Reduce Bias in Einstein Bots

Build bots with your target audience in mind.

Teams run the risk of losing sight of the target audience if they go too wide or too narrow in implementation. For example, bots that are written generically to handle the widest range of audiences can lead to frustrating interactions for groups whose speaking patterns vary from the assumed standard dialect. Teams can also design too narrowly and use too much slang or jargon, which users can interpret as out of touch, or even on the verge of cultural appropriation.

Fine-tune your bot script by sharing it with stakeholders from your target audience and revising it based on their feedback. You can also learn more about the best ways to craft bot conversations by taking the Conversation Design Trailhead and hiring a conversation designer to review your work.

Build a balanced and diverse intent model for accurate intent recognition.

Intent models with unbalanced data or a lack of diverse data can introduce bias if the bot moves the conversation with users in unexpected directions. Unbalanced data models favor the intent with more utterances, potentially leading to incorrect intent recognition and drop off. Bots that don’t understand the many ways to phrase an intent move into a confused state, which can frustrate the user.

Intent models are balanced by including roughly the same number of utterances for each intent. The minimum number of utterances varies by language, but in all cases more utterances result in a more accurate and repeatable intent recognition. You want your utterances to include many different ways to say your intent. Additionally, you want them to be varied across multiple dialects, slang, misspellings, and speech impairments to ensure that all audiences are included.

Many template bots in Einstein Bots come preloaded with balanced intent data for specific intents, but you can source your own through internal crowdsourcing or with a survey company. Make sure that the people that you ask to generate utterances for your intent model capture the makeup of your target audience. For example, build a bot for college students that reflects the audiences that are attending college, not just people between the ages of 18–24. Learn more about building a great intent model at Use Intents to Understand your Customers.

Build a bot that considers accessibility and channel specifics.

Bots are designed to exist on a specific channel. Each major channel, including Chat, SMS, Facebook Messenger, WhatsApp, and Slack, has their own ways to make their channel accessible, so building a bot without researching these methods can lead to a poor final product. This issue is problematic when your bot deploys on more than one channel at a time.

We recommend that teams test their bots on the intended channel before deployment and turn on accessibility tools to confirm that the bot works for all customers. If a bot is deploying to multiple channels, it’s best to design the bot to fit the most restrictive channel. For instance, customers viewing a bot with a menu deployed on SMS must type their choice instead of clicking the menu item. In rare instances, the customization can be too different between channels to share one bot. In those cases, we recommend cloning the bot before customizing.