Colleen Callaway Eager

So, you know how when you ask ChatGPT something, and it goes and reads the whole internet, weeds out unnecessary information, and comes back with just the important stuff, in the language of your choice? Well, it can do that for the thousands of emails in your work inbox too! I kid you not.

That’s right, today I bring you a practical, universal business case for leveraging AI in customer service without risking your customer experience.

Business Use Case: High Volume Customer Emails

This is one we see most often in B2B sectors like manufacturing, logistics, and automotive service that adopted email customer communication decades ago and are now struggling to transform and adopt new technology. Customer service agents are bogged down with the highly inefficient task of reading emails, categorizing them and responding fast enough to meet customer expectations. This series of tasks is perfect for AI, which can execute all the same tasks at light speed.

Solution: Pre-Trained Large Language Models (LLMs)

Instead of building and training a custom AI model, the advent of pre-trained large language models, or LLMs, like those available in Amazon Bedrock, enable fast deployment. The advantage of pre-trained models is the ability for your team to focus on prompt design and refinement, instead of complex, time-consuming custom model development. Keeping the human in the loop to engineer prompts tailored to desired business outcomes empowers businesses with the efficiency that AI tools promise with good guardrails.

Practical Example: Where’s my order?

In this case, we prompt the AI tool to identify why customers are emailing, and then power automation tools to respond to or route the email based on the appropriate resolution. For industries like manufacturing, logistics and automotive service, the highest volume of customer communication is usually related to orders, products and invoices: When will my order arrive? What’s my tracking number? Is this product available? I need a refund on this item.

For the bot, identifying why a customer is emailing is called intent detection or classification. With prompts tailored to identify targeted intents, the bot can quickly “read” emails and tag them with the correct intent or classification so they can be routed to the correct agent to resolve the issue, or automate the best response with the information, or direct the customer to self-service solutions.

How It’s Done: Making the Model Work for You

Unfortunately, a pre-trained large language model doesn’t just do the work for you out of the box. I know, sad. The good news, though, is that all you need to “train” the LLM is to engineer prompts that align to how your business is structured. As the bot is ”reading” an email, it’s looking for identifiers that indicate intent, like ”tracking number,” and making a series of decisions based on your prompt.

Let’s look at the example of customers sending an email about a refund. For this we use a hierarchical approach, making a series of nested prompts that categorize the intent in detail. The first level of intent that the bot will detect is when it finds the word refund. The next decision you want it to make is to identify what kind of refund — a cancelled order, damaged product, invoice error.

The human-in-the-loop training exercise is to engineer a prompt that directs the bot to follow a decision tree based on hierarchies of intent or reasons that customers are emailing.

Value Add: Language Translation & Customer Self-Service Adoption

Also baked into pre-trained LLMs is the capability for multilingual translation, simultaneously with the intent detection process. With this capability, your agents can serve customers beyond their language of choice, and your customers can continue to communicate in their preferred language.

Automating responses that direct customers to self-service options creates an effective new channel to drive customer adoption of self-service capabilities. With your new email AI Bot, customers receive speedy responses to their emails that can include things like, “Here’s your tracking number. Here’s a link to set customer preferences to receive tracking notifications on your device of choice for future orders.”

De-risk AI Adoption

Perhaps the biggest value-add of pre-trained LLMs is the opportunity to implement them within the walls of your contact center, which means no risk of hallucinations or frustrating response errors. During implementation your team refines your bot by carefully optimizing a single prompt rather than iterating with revised data labels and re-training your model.

It’s important to note that “your team” means humans. This example demonstrates the human-in-the-loop necessity for AI implementations to provide oversight and deliver a tool with high confidence and accuracy. With your final prompt, emails can be “read,” classified and translated with a single LLM call, saving both usage cost and processing time.

For a more technical deep-dive on this topic, check out Multiclass Text Classification Using LLM by my expert colleague Uday Yallapragada.

About the author:

Colleen Callaway Eager is Senior Strategist, Customer Care, at digital agency Perficient.

Sign up

Sign up for a complimentary subscription to Digital Commerce 360 B2B News, published 4x/week. It covers technology and business trends in the growing B2B ecommerce industry. Contact Mark Brohan, senior vice president of B2B and Market Research, at [email protected]. Follow him on Twitter @markbrohan. Follow us on LinkedIn, X (formerly Twitter), Facebook and YouTube.