Despite challenges, Salesforce says chatbot adoption is accelerating

Chatbot usage has exploded during the pandemic as organizations look to bridge emerging gaps in customer service and onboarding. In 2020, the chatbot market was valued at $17.17 billion, and it is projected to reach $102.29 billion by 2026, according to Mordor Intelligence. There’s also been a 67% increase in chatbot usage between 2018 and 2020.

This uptick correlates with chatbots’ increasing capabilities, as they enable brands to tailor offers and recommendations without humans in the loop. Chatbots leverage customer, product, and interaction data to improve experiences in real time, leading to reduced wait times, service costs, and customer churn.

To discuss trends in chatbots and conversational AI more broadly, VentureBeat talked with Greg Bennett, conversational design principal at Salesforce. Bennett believes the technology presents an opportunity for businesses to express their brands through words and languages, creating a greater degree of intimacy with customers.

Accelerated adoption

According to recent estimates, Gartner predicts that by 2022, 70% of customer interactions will involve emerging technologies such as chatbots — an increase of 15% from 2018. That’s not surprising, considering a significant portion of consumers say they prefer chatbots to other virtual agents.

“At Salesforce, we’re seeing more than a 700% increase in sessions with our Einstein bots products. I think a lot of that is due to the fact that we’ve experienced isolation as a result of a pandemic, but it also points to the need to scale up business,” Bennett said. “It may not necessarily be that businesses got the idea because of pandemic, but rather the pandemic accelerated their timeline.”

One example is Lee’s Famous Recipe Chicken Restaurant in Englewood, Ohio, which partnered with startup Hi Auto to build a conversational AI experience for its drive-thru customers. As a result of the pandemic, drive-thru orders in the U.S. saw an uptick of 22% in 2020. Consequently, drive-thru wait times increased by an average of 30 seconds, putting additional strain on employees.

Hi Auto worked with Lee’s on a solution to the challenge. At the restaurant, the company’s chatbot greets guests, answers questions, suggests menu items, and enters orders into the point-of-sale system. If a customers asks an unrelated question — or requests something that’s not on the menu — the chatbot automatically hands them off to a human. It also integrates with Lee’s employee headsets, allowing employees to provide real-time updates to inventory, as needed.

Lee’s plans to implement the chatbot at more of its drive-thrus, and Hi Auto says pilots with other restaurants are underway.

“The automated AI drive-thru has impacted my business in a simple way. We don’t have customers waiting anymore — we greet them as soon as they get to the board, and the order is taken correctly,” Lee’s owner Chuck Doran said. “We see improvements in our average check, service time, and improvements in consistency and customer service. And because the cashier is now less stressed, [they] can focus on customer service as well.”

Internal use cases

Chatbots can have value beyond customer service. For example, they can assist in the employee onboarding process, fielding screening questions, recording answers, and guiding new employees through company policies and protocols. Chatbots can also address common problems, which gives IT service desk agents the opportunity to fix more complicated issues.

Salesforce took a step toward addressing these use cases last year, according to Bennett, with the introduction of the Einstein bot intro template. Available in beta, the intro template lets developers create chatbots for onboarding, with popular Salesforce actions like creating a case or a lead, looking up an order, and adding a comment to an existing case.

“Companies can take this baseline conversation design and customize it to fit their needs. That’s really what we’re seeing — we’re seeing a shortened time between developing and deploying chatbots,” Bennett said.

The data bears this out. According to a McKinsey survey, at least a third of activities could be automated in about 60% of occupations. And in its recent Trends in Workflow Automation report, Salesforce found that 95% of IT leaders are prioritizing workflow automation technologies like chatbots, with 70% seeing the equivalent of more than four hours of savings per employee each week.

Challenges in design

Asked about trends in the chatbot industry, Bennett pointed to growing awareness of inclusive approaches to design. He’s worked with teams at Salesforce to ensure chatbots don’t discriminate against certain vernaculars, like African-American English or Chicano English.

“We ask ourselves, how can we make sure that, for example, a Black woman from the South in the U.S. doesn’t have to change their language in order to get the chatbot to react in the way that they don’t expect? We as research scientists, designers, product managers, and engineers have a responsibility to not only think about the bottom line, but also think about a total addressable market and consider the users that are being left behind.”

Natural language models are the building blocks of apps, including chatbots. But there’s growing evidence showing that these models risk reinforcing undesirable stereotypes, mostly because a portion of the training data is commonly sourced from communities with gender, race, and religious prejudices. Detoxification has been proposed as a fix for this problem, but the coauthors of newer research suggest even this technique can amplify rather than mitigate biases.

The increasing attention on language biases comes as some within the AI community are calling for greater consideration of the affect of social hierarchies like racism. In a paper published last June, Microsoft researchers advocated for a closer examination and exploration of the relationships between language, power, and prejudice in their work. The paper also concluded that the research field generally lacks clear descriptions of bias and fails to explain how, why, and to whom that bias is harmful.

“As a linguist, I look at conversation as really the fabric or the currency with which we negotiate relationships in society. Technology has now reached a point where this sort of traditionally human behavior — conversation — is something machines can partake in,” Bennett said. “The challenge now is to design a chatbot in such a way that that it adheres to human expectations about what was once an exclusively human behavior.”

Potential solutions

Bennett suggests one solution to models’ shortcomings might be developing tools for customers to evaluate quality. He points to Robustness Gym, a framework developed by Salesforce’s natural language processing group, which aims to unify the patchwork of existing robustness libraries to accelerate the development of novel natural language model testing strategies. CheckList — from Amazon, Google, and Microsoft — takes a task-agnostic approach to model benchmarking, allowing people to create tests that fill cells in a spreadsheet-like matrix with capabilities and test types, along with visualizations and other resources.

In a recent paper submitted to the Association for Computational Linguistics (ACL) 2021 conference (“Reliability Testing for Natural Language Processing Systems”), Bennett and Kathy Baxter, Salesforce’s principal architect of ethical AI, argue for reliability testing and contextualization to improve accountability. They explain that reliability testing, with an emphasis on interdisciplinary collaboration, will enable rigorous and targeted testing, aiding in the enactment and enforcement of industry standards.

Bennett also advocates including key stakeholders throughout the chatbot design process so biases can be accounted for and mitigated — at least to the extent possible. A recent attempt at this is the Masakhane project, a grassroots organization of 400 researchers from 30 African countries (and three countries outside Africa) whose mission is to strengthen natural language research in African languages. As of February 2020, the group has published more than 49 translation results for over 38 African languages on GitHub, many of which had never been translated at scale.

“Any institution has the opportunity to use a chatbot to essentially extend itself in a relationship with a customer — with prospective students, with job applicants, the list goes on. These are opportunities to create relationships and have a meaningful exchange,” Bennett said. “There’s a linguistic reason why someone uses a period versus an exclamation mark or emojis versus not. These things conveys additional meaning about the state of the relationship at hand, which is the kind of thing that will continue to be really important on the product and engineering side. With chatbots, we need to think through the conversational aspects of the conversation besides everything about the text, whether the bot or the user takes the first turn.”

Source Link