Robot child lying on the floor and reading a book. Bluish background
PHOTO: Shutterstock

Think training a chatbot is easy? Think again. According to Ashok Kumar, vice-president of digital for Verizon, it’s a process that never ends. Operating through existing messaging platforms like text, Facebook Messenger, or Slack, bots use natural language processing to connect people with information. Verizon’s customer-facing chatbot, for example, operates through the company’s My Verizon and My Fios apps. The bot helps consumers with billing, device troubleshooting, and customer service. My Verizon engineers did the initial development and, today, a team of 50 people maintain the bot. From the customer’s perspective, though, the experience feels flawless.

But every time the company issues a new product, the chatbot has to be retrained. “Normally when you build a product, you work in parallel to build a website,” Kumar explains, discussing how product and marketing teams often partner pre-launch so a business can advertise new offerings. Chatbot development ideally works the same way. To answer customer questions about a service, he says, product and bot developers “have to know [and] work in parallel to build scenarios.”

Understanding the Necessary Context and Communication Complexity

Take Verizon’s latest product for example, which Kumar says required months of chatbot training: a family data plan called Mix-and-Match. If a customer asks, “Can I mix and match services?,” the bot needs to know whether she’s asking about the new plan or wants to combine existing products — like maybe adding cable to lower her phone bill.

Context doesn’t just dictate the kind of training a chatbot gets, but how much: What does the bot need to do and how complex of a communication channel will it interact in? If a chatbot connects through a multiplayer video game, Andrew Rollins, CEO of tech company Everbloom, says, initial setup can take “something like 180 years worth of compute time.” And that’s just to teach the bot how to play the game — that doesn’t include training for actual conversation. “That’s not something that just a random start-up off the street is gonna be able to pull off,” he explains. “You can certainly put a bot out there, see how people interact with it. But there are some things that go on behind the scenes… that require a large capital investment.”

Related Article: Top 14 Chatbot Building Platforms of 2017

Be Prepared for Constant Iteration and Refinement

Of course, open-sourced code could help get chatbots started, but at Verizon, the challenge isn’t creating these bots; it’s maintaining them as new products and services are launched.

In addition to training for new products, Kumar’s team constantly keeps an eye on the way language changes over time. A team of computational linguists monitors conversations for what Verizon calls “fall-out”: words and expressions the company chatbot doesn’t yet understand. He says, “You’ve got to have teams who are looking at these unrecognized phrases.”

A chatbot also “needs to be aware of what’s going on currently in and around it because people are going to ask questions — things like...just last year when we had this major hurricane,” Kumar continues. During Harvey, Irma, and the California wildfires, Verizon’s bot saw a dramatic use increase from “customers who wanted to know if they were eligible for payment relief,” according to company spokesperson Christine Amodio. “Hurricane” and “wildfire” might not be at the top of any cell phone provider’s training list, but had Verizon’s chatbot not recognized the words, the company would have seemed unfeeling and out of touch.

Set User Expectations

It’s important, Kumar says, to have some idea of “what is happening to the customer before this customer started chatting with you.” For the times when you can’t, though, he stresses the importance of making sure people know when they’re chatting with a man vs a machine: “I would never advise a bot...to pretend that it is human. It’s okay to let customer know that you’re talking to a bot.”

In this way, Kumar explains, you don’t just train the chatbot, but your customer as well: “There is an understanding there’s a machine. It has limitations.” A bot will never get 100 percent of a conversation, so aligning expectations keeps people from being disappointed. And they engage with chatbots more in daily life, people — like technology — are always learning.

Related Article: 8 Ways to Measure Chatbot Program Success

A Need for Open Source

Many companies don’t have the money or manpower to invest in proprietary training. That’s why Seed Vault CEO Nathan Shedroff says chatbot developers need open-source access to each other’s work. “One of the reasons why the web expanded so quickly is that anyone could view [website] source [code]” and paste it on their own site, he explains. “That ability to rapidly build off of what other people have built was a really important component of the whole web. That doesn’t exist for bots yet,” he continues, which is why he founded Seed Vault — a decentralized marketplace for conversational user interface development.