5 steps to training a chatbot

So you’re thinking of implementing a chatbot, like every other company on the planet. Now, the issue is training the thing.

There has been an endless stream of coverage about all of the wonderful things chatbots are going to do for business by automating conversations with customers. The problem is, bots are only as good as the training that goes into them, and training isn’t something you just knock off in a few days.

Many bot startups seem to want to treat bots like an IVR — choose the top 20 use cases and train your bot around those needs. Considering IVR is probably the most-hated piece of technology invented in the last 50 years, replicating that process is probably not a great idea.

Very few, if any, bot vendors have actually operated these things out in the wild — with millions of customer interactions over a number of years — so it makes sense that we’re not yet hearing about the best practices when it comes to training a bot. And the reality of training a bot is a little more complex than might be apparent at first glance.

At the highest level, you need to consider all of the various channels that customers are using to contact you (such as Facebook Messenger, Twitter, Kik, WeChat, web chat, mobile apps, Amazon Alexa, and many others), the unique UI capabilities of each channel, and the multiple languages to be supported. (You definitely don’t want to train in multiple systems.)

This is a little more complex than just plugging in the top 20 use cases, so let’s look at the best way to train a bot to deliver real value to your enterprise and customers.

Step 1: Pre-training your bot

This is what many bot companies are pitching as the entire training program, when it is really only the first step of a real training plan. This is where you go to your customer care or tech support department and see what they think are the top reasons your customers are contacting your company. By sampling the chat logs and call recordings and digging through the typical support material that the reps use you will get closer to understanding the customer intent. This is a decent place to get started.

Step 2: Analysis of all customer interactions

Many bot companies assume that you know what your customers want and that if you don’t it is your problem to figure it out. However, a tool that analyzes customer questions in real time is a key piece of any bot solution. The analysis will tell you what your customers actually want to do with your bot and will let you focus your training efforts where it pays the biggest dividends. This is the step in determining bot priorities relative to customer needs, and it is the most important part of any bot training framework.

Step 3: Training the bot

With all of the rich information coming from analytics, you can get a handle on how to train the bot. A good bot solution will now let you match your bot to the best platform and take into consideration each platform’s capabilities and limitations. What you present in Facebook Messenger won’t work in a pure text messaging bot, and what you can present in Twitter may not work in your web chat UI.

A well-architected system will let you train for both conversational and Q&A types of platforms in one system. Multiple languages should be automatically generated or updated if you have created a new scenario or just updated an existing one. Training multiple bots to solve this problem is insanity, as described in this previous VentureBeat article.

Step 4: Measuring effectiveness

Use every question from every customer from every channel every day to determine how effective your bot is. There are many ways to measure your customers’ level of satisfaction during their interaction with the bot. Over time, you will determine normal thresholds for each parameter, and you will then be able to use anything outside the norm as an indicator of failure. Your false positive rate should be monitored closely here. Delivering the wrong answer is far worse than not delivering an answer at all, so this is critical. If you are delivering the wrong answers more than 3 percent of the time, your system should be taken back to the drawing board.

Step 5: Continuously improving

The process of training your bot never ends. Even if you have a simple pizza-ordering bot, you’re going to have to continuously learn from your customers how they want to order and add new platform support and new products. If you have a more complex business and are using your bot for customer service, you should plan to invest considerable effort in ongoing training.

Don’t get caught up in the bot hype and deploy something that is going to create more work in the long run. Remember, a fully thought out customer strategy is the key to customer satisfaction and economic benefits.