If you’re a retail leader, chances are you’ve been thinking about how AI can improve your bottom line. Sixty percent of cross-industry leaders plan to spend at least 5% of their IT budgets on AI initiatives, and over 35% of retail CIOs say their organizations are implementing AI to improve areas like customer experience.
For technical and nontechnical leaders alike, understanding where AI delivers the most value to brands and retailers is a complicated question. At Caravel, we focus on a key area of the retail process, Customer Experience (CX). Within CX there is a big opportunity to engage in conversation with customers to meet their individual wants and needs, easier and at scale.
We've written this article as a way to break down the challenges of conversational AI and how the Caravel product approaches making it easy to use and highly personalized. We've split the process into existing "challenges" and AI-enabled "solutions." Within each solution, we share some Caravel examples to illustrate what's going on in the background and introduce the results you might expect.
Challenge 1: Understanding the product catalog
In order for inventory to be easily discovered by customers, it is necessary that product catalog data be clean and up to date. It's also important that the process of updating it is streamlined; otherwise, it can turn into a never-ending process that distracts internal teams from other important tasks.
There’s a wealth of knowledge hiding in unstructured product data like descriptions, text, images, and reviews that merchants put a lot of time and effort into producing. Until recently, this unstructured data was difficult to incorporate for use in your customer experience.
The first benefit AI should provide is reducing routine tasks internal teams must do, like endlessly managing and replicating product data that already exists. Computer Vision and Natural Language models can automatically extract rich information from unstructured data to inform customer-facing interfaces with minimal intervention from you. This limits the reliance on product data wrangling—reducing internal workload and allowing teams to act more nimbly as they get products out to customers. With Caravel, our model might analyze a product image for jeans in this way:
Challenge 2: Understanding what customers need
On the front-end, eCommerce has traditionally required customers to fit their interests or needs into a retailer’s predetermined categories and attributes, but what if those paths don't match a customer's understanding? The task of matching needs to products falls onto the customer.
Looking for wedding-appropriate attire? With a traditional filtered search, a customer might need to navigate through dresses>summer dresses>cocktail dresses to arrive at the appropriate products. It may take the customer several steps to figure this out, can lead to frequent dissatisfaction and drop-off, and the merchant is none the wiser about the customer's true intent.
At Caravel, we approach this challenge by applying Semantic Search, a process that reviews the context or meaning of a customer’s inquiry—in their natural language—and uses probabilistic matching to showcase relevant products that fulfill their needs. The customer's search terms can further be matched against a “complete understanding” that has been formed of products from your unstructured data (as explained in Challenge 1). As a result, customer questions and inquiries are addressed with less effort, they find corresponding products more quickly, and, by enabling your customer to naturally communicate their needs, you get greater insight into their needs and motivations.
In the next few examples, we'll dive deeper into the AI behind matching your customer's needs to your product catalog through conversation, and how this improves your overall CX.
Challenge 3: Deciphering a customer’s own words
In order to enable natural engagement that doesn't feel bot-like or frustrating, conversational AI must be good at understanding the nuances of everyday vernacular and accurately deciphering intent to determine the best course of action. Modern conversational AI leans heavily on Natural Language Understanding (NLU) to analyze what the customer is asking for, using their own words. With Caravel, we use domain-specific NLU (trained on conversations in your retail vertical) in order to be as accurate and helpful as possible. Let's look at how Caravel would determine a shopper's intent and their key search terms within a conversation.
Once we identify the key search terms, Caravel's language models are able to break down the terms into domain-specific categories and attributes like price, occasion, size, material, and/or color preferences relevant to the product they're searching for.
Challenge 4: Knowing what to ask in the moment
Let’s keep working through customer language and intent. Getting the customer from initial inquiry to the right product is often a multi-step process. While traditional searches give the customer one chance to get it right and present pages of results, conversational interactions have an ability to guide customers to the most relevant match. This curation process reduces anxiety for your customers, but it’s hard to do well.
A big step in that curation process is determining the next best question to ask. Let’s look at an example:
With traditional automated solutions, you have to pre-define all your responses on the platform's backend. However, the number of responses you have to plan for becomes overwhelming when you have more than a few products, all with many attributes.
Instead of overwhelming yourself with the manual work of designing dialogue for every possible outcome, you can make use of Machine Learning to build a map of relationships among products and their attributes, and define a heuristic (a scoring function) to determine the optimal conversational flow in real time based on the current context. Heuristics can also make use of behavioral data and other data points from customers. Now you only have to design snippets of dialogue because the flow is optimally designed for you as the conversation happens. A Caravel heuristic for the above example might go like this:
Generative modeling can go a step further and generate responses from scratch for you, completely removing the need to write any dialogue. However, this is a very active research area as even state-of-the-art results can be unpredictable and are not suitable for business applications without human moderation. Generative text is a topic we'll save for another post where we'll share some of Caravel's active research efforts.
Challenge 5: Adjusting to the customer
Customers quickly become frustrated when a conversation breaks down because they changed their mind or made a mistake in their response.
This is because traditional bot platforms rely on tree structures that operate similarly to if/else statements: if a customer gives this response, then reply with this. Customers are led down a pre-defined path—but if they change course or deviate from it, the experience fails.
In order to make conversation truly natural for customers, a chat platform must allow customers to change their mind, ask non-linear questions, and still get useful responses. To allow for this flexibility, you can't rely on decision tress or pre-built dialog.
Instead, automated conversational interactions must be customer-centric: driven by the customer, allowing them to jump around in their responses just as they would in natural, in-person conversations.
To illustrate, let’s return to the sunscreen example:
At best, a traditional bot would start the customization process over again, frustrating the customer. A customer-centeric approach would recognize the change, recall past preferences, and adjust the total current order without breaking stride.
Bonus Challenge: Overcoming a cold start
While this challenge isn’t AI-specific, we wanted to mention it as we’ve heard it come up in many conversations. Whenever personalization platforms encounter a brand new customer, they lack any relevant history or foundation to direct recommendations or personalized content. Without a baseline of customer behavior, engines can’t apply analysis to predict what a customer might need or want. Like a car in the winter, platforms are forced to “cold start” an interaction in a way that might feel jarring.
While many companies look to solve the cold start problem with behavioral predictions, approaching it with a natural language solution invites the customer to share what they’re looking for—yielding a more accurate understanding of current and future intent than any behavioral AI could predict.
What to expect
So what outcomes could you expect by adopting conversational AI? First, you'll free up time and resources for high value tasks that can make your CX exceptional. Second, addressing customer needs instantly and guiding them to the products they'll love builds their confidence and overall satisfaction.
To give you an idea of potential impact that confident shoppers can have, merchants that use Caravel have seen:
- A boost in digital sales up to 111%
- Conversion rate increase 64%
- Reduced return rate by as much as 58%
Not to mention, you'll be equipped with an incredibly accurate source of insights about your shoppers.