The Trials of Training Your Chatbot and How to Avoid Them


Bookmark and Share

Puppies are adorable. Puppies are also a lot of hard work. Before you can reap the benefits of man's best friend, you need to train your new four-legged family member to obey specific commands (sit, stay, etc.), teach it to walk on a leash, and housebreak it.

A brand-new chatbot is a lot like a puppy. To succeed in conversational commerce, you either need to train your bot to interact most effectively with customers or use a commerce-specific solution that is pre-trained and equipped to resolve issues. The danger is that your customers aren't going to be as forgiving of an unhelpful chatbot as they would a fluffy little ball of joy who had an accident on their favorite pair of shoes.

In fact, lack of appropriate pre-deployment training is one of the key sources of failure for companies that rushed too quickly to embrace chatbot technology without doing their homework.

With that warning in mind, let's take a look at the challenges you'll need to meet to train a successful chatbot.

Training a chatbot isn't necessarily a smooth process, and simply turning your bot loose and hoping for the best doesn't end well. Just ask Microsoft. When the tech giant launched Twitter chatbot Tay last spring, it took less than 24 hours for Twitter users to teach the bot how to be a raving racist.

You can't let customer service slip.

The catch-22 companies face is that bots need exposure to real (or seemingly real) customer interactions to learn, but as the bots are learning, they aren't smart enough to provide the level of customer service that customers expect. Exposing customers to a poorly-trained or under-trained chatbot that delivers an unsatisfactory experience can have long-term consequences for your company. More than 80 percent of consumers report that they've left a company after a single negative experience, with 55 percent saying it was the inability of the company to solve their issue in a timely manner that prompted them to leave.

Training is labor intensive

Just like you can't expect your exuberant new puppy to be fully housebroken after a trip or two outside, you can't expect a chatbot to learn its role overnight. Check out the detailed documentation IBM provides for specific performance metrics (which you'll need to establish before training begins). Getting a chatbot ready to make its customer debut is an iterative process that requires time and energy for testing, analyzing performance, and, most crucially, refining the bot's natural language capabilities based on previous results.

Domain expertise is necessary.

How do your customers talk about your product? What are their most common customer service questions or issues? If you aren't already an expert in the conversations of the domain in which your chatbot is going to operate, feeding it the accurate natural language data it needs to start recognizing and responding to common customer queries is going to be difficult.

Consider, for example, the weather chatbot Poncho, one of the early Facebook Messenger bot failures. Users ended up frustrated with the bot because it couldn't handle seemingly straightforward weather-related queries or requests, instead responding with pithy non-sequiturs, like a hard-of-hearing elderly relative who can't quite make out what you're trying to say but is determined to carry on the conversation anyway. As the CEO of Poncho astutely noted after announcing that beefing up the bot's natural language processing capabilities was a priority, "Tolerance for a mediocre bot is much less than for a mediocre app."

Understanding customer needs isn't enough.

Although launching a bot that truly understands customers is no small achievement, recognizing intent alone isn't enough to solve customer problems. Think of the frustrating reality of complaining to an entry-level customer care rep about a poor experience with a particular company. While she is able to understand and empathize with your plight, she might not be empowered or have the authority to give a refund or otherwise solve the issue to your satisfaction without involving a superior. For a chatbot to be truly effective, it needs to be connected to the specific tools it can use to solve customer problems in the moment, with no buck-passing involved. That means connecting it to order management, customer data, product information, etc. Customers expect a clear path to problem resolution when they interact with your company, so your bot should be equipped with the resources it needs to deliver that.

The key component to chatbot success is deploying within a specific, well-defined domain, with specific skills the bot can use with a high success rate. Though building, training, and deploying is model to launch your chatbot, other options are available, from working on a fully custom bot with an agency partner to using pre-equipped and trained platforms that have specific use cases and focus. Both of these options will substantially reduce your risk for budget and timeline overruns and get you launched with what you have in mind quickly and with a strong opportunity for success.

Investment levels differ depending on how much functionality you desire and whether your selected partner has already deployed similar bots. Be especially diligent when identifying the connections your bot will need to make to your existing systems and solutions and whether your partner has built components for these or will be building all connections from scratch. This will significantly impact budget and timeline.

Ultimately, the decision is yours, but don't let the need to train AI stop you from taking a proactive approach and taking steps toward identifying your best go-forward strategy.


Fang Cheng is co-founder and CEO of Linc Global. She previously co-founded a business acquired by Amazon and had multiple years of experience in the investment world.