Last year, I was working with a telecommunications software company that wanted to integrate a chatbot into its solutions. Like many companies, it was looking to reduce support costs including calls and TWC (time with customer). The company researched a handful of chatbots and quickly settled on one. Living by the fail-fast methodology that so many agile shops love to embody, it launched the chatbot into its software.

Of course the chatbot was not positioned as a bot; it was given a name (Sally) with a nice headshot. If you were on a screen and idle for a spell, Sally would pop up in the lower right-hand corner and ask if you needed help. True to the fail-fast mantra, my client didn’t test Sally—it launched Sally. Then they ate cake in the break room and celebrated how Sally would save them so much money. After that, the calls started.

At first, it was just the usual customers—the ones who always call when any of the software changes. But more calls started coming in from previously happy customers. Many customers initially engaged with Sally, but quickly learned she was limited in her ability to help. For example, if customers needed assistance with their account info, they could ask Sally how to change their password. Sally would answer a question, but then try to make conversation. Sally used the customer’s name, location, time of year, and other vital datapoints to construct a seemingly lifelike conversation.

Problems arrived when Sally started profiling customers. In one instance during the holiday season, Sally was communicating with a customer named Rhonda. Using the time of year and Rhonda’s gender, she asked if Rhonda was busy making dinner for her family. As a single, career-focused  non-cook, Rhonda took tremendous exception to Sally’s assertion that a female would be busy in the kitchen during the holidays.

In the ensuing weeks, the calls, emails, and negative social media outcry persisted. Sally lasted less than a month. Limited resources and the associated costs have not allowed my client to fix Sally.

Whenever I relay this story folks always ask why Sally was a flop. Sally was not the problem; instead, it was the data she inherited that was the issue. The problem with Sally—and so many of her chatbot cohorts—is that she was fed bad (in some cases, harmful) data that was merely transferred over to her codebase.


Fixing AI With Content Strategy

We crave Big Data, but increasingly invalidated data is flooding our AI-assisted systems. AI relies on algorithms, but in many cases, those computations contain poor legacy data. In Sara Wachter-Boettcher’s book, Technically Wrong: Sexist Apps, Biased Algorithms, and Other Threats of Toxic Tech, she writes, “Reliance on historical data is a fundamental problem with many algorithmic systems.” You think many of today’s chatbot startups are checking the validity of the data being fed into their algorithms?

One way to combat this is to use content strategy artifacts. Personas, research, voice and tone style guides, and a fully vetted taxonomy are vital tools that can be reused when configuring chatbots. Any chatbot should be internally piloted and tested before being put in front of customers. Usability testing your chatbot is a prime area in which you can revisit your voice and tone guidelines to ensure its language is adjusting, depending on where users are in their journey. And just to be safe, have your chatbot skip the small talk.