Bot” is able to handle approximately 29% of all customer queries itself, with no humans involved at all.
Specialized chatbots are also gaining traction in healthcare and wellness. Autonomous therapy bot Woebot, for instance, can pick up data about users’ specific needs, and over time use that information to offer a more personalized service.
At law firms and banks, chat is becoming a powerful medium for research, which can eat up valuable billable hours or necessitate hiring armies of associates. In 2017, JPMorgan Chase announced that its “COiN” bot was reviewing about 12,000 financial contracts every year — saving about 360,000 hours worth of labor from lawyers and loan officers.
Rather than try to boil the ocean, companies are picking specific functions that it makes sense for a chatbot to solve — and they’re becoming increasingly useful doing it. But chatbots in general have experienced quite a bit of tumult over the last couple of years.
The chatbot movement gains momentum
In 2016, virtually all of the world’s biggest tech companies were announcing competing chatbot projects.
For example, Microsoft had Tay, an AI bot for Twitter. CEO Satya Nadella proclaimed that “bots are the new apps” in March 2016. Amazon made Polly, its text-to-speech service, and Lex, the behavioral engine that Alexa uses to trigger actions, publicly available for chatbot developers to build on.
Messaging services Kik, Line, and Telegram all launched their own bot platforms. Facebook revealed at its F8 conference in April 2016 that it was opening up Messenger as a chatbot development platform and would work with companies like 1-800-Flowers and CNN News on early trials.
Facebook promised developers the ability to harness all aspects of the Messenger platform, including “Bot Engine,” its new natural language tool that developers could use to train their bots. By mid-2016, 11,000+ Facebook Messenger bots had launched, including trials from Burberry, British Airways, and Starbucks. More than 100,000 chatbots would go on to be built for the Messenger platform over the course of a single year.
At the same time, VC funding to chatbot startups accelerated, with firms investing approximately $85M across 30 deals from 2015 to mid-2016. Slack launched its own $80M chatbot fund, partnering with leading VCs including Spark Capital, Index Ventures, Accel, KPCB, Social Capital, and Andreessen Horowitz.
This flurry of activity generated a lot of excitement about the technology’s potential. Gartner predicted 80% of all new enterprise applications would be using chatbots by the year 2020, and that by 2021 chatbots would be the “most important platform paradigm” for the enterprise.
Chris Messina, inventor of the Twitter hashtag and product designer for Google and Uber, wrote,
“[Y]ou and I will be talking to brands and companies over Facebook Messenger, WhatsApp, Telegram, Slack, and elsewhere before year’s end, and will find it normal.”
In reality, late 2016 turned out to be the peak of chatbot hype.
Despite a flurry of investment, product launches, and media chatter, chatbots in general disappointed both users and the tech companies betting on them to succeed.
The hype dies out
Building a real “virtual assistant” capable of understanding context and responding to unclear queries proved to be much more challenging than building a simple chatbot.
Many chatbots couldn’t understand enough human language or process enough data to complete the kinds of complex requests companies promised they could handle.
Facebook scaled back work on its flagship “M” virtual assistant after it revealed that M had failed in handling 70% of all user requests. Google similarly admitted that it used humans in call centers to make its assistant work.
By 2018, news coverage of chatbots was in decline. Mentions of chatbot tech on public company earnings calls peaked in early 2018, but fell to nearly half that by the end of the year.
Source: CB Insights
The chatbot industry remains nascent, with most deals in the space occurring at early-stage (seed/angel and Series A) rounds. Given the number of angel and seed rounds that took place between 2014 and 2017, this lack of later-stage fundraising rounds suggests that few of those early chatbot investments were successes.
Source: CB Insights
In January 2018, Facebook closed down M. That same month, Inc Magazine ran an interview with Ethan Bloch, founder of chatbot company Digit, who bemoaned:
“I’m not even sure if we can say ‘chatbots are dead’, because I don’t even know if they were ever alive.”
Several other high-profile chatbot discontinuations followed. By August, many of the media companies whose chatbot experiments had been held up as proof that the trend was real — including Washington Post, Business Insider, HuffPo — had all scaled back or shut down their work on their bots.
Microsoft’s Tay shuttered in March 2016 after it started expressing troublesome political views within 24 hours of its launch, and Google’s Allo was dead by December 2018.
But the stories from the chatbot boom don’t all end in failure.
Fanta, which used a Messenger bot to drive awareness about a new product, hit a 46% higher-than-expected redemption rate on the coupons it advertised. Kia reached 33,000 potential customers and doubled it conversion rate on quote requests and test drive requests with its bot.
LEGO’s “Ralph” shopping assistant bot did similarly well, producing a 3.4x return on its ad spend.
These success stories suggest that, despite the sensationalism, there are legitimate applications for chatbot technology — especially for bots with a more specific focus.
Fintech: Contract review & bill negotiation
From crawling financial documents to simplifying how users manage their money, chatbots are creating more efficient financial technology on both the back- and front-end.
Managing money is a complex undertaking. For banks, that complexity is in the form of things like long, dense loan agreements and derivatives; for consumers, it’s in budgets, bank accounts, credit card statements, and subscriptions. Chatbots can help users keep track of their debts and liabilities.
On the personal finance side, the most successful implementations of chatbots have focused on fairly routine information retrieval and management tasks that seem fairly routine — like checking a balance or canceling a subscription — that can be made more efficient through chat.
On the corporate side, chatbots are helping to automate highly time-consuming and expensive processes, saving hundreds of thousands of hours and associated costs.
Personal finance: Chatbots to help manage spending
Crafting a simple UX for a personal finance app is a tall order. There are multiple inputs to juggle — bank accounts, credit cards, bills — and complex data that needs to be communicated in a user-friendly way.
Startups and financial institutions have looked to chatbots to help solve the problem, betting that they can create a more intuitive experience for users.
Manually monitoring bank accounts and credit card statements to identify unnecessary spending can be time-consuming, and it’s easy for small recurring expenses to go undetected — especially if transactions go through a service like iTunes or Paypal that can conceal how funds are being spent.
Chatbots, however, can continually monitor users’ bank statements for recurring monthly expenses and streamline the management process, using simple yes-or-no questions to determine if users want to pay for something each month.
Fintech startup Trim launched its SMS-based chatbot product in 2015. The bot analyzes users’ bank statements, then asks plain language questions about whether they want to cancel any recurring subscriptions, such as Netflix and Dropbox. Trim can then cancel memberships on a user’s behalf.
Bank of America’s chatbot, Erica, can help users with tasks like paying down debt and checking account status. Swedbank, Capital One, SEB, and Wells Fargo have all introduced bots for similar purposes. Wells Fargo’s iteration can understand and answer questions like ”Where is the nearest bank ATM?“ and ”How much money do I have in my account?“
But not all companies are embracing chatbots for personal finance.
Automated savings app Digit was one of the earliest companies to introduce a chatbot to personal finance. In 2018, however, the company announced that it was redesigning its interface away from the chatbot model, with CEO Ethan Bloch citing the format’s inefficiency when helping users find important financial information.
In an article for American Banker, Bloch discussed removing users’ bank account balance from the top of the Digit app to encourage users to engage with the chatbot to retrieve that information. The change only added friction for users. Customers quickly pointed out that the change was nonsensical, and Bloch said he took “full credit for the blunder.”
Companies in the personal finance space looking to leverage chatbots in their products should pay close attention to the types of tasks in question. For routine tasks, chatbots can add simplicity and clarity; with any more complexity than that, however, their value can quickly disappear.
Commercial banking: JPMorgan chatbot saves the company 360,000 hours
Promising chatbot applications in finance aren’t limited to the consumer-facing side: financial institutions are also exploring how the technology could be used to improve operational efficiency, especially by automating repetitive, routine, and time-intensive tasks.
Interpreting commercial loan agreements is one task that costs banks hundreds of thousands of hours in attorney and loan officer hours. This kind of labor is expensive, and despite requiring high levels of expertise, not invulnerable to errors.
Chatbots allow banks to process loan agreements in a fraction of the time that it would take a human — without making mistakes.
In 2017, JP Morgan launched COiN (short for Contract Intelligence), a bot that uses natural language processing to parse human requests and process commercial loan agreements, reviewing in just a few seconds documents that previously would take a team of human workers hours to interpret.
According to JPMorgan’s 2016 annual report, the bot has been trained to “extract 150 relevant attributes” from the commercial credit agreements it analyzes, helping mitigate the 80% of loan servicing errors brought about due to “contract interpretation errors.”
The banking giant reported that the software was reviewing about 12,000 contracts per year, saving about 360,000 hours worth of labor.
“People always talk about this stuff as displacement. I talk about it as freeing people to work on higher-value things, which is why it’s such a terrific opportunity for the firm,” — Dana Deasy, CIO at JPMorgan.
JPMorgan now says it plans to use similar software to handle even more complex financial filings in the future, including credit-default swaps and custody agreements, as well as to perform more natural language-based tasks, like interpreting regulations and analyzing corporate communications.
Companies interested in leveraging chatbots would do well to look beyond consumer-facing applications. Any task that is language-based, highly structured, and time-consuming could benefit from chatbot systematization.
Healthcare: Patient engagement & telemedicine
Healthcare providers today are overburdened and under-resourced, with rural areas especially struggling to provide consistent access to care. At the same administration costs account for 25% of all US hospital spending — and 86% of all healthcare mistakes.
Chatbots could help mitigate some of these problems by streamlining patient engagement and increasing access to care in underserved regions.
For example, chatbots can use natural language processing to process a patient’s reported symptoms and return a diagnosis without the patient needing to set foot in a doctor’s office. When a patient does come into the doctor’s office, a chatbot can collect their basic information at check-in, eliminating expensive human error.
They can also follow up with patients after discharge, answering questions, providing aftercare, and even checking whether patients are following plans set out by their doctors.
Chatbots leveraging cognitive behavioral therapy are even proving useful in providing first-line treatment to people who might not have the time or money to visit a medical professional in person.
All through the medical process, chatbots are helping collect information, educate patients, and even emulate the core diagnostic and conversational functions of a regular doctor visit — already bringing better care and accessibility to millions of patients worldwide.
Diagnostics: Chatbots assist with triage and reduce the burden on GPs
One of the biggest bottlenecks in the healthcare industry is physician time. There are only so many doctors, and those doctors can only see so many patients in a day.
However, not all conditions need to be physically assessed by a doctor to be successfully treated. Some healthcare providers are using chatbots as a first line of defense to sort out critical cases from less urgent ones.
Diagnostic chatbots essentially function as a “smart WebMD.” Chatbots prompt patients to report their symptoms, and then the bots return potential diagnoses and recommend next steps.
Microsoft has introduced a Healthcare Bot Project with the goal of “bringing conversational AI to healthcare.” The system uses language recognition and information retrieval rather than dynamically learning about the patient over time.
Premera Blue Cross and Aurora Health Care are using Microsoft’s platform to cut down on calls to medical centers and triage patient symptoms before they have to come into a doctor’s office.
Ada Health, a London- and Berlin-based health startup, raised $47M in 2017 for its natural language processing-enabled diagnostic chatbot project, Ada.
Ada users enter their basic information and symptoms, then answer a few follow-up questions. The bot then provides a few likely diagnoses, and, if necessary, instructions to see a doctor in-person.
The core process that Ada follows to return a diagnosis is not all that different from visiting a site like Web MD or searching symptoms on Google. But Ada adds value for users by automatically filtering results by age, gender, medical history, preexisting conditions, and more. Additionally, the bot can use information from previous conversations with patients to inform its diagnoses.
Users can also opt to smoothly transition from their conversation with Ada into a conversation with a doctor for a flat fee of $25. Any medicine prescribed can be delivered to the user’s home.
Ada is active in over 130 countries now, and the Bill and Melinda Gates Foundation has partnered with it to examine how the chatbot could be used to offer care in more remote parts of the world.
Babylon Health, the UK-based health tech company behind the National Health System (NHS) mobile app, released its diagnostic chatbot in 2018 and received significant media attention. At the time, the company reported that its chatbot AI had scored higher on a clinical exam than the average British GP: the Babylon chatbot correctly diagnosed 81% of patients, compared to an average score of 72% for human doctors.
In addition to its NHS partnership, Babylon has been working with Rwanda’s healthcare system since 2016, and counts more than 2M registered users in the country.
As of August 2019, Babylon representatives say the service “now covers 4.3M people worldwide, with more than 1.2M digital consultations completed to date, with more than 160,000 five-star ratings for appointments,” according to TechCrunch.
But Babylon chatbot’s commercialization has not been entirely smooth. The app has faced criticism for potential inaccuracy of its answers. Additionally, despite its stated purpose of helping people avoid going into the ER, Babylon’s chatbot sent 30% of all users to the emergency room — about 10% more than Britain’s national health advice line.
Companies looking to optimize triaging care and lessening physician workload should be wary of depending too heavily on chatbots to navigate potentially sensitive diagnostic situations. Chatbots are a better suited for handling routine situations where next steps are clear-cut. For serious or more subjective diagnostic situations, referral to a human physician should always be the next step.
Patient engagement: Reducing human intervention without sacrificing patient satisfaction
Data suggests that patient adherence — whether patients successfully follow the treatment plan set out by their doctor — is an ongoing problem in healthcare. As many as 60% of patients with chronic illnesses do not accurately follow their treatment plan, which leads to avoidable repeat hospital admissions, additional treatment costs, and even preventable death. Studies have placed the cost of patient non-adherence as high as $300B.
Traditionally, monitoring patient adherence is a labor-intensive process. As a result, healthcare providers are looking to chatbots to reduce the amount of human labor without sacrificing outcomes in the process.
Healthcare chatbot Florence, named after famous nurse Florence Nightingale, operates on Facebook Messenger, Skype, and Kik. The bot can remind patients to take their medication and track basic health information such as body weight, mood, or menstrual cycle to monitor patients’ progress toward their goals. Florence can also direct users to the nearest doctor’s office or pharmacy.
Nonprofit healthcare provider Northwell Health has also begun implementing chatbots as a means to improve patient engagement and adherence.
Available in English and Spanish, Northwell’s bot prompts users with conversations about general discharge issues and specific conditions. For example, someone discharged from the hospital with a heart issue might receive follow-up questions from the bot about their weight.
Source: Consumer eHealth Engagement
While skeptics have voiced concern that reduced human interaction could negatively impact quality of care, reception by patients to date has been positive: 96% of all patients who used a post-discharge care management chatbot after leaving a Northwell Health treatment facility found the conversation useful.
For healthcare providers overwhelmed by logistics and administrative tasks, chatbots can be a valuable tool for freeing up harried professionals’ time — and assuring patients that their needs will be addressed, even when a human specialist is not immediately available.
Mental health: How CBT chatbots help patients reframe negative thoughts
Approximately 1 in 5 Americans (nearly 47M people) lives with some form of mental illness, according to the National Institute of Mental Health.
For chatbot creators, the mental health space has proven to be just as amenable as physical healthcare when it comes to expanding access, reducing labor costs, and improving outcomes.
Part of what makes mental healthcare appropriate for chatbot experiments is the fact that therapy is an inherently verbal process — especially cognitive behavioral therapy (CBT), the most popular and extensively researched form of therapy.
CBT’s goal is to teach patients to recognize negative thought patterns (aka “cognitive distortions”) and then reframe their thoughts in a less harmful, more productive manner. Chatbots can be useful here because they can use natural language processing to recognize certain kinds of distortions and prompt users to rethink them.
One example is Woebot, which launched in 2015. Woebot asks users about how they are feeling, then analyzes their responses to recognize examples of cognitive distortions. The bot then coaches the user on how to shift their perspective following a pre-designed decision tree that determines the correct response to user input.
Source: Business Insider
Woebot was founded by a team of Stanford researchers. Prior to launching the service, founder Alison Darcy published the results of a study featuring an early iteration of the tool in the Journal of Medical Internet Research, Mental Health. The study, which followed 70 college students with reported symptoms of depression for 2 weeks, found that the participants assigned to chat with Woebot saw a significant reduction in depressive symptoms compared with participants who read an e-book on CBT instead. According to Dr. Darcy,
“[Participants] said Woebot felt much more like a friend rather than an app or a technology.”
Another mental health chatbot, Wysa, also incorporates principles of CBT, but with an added focus on mindfulness and meditation. The app provides a “toolkit” of exercises designed for purposes like focusing while overwhelmed, managing conflict, and relaxation.
Wysa “learns” from its interactions with a user to recommend certain tools; however, Wysa’s website stress that the bot is “restricted in the means of response” and that its “intended usage is as an early intervention tool,” not as a replacement for in-person therapy.
Other clinical research suggests chatbot technology may have an important role to play in fighting the growing mental health crisis in the United States.
Chatbots can support users on their own schedule, not their provider’s. That’s a significant value-add in an area where so much of the work happens during the patient’s everyday life, outside of the therapist’s office.
But mental health chatbots are most effective as an accountability tool, reminding users to apply the mental health strategies they have learned. The primary analysis, diagnosis, and education is still best undertaken by a trained mental health professional.
Sales & CRM: Qualifying leads and answering questions
In the world of SaaS, chat has quickly become one of the primary ways to interface with a company’s website. In mid 2019, a third of respondents to a Drift and Audience survey reported using a site’s online chat feature in order to communicate with a company in the last 12 months.
As customers grow more comfortable with chatbots, enterprises are investing in chat as a sales and customer management tool to help triage customer problems, move users in and out of the sales funnel, and offer a more personalized experience to select customers.
Going beyond rudimentary chatbot functions, bots at the cutting edge of sales and CRM are starting to perform complex tasks — bringing revenue and productivity gains in the process.
Sales: Automatic lead qualification makes reps more efficient
Businesses that rely on inbound marketing for new customers are likely to have some form of lead qualification effort — it can be costly for sales reps to spend time on calls with people who aren’t a good fit for the product they’re selling.
Complicating this effort, some SaaS companies find that people end up in their sales funnel who are actually current customers just looking for an answer to a question.
Chatbots in the sales industry today are designed to help solve these challenges by separating real prospects from customers looking for help, as well as helping sales reps deal with potential customers more efficiently.
For example, Intercom’s Operator chatbot asks site visitors if they’re an existing customer to help prompt them to different teams in the company.
Implementing this question for new visitors to the Intercom site led to “significant time savings” for Intercom’s SDRs, according to the company.
“That might seem like a simple little step, but it dramatically changed our sales team inboxes — more than two-thirds of anonymous leads answered the [initial] question, and about half were existing customers. Needless to say, they weren’t looking to talk to the sales team, meaning a dramatic drop in the number of conversations we were having to engage in.” —Sebastian McKenna Long, Sales Development Representative at Intercom
In short, chatbots are being used to automate relatively simple but time-consuming tasks like capturing leads’ details and asking qualifying questions. This can help free up human workers’ time for more complex tasks, such as understanding customer pain points and building relationships.
Customer service: Making complex websites easier to navigate
When asked about the most frustrating parts of an online experience, one of the most popular responses is “can’t get answers to simple questions” (cited by 34% of respondents to an Audience and Drift survey). This was followed by “sites are hard to navigate” (30%) and “basic details about a business are hard to find” (25%).
All 3 of these top frustrations reflect that websites today are bigger and more organizationally complex than ever.
The median size of a single page on a desktop website is about 1,500 KB, a 3x increase from 2011, according to HTTP Archive.
Source: HTTP Archive
Ideally, websites should help customers answer questions themselves. In reality, 90% of the calls placed to any given company’s contact center reportedly come after a visit to that company’s website, according to TechCrunch.
Chatbots could help address this as they’re a well-suited medium for offering quick customer service.
Customer service ranks high on the list of functions people report they would like to use a chatbot for. Around a third of respondents to an Audience, MyClever, and Drift survey said they would use a chatbot for “a quick answer in an emergency,” “resolving a complaint or problem,” and “getting detailed answers or explanations.”
There’s evidence that suggests customer service chatbots can be a useful tool for these kinds of user problems. Globe Telecom, a Philipennes-based telecommunication company, says that building a customer service chatbot reduced the overall amount of calls being made into its call center by half, cut call center costs by 10%, and helped enable a 3.5x increase in employee productivity.
In addition to lowering costs and raising productivity, chatbots can also improve customer satisfaction. Globe Telecom’s customer service chatbot resulted in a customer service satisfaction rate that was 22% higher than that of its call center, while Dutch airline KLM found that its service chatbot got a higher Net Promoter Score — an index for gauging customer loyalty — than typical channels.
Companies using chatbots vary in their approaches to supplementing them with humans. KLM reports that its chatbot software is trained to answer 60,000+ different questions, though a human agent still checks each response.
Intercom’s “Answer Bot,” on the other hand, uses AI-based tech to draw from the company’s existing FAQ database to answer about 30% of all customer queries itself, with no humans involved.
Companies interested in chatbots as a sales and CRM tool will need to evaluate how their available data — such as help center documentation, FAQs, customer feedback on responses, and more — can support a bot as a customer service medium. They will also have to consider to what degree chatbot responses should be monitored by humans.
Commerce: Convenient browsing & personalized shopping
One issue for many e-commerce retailers is that user experience challenges can make online shopping a frustrating experience. Difficult to navigate layouts, intrusive autoplay videos, distracting ad carousels, and a lack of accurate search capabilities can be a turn-off for would-be shoppers.
Chatbots may be able to help. Acting as a virtual store concierge, they are being used to help answer customer questions and direct them towards relevant products. Shopping via chatbots is already becoming a more common occurrence, especially for younger consumers.
Source: Worldpay via PaymentsSource
Gifts: Increasing conversions by helping customers find gifts faster
Holiday gift-giving is a stressful activity for 84% of shoppers, even as more people skip the line to shop online, according to live chat service Needle.
Compounding the difficulty of choosing the right gift is the challenge of navigating the goods available on many e-commerce sites. (Recall that hard-to-navigate sites are one of the top frustrations reported by consumers.)
Retailers often have a huge product catalog to present, and making that information easily searchable is a difficult task. For example, click the “Theme” option on the LEGO site, and you’re met with 40 different styles of LEGO products to choose from.
Aware of this difficult-to-navigate layout and shopper pain points around gift giving, LEGO turned to a Messenger chatbot.
The goal of the chatbot was to shepherd shoppers through LEGO’s extensive catalog of products (the company says it has produced 3,700 different varieties of LEGO brick) and recommend gifts that fit the recipient’s age and personality as well as the buyer’s budget.
Rolled out during the 2017 Christmas season, LEGO’s “Ralph” bot is now available year-round. After asking a series of predetermined questions about personality, age, and budget, Ralph presents users with a link to either buy a recommended product.
During the holiday season, LEGO says it reaped a 3.4x higher return on investment by advertising the bot on Facebook than it did from ads that merely linked back to its website. What’s more, the value of the products sold through the Messenger bot was 1.9x higher than purchases made through the website directly.
The Messenger bot didn’t just help LEGO sell more products — personalizing the shopping process helped them sell more expensive products and drive more revenue.
In an era where curation and customization are defining features of the shopping experience, and where retailers are carrying more products than ever before, chatbots can be a powerful tool for adding a more personalized touch to the consumer experience — and for converting that customer-centricity into sales.
Customer journey: Using Messenger to drive people back to brick-and-mortar
The “retail apocalypse” of the last several years has continued into 2019, with Payless, FullBeauty Brands, Charlotte Russe, Gymboree, and others shutting down operations. The problems that store owners around the US are facing — low foot traffic, debt, inefficient operations — are becoming existential threats for more and more retailers.
Nevertheless, around 90% of retail sales still happen offline, according to the US Census Bureau, making brick-and-mortar a key priority for many retailers.
One company that has actually expanded its brick-and-mortar footprint over the last few years is Sephora. A reason for that success is that the company uses online offerings to supplement its in-store experience — including using chatbots to help customers find new products and lure them back into the store.
Sephora was one of the first retailers to start using a chatbot as part of its e-commerce strategy. Sephora Assistant, which launched in November 2016, was built to get more customers into stores by making it easier for them to book a makeover. Through the conversational interface, customers can set up an appointment with their local Sephora store in just three steps (five steps less than booking through Sephora’s app took).
After launching Sephora Assistant, the company says that its conversion rate on in-store bookings through the chatbot reached 11% — higher than the other mediums used for booking appointments. Additionally, Sephora saw in-store spend increase, with customers who booked appointments through Sephora Assistant spending over $50 in-store on average.
“We were excited about combining ease and utility for Sephora reservations — enabling our clients to book a makeover with us in seconds, just by messaging Sephora.” — Mary Beth Laughton, EVP Omni Retail, Sephora
The Sephora Assistant chatbot has since evolved to provide users with personalized product recommendations for their skin tone to look up product reviews.
Ultimately, the Sephora Assistant chatbot aims to be a bridge between product marketing and in-store shopping, helping customers identify the products they want and then encouraging them to make in-store purchases as well. This progression — from in-app curation to in-store purchase — underscores the opportunity for retailers to explore additional use-cases for chatbots.
Personalized shopping: How an on-page concierge helps Levi’s sell more jeans
Buying a good pair of jeans can be tough. Sizes and fit aren’t consistent between brands, and shoppers have a huge variety of jean types to choose from.
For many consumers, the jean shopping experience could be improved by increased personalization — a function chatbots are well-suited to.
Levi’s was well-positioned to introduce a jeans-buying chatbot, given its coded system for identifying specific details about how a pair fits.
While organizing jeans into numbered codes has helped Levi’s build a more extensive catalog that can cater to a range of body types and preferences, it has also made it more difficult for online shoppers to choose what to buy.
Levi’s chatbot helps customers sort through its catalog using simple language.
“We have over 20,000 denim [stock-keeping units] on our site, so we need to give our consumer some guidance so they don’t get lost in that sea of denim.” — Brady Stewart, SVP, LSA Digital, Levi’s
To help customers find a suitable pair of jeans, Levi’s chatbot asks simple questions — like “How do you want your jeans to fit through the hip and the thigh?” — then uses shoppers’ answers to recommend jeans that match their preferences.
Originally developed by Mode.ai for Messenger, Levi’s removed the chatbot from Facebook in 2017 and placed it on its own site instead.
So far, the results of Levi’s on-page chatbot experiment have been positive for the company — customers who use the chatbot are up to 80% more likely to make a purchase, according to Brady Stewart, an SVP at the company.
Levi’s bot is another example of the power of chatbots as a personalization and curation tool. Customers are more likely to buy when they can find exactly what they’re looking for, and a chatbot can guide them there directly.
Law: Automating small claims and powering research
Like finance and healthcare, the legal system is complex and difficult for many people to understand — but it is also guided by a set of clear, well-documented rules.
Whether you’re a lawyer researching legal precedents for some obscure offense, or a typical person trying to figure out how to contest a traffic ticket, navigating legal documents and procedures can be time consuming and expensive.
Enter chatbots. The highly structured, routinized nature of legal tasks and documents makes them particularly suited to the chatbot value proposition: chatbots can parse digitized documents, such as court filings and judgments, and extract relevant information in a fraction of the time it would take a human associate or paralegal.
Research: ‘Robot lawyer’ ROSS aims to reduce research time
One example of chatbot technology in the legal sphere is the AI-driven “robot lawyer” ROSS.
Powered by IBM’s Watson, ROSS uses natural language processing to answer queries from human lawyers doing research for cases.
Legal research is a highly time-consuming and complex task, requiring lawyers to slog through cases, statutes, and regulations from the federal to the local level, as well as secondary sources like law reviews. Legal researchers must look at not just the text of a statute, but also the documents that preceded it to understand the legislative intent behind it.
ROSS tries to speed this process up by combing through the text of thousands of past legal cases, looking for relevant precedents, finding matching language, and distilling the last court’s decision-making process.
With the right input, the chatbot can tackle relatively complex questions — like “what is the standard for ‘false by necessary implication’ advertising in New York since 2010?” — and retrieve relevant information for the user.
This search process can be done in a fraction of the time it would have taken a human, saving time and money.
“I work on a lot of complex cases where the research I need to perform falls into a needle in the haystack search,” says Alejandro Miyar of Berger Singerman, “and while I have spent a lot of time in my practice using more traditional legal research tools, I was interested to see how natural language processing could be used for legal research, so that I could spend less time finding the answer and more time putting it to use for my clients.”
Similarly to JPMorgan’s COiN system in finance, ROSS demonstrates how the utility of chatbots extends beyond customer-facing applications. Reducing the time spent on research tasks and busywork is also powerful, and an area that may be attractive to law offices looking to cut back on costs.
Small claims: Reducing headaches & improving access
Lawyers are not the only ones who could potentially gain from chatbot adoption in the legal sphere: consumers could also benefit, as the technology could help streamline more routine legal interactions, reducing the time and cost for consumers as well as law professionals.
For example, while contesting a parking ticket may be a relatively routine legal procedure, it is mired in enough bureaucracy to make the process intimidating for many people. The process can require navigating communication with the relevant court, assembling paperwork, and pleading a case to the judge, among other steps.
The free parking ticket-contesting app DoNotPay aims to make this process much simpler. The chatbot guides the user through a series of pre-programmed human language questions — e.g. “were you parked illegally because of a medical emergency?” — and uses their answers to determine whether the user is eligible to contest their ticket.
The app can offer suggestions and help collect evidence, such as a lack of clarity on the street parking signs or an illegible ticket. It then automatically goes through the process of submitting a ticket contestation.
By mid 2017, DoNotPay said it has saved 375,000 people around $9M on parking tickets.
DoNotPay’s ambitions go beyond parking tickets. The app’s website states that it is extending its service to include small claims, customer service issues, and government paperwork, among other areas.
Some proponents believe that down the road, legal chatbot technology could further ease the load on the court system, which is frequently overburdened and under-resourced.
According to Roland Vogl, Executive Director of the Stanford Program in Law, Science, and Technology, “We’ll see cases that get navigated through an artificially intelligent computer system… lawyers will only get involved when it’s really necessary.”
Chatbots’ ability to help automate otherwise tedious and time-intensive tasks is one of the technology’s key value propositions. Law has more of those types of tasks than most disciplines, providing opportunities to leverage chatbots to increase efficiency. But firms should be cautious about expecting chatbots to do too much too quickly — in many cases, a human perspective and education are still essential.
Key takeaways from the ‘chatbot revolution’
Much of the early hype around chatbots didn’t pan out. But despite the drop-off in excitement, chatbots today are making an impact across many industries, including some unexpected ones.
Instead of booking our dinner reservations, chatbots are helping lawyers research precedent. Instead of planning our daily schedules, they’re helping take care of patients after they leave the hospital.
While chatbots’ failure to live up to the hype may have been disappointing for early adopters, it’s now clearer than ever where chatbot technology can be useful.
Tasks: Chatbots are best at repetitive tasks, not novel ones
The most hyped up early chatbot experiments advertised full-featured AI butlers, like Facebook’s M. While these tended to disappoint users, chatbots that focused on automating relatively simple tasks (often undertaken by sales and customer support teams) have performed better.
Chatbots forced to deal with a constantly changing pattern of requests, like Facebook’s M, have been forced to rely on human handlers to fill in on a large percentage of queries.
But chatbots that accept only specific types of requests have demonstrated an ability to help even with complex tasks in many cases.
At JP Morgan, a bot has taken over the time-consuming and expensive process of document review for commercial loan agreements — a function that used to require a team of highly paid lawyers to finish.
Within a certain range of use cases, chatbots can also make use of artificial intelligence and machine learning.
Intercom’s Answer Bot, for example, learns through iteration how to better surface FAQ information for users seeking help. When users indicate that an answer was useful, it flags both the query and the response, so it can learn to more reliably deliver useful answers to people with similar problems.
Finally, chatbots especially excel in repetitive tasks where humans often don’t. Repetition creates boredom and distractions, laying the foundation for mistakes. Some studies have indicated that even veteran manufacturing workers can make errors when performing repetitive, manual tasks. Chatbots, as long as they’re programmed correctly, can’t get bored, a trait that could help mitigate issues stemming from human error.
Complementary approaches: Most chatbots work best alongside GUIs
While text is the primary method for communicating with chatbots, many of the most successful chatbots utilize a graphical user interface (GUI) to give users more freedom and power in how they interface with the bot.
Babylon Health’s GP at Hand service, for example, doesn’t only record patient symptoms via text; it also aims to connect patients for follow-ups with doctors over video. The data collected by the bot can then be shown to doctors through the app’s visual interface, allowing them to follow along with any diagnoses the chatbot has made while they talk to the patient one-on-one.
Source: Babylon Health
Retail chatbots like Levi’s typically incorporate a GUI into the chat window itself. While communication is useful for clarifying how a customer wants their pants to fit, and what their preferred style is, actually choosing a pair of jeans to purchase is not something shoppers can do purely through text.
Using shoppers’ answers to its questions, Indigo provides users with inline product photos linked to product pages for recommended items.
Data: Chatbots are best at dealing with structured but limited information
Early chatbots often failed because they couldn’t understand user requests. There was an almost endless variety of questions that M could be asked, creating a virtually impossible task for the bot’s natural language processing abilities.
Today, the chatbots that are proving most practically useful are in fields like law, health, and banking — fields that all run on highly specialized, yet constrained, vocabularies and documentation.
Non-experts may have trouble navigating legal or medical language, but chatbots can be programmed to do so because this type of language is so formally structured. For chatbots, technical language is often easier to parse than open-ended human speech: even arcane terms have settled definitions.
Chatbots are similarly well-positioned in retail. While the Nike catalog may not be highly complex, it is extensive. Sorting through hundreds of items is a challenge that GUIs often try to solve using nested, dropdown menus — not always the most efficient way to help a human find what they’re looking for. A chatbot, on the other hand, can reduce the burden of finding an item by making the catalog accessible via language.
The road ahead
When Facebook launched its chatbot project M in 2016, its fatal flaw was expecting too much, too soon. An all-purpose chatbot that can handle everything from our schedules to our sanity sounds appealing, but a tech application that does “everything” can end up doing nothing particularly well.
Ultimately, chatbots benefit from the same kind of bodies of knowledge that regular software applications draw from. The difference is that chatbots can make those applications simpler to use and navigate because they can harness the immediacy of human communication.
The reprieve in chatbots came from trading aspiration for application — moving from what sounded great on paper to instead focus on what delivers clear, measurable success for the user. Nearly two years after publications were trumpeting the death of chatbots, businesses in retail and sales to finance and medicine are finding that there’s something more to speak about.