Photo Credit: Pixabay
Customer Service is a holy grail in retail. It is an integral part of customer experience, where customers interact with a business during their pre-purchase and post-purchase journey for a myriad of reasons including queries about products, checking their order status, modifying order details, providing feedback, and more.
At Walmart, millions of users reach out to our Customer Care through different channels like chat, call, email, etc. for varied requirements ranging from order-placement help to order-status checks to returns and more. How we service the user at this juncture in their shopping journey is a crucial factor in determining customer satisfaction. Interestingly, a significant percentage of users reach out for relatively simple queries like checking their order status, which may be easily automated. Also, a considerable segment of users interacting with customer care are existing customers with prior order history. These imply a large scope for automation to personalize and build seamless conversational experiences for our users.
A user may browse help center or reach out via channels like chat, email, call etc. to get their queries resolved. In some cases, the user may also connect with an agent to get the resolution.Converse, Walmart’s in-house conversational AI platform, is uniquely suited to help our customers, associates and sellers resolve their queries by using the platform’s NLU (natural language understanding) and personalization capabilities. In our earlier blogs, we have discussed how Converse enables conversational experiences for voice shopping, text-to-shop, AskSam assistant for stores associates, etc.
In this blog, we dive into how Converse powers conversational assistants for customer, seller, and associate care. The goal for these conversational assistants is to anticipate user needs and understand any feedback, to provide intuitive and personalized experiences. Understanding user intent enables us to route users into appropriate self-service flows e.g., starting a return or checking order status. These experiences are enabled across platforms such as Chat, IVR (Interactive Voice Response) systems, Help center pages and more. This automation helps not only in lowering wait time for customers but also leads to a better customer experience and increased sales. It also enables our associates, agents, and partners to quickly serve and delight customers anytime, anywhere.
Let us now dive into Walmart’s in-house AI solution that powers the conversational assistants for ‘Care’.
AI Solution
In its simplest form, our AI solution is two-fold:
1 — Natural Language Understanding Models
Converse’s natural language understanding models aim to understand a user’s ‘utterance’ as they type/speak. This facilitates building intuitive experiences for users, where they will be presented with and routed to relevant flows.
Chatbot without/with NLP Models | Screenshots from two versions of Walmart Help Center’s chatbot. In the chatbot without NLP (on the left), whenever a user types, they are immediately escalated to agents. Whereas the chatbot with NLP in the backend (on the right) can comprehend the user’s utterance and suggests appropriate flows to the user for an intuitive experience.We have an NLP (Natural Language Processing) model suite which includes the following models:
- Intent detection: predicting intent using utterance and any context from current/previous interactions (e.g., start a return, check order status etc.)
- Entity detection: detecting entities within an utterance (e.g., item name, order-id etc.)
- Sentiment analysis: detecting user sentiment during a conversation (e.g., happy, neutral, etc.)
- Q&A over knowledge base: matching user queries to relevant knowledge-based articles to extract and provide answers (e.g., people interested in ‘Walmart+’ could be routed to a relevant article)
The NLP-suite helps us understand a user’s query to route them through relevant automation flows, which can resolve the user’s issue. This significantly reduces the query resolution time for users and increases the automation rate* for the channel, creating a seamless experience for the users.
(*Automation rate is the percentage of contacts that get resolved via automation flows without the conversation escalating to an agent.)
2 — Proactive Assistance Models
Even before a customer explicitly speaks or types anything — Proactive assistance models predict the customer’s need/intent based on the customer’s transaction and interaction data. This enables us to personalize the customer’s experience, by proactively offering relevant flows/options that will delight them.
This is an all-encompassing approach for personalized customer care across all touchpoints — whether it is in chat, on the help center, through IVR (interactive voice response) or others. For Chat, it surfaces as personalized options shown to a user as soon as they open it. On IVR, it surfaces as personalized welcome messages. And as different other flavors across channels and platforms.
Personalization is a powerful tool for building seamless user experiences, especially because a significant percentage of users have some order history or engagement before they reach out to customer care. So, rather than offering generic flows (like order status) to all users, we now show the most relevant and personalized options e.g., a customer whose order is in transit receives a different message from a customer who has returned items in their order. This ‘proactive’ messaging requires us to leverage the user’s interaction and transaction history.
We have observed a 17–33% increase in automation rates by launching ‘proactive care’ solutions across channels. In addition, it also allows more users to ‘discover’ new and relevant flows. We will cover the interesting details and insights in an upcoming blog. Meanwhile, you may check out our blog on how proactive assistance works within an ongoing conversation for shopping experiences.
Chatbot without/with Proactive Models | Screenshots from Walmart.com Help Center: Baseline/Fallback flow options (on the left) are the same for all users v/s Proactive model flow options (on the right) are ‘personalized’ for a user based on their latest transaction and interaction data. A similar experience is serving customers on Walmart’s IVR, where they get ‘personalized welcome messages’ using the Proactive models.A user may be greeted by ‘personalized’ intents/flows predicted by ‘Proactive models’ and as they interact further, the ‘NLP models’ kick in and take the conversation forward.
Appropriate problem definition and dataset creation are important challenges in building any AI solution. They manifest themselves in interesting forms for the care assistants as well.
- Problem definition: Which problem are we solving and at what touchpoint? Is it personalization or understanding single user utterance or understanding complete conversation? Are we creating experiences for the instant when a user opens our website or when they open our chat or when they type in the chat? In the example screenshots above, the Proactive model is invoked when a user opens a chat, whereas the query understanding model is invoked when a user types into the chat.
- Dataset creation: A multitude of data challenges exist given the sheer scale of e-Commerce data with multi-faceted customers, omni-channel offerings, and multiple humans in the loop (customers, agents, sellers etc.).
Each of these challenges is interesting in its own right and we will touch upon them in later posts. For now, let us dive more into the design, scale, and impact of Converse for ‘Care’.
Design, Scale & Impact
‘Converse’ powers care across US and international markets. The AI solution is already deployed for our customers across markets and channels including US’s help center Chat and IVR, Canada’s help center Chat, ASDA help center Chat, Chile’s WhatsApp, and Mexico’s WhatsApp. We have also rolled out the solution for our marketplace sellers in the US.
The system architecture and the AI solution powering it have been built as pluggable components which are leveraged across markets (e.g., US, Canada), channels (e.g., Chat, IVR), languages (e.g., English, Spanish) and end-users (e.g., customers, agents). This allows us to leverage different flavors of the same models across a multitude of use-cases.
‘Converse for Care’ provides an intuitive and seamless conversational experience to our users, enabling millions of conversations to be resolved via automation itself i.e., users are empowered to solve their queries via self-service flows and do not need to be escalated to agents. This also results in shorter query resolution times for our users and saves Millions of dollars yearly.
What’s Next?
In this blog, we discussed one flavor of AI for Customer Care — related to building smart conversational assistants, but we are just getting started. There is so much more: smart routing of customers to appropriately skilled agents, AI-assisted training/re-training of human agents, proactive defect discovery and issue resolution for our customers, etc. We are working on these interesting dimensions while also scaling and generalizing our models further for deployment across devices, geographies, Walmart businesses, languages, and end-users.
For folks interested in the specifics of the models mentioned above, we will discuss in more detail the algorithms, challenges, and solutions powering the AI for Customer Care at Walmart. Stay tuned!
AI for Customer Care– Building smart conversational assistants was originally published in Walmart Global Tech Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.
Article Link: AI for Customer Care– Building smart conversational assistants | by Priyanka Bhatt | Walmart Global Tech Blog | Medium