How to Keep Your Customer Support Chatbot From Looking Stupid

Posted by Alyssa Verzino on Feb 19, 2019 3:30:00 PM


We've all had interactions with clunky or incoherent chatbots that made their owners look dumb, but few companies know what steps to take to ensure their own bots don't make the same mistakes. Before you entrust a chatbot to help with customer service and support, make sure you've given it the tools it needs to avoid embarrassing itself.

Most modern chatbots run on some form of artificial intelligence and, like all AI, they need a solid dose of well annotated training data before they should be let loose on your customers. The first mistake most chatbot adopters make when training their new AI assistants is in not refining their training data to a sufficient degree.

You can't just dump all your old support emails into an algorithm and assume the AI will learn how to provide good customer service. You'll need to mark up those conversations to identify the customer, the support rep, the solutions being supported, the issues being addressed, and the outcome of the support ticket. Without that annotation, the data is just noise.

You shouldn't release your chatbot into the wild until it can pass three basic tests pretty consistently.

#1- Can your support bot recognize your products?

Your marketing department probably has a nice set of formal branding guidelines for discussing your products and services, but, much to Microsoft's chagrin, almost no one casually refers to their spreadsheet program as "Microsoft Office 365 Excel, Home & Business 2019 Edition." They'll just type "excel" and they likely won't even capitalize it. (And if you ask them what browser they're using, they're just as likely to say "Yahoo" as they are "Internet Explorer 11.")

Your support chatbot needs to understand all the common ways your customers describe your products and services so it can properly respond to any support requests. If it can't consistently do that with real, often sub-optimal customer correspondence, it isn't ready for direct customer interaction.

#2 - Can your support bot recognize your known issues?

Every product, service, and solution have a common set of known issues or limitations that generate support calls. These "greatest hits" are exactly the kind of support tickets you should offload to a support bot rather than a human support representative, but the bot has to be able to recognize these issues first. To do that, a bot has to understand that "can't download" might mean "can't install the app" in one context, and "can't export data" in another.

If your bot can't use context, or ask follow-up questions, to consistently identify known issues given the way your customers typically describe them, it needs more time with your training data. (Also, you should have a knowledge base that has well annotated version of your known issues and their known solutions.)

#3 - Can your support bot recognize your customers?

When someone interacts with your support bot, it's important that the AI can distinguish between prospects, trolls, and your actual existing customers. If someone contacts you to ask for help with your solutions, they are likely an existing customer with a record in your CRM system -- and your support bot should be able to tie any support tickets to that record. That way, the bot creates its own annotated training data for its own future use, so it gets smarter over time.

If your support bot can't consistently identify existing customers and tie support tickets to their CRM record, it can’t learn from its performance, and it can't pick up an in-progress support call if it’s interrupted. While you can release a support bot before it's ready to connect every ticket to a customer account, you'll slow down its learning curve if your rush it past this milestone.

Talla is building a customer support automation solution that can pass these tests and keep your organization's chatbots from looking dumb. If you'd like to begin your journey towards advanced support automation, contact Talla today.

New call-to-action

Topics: AI, Customer Support, chatbot, automation