The logo of, the AI powered sales team coach
Product | March 18, 2021

Improving AI Models Through Feedback in Plain English

Today we are announcing the launch of the Suggestions Director which empowers contact center managers to direct the suggestions towards ideal behavior. Suggestions are the real time suggested responses that contact center agents can send to their customers instead of typing the whole response. Now managers will be able to provide feedback in plain English giving us ideas on how to improve the current agent messaging and provide alternative responses saying what the agent should have said instead. Our expertise engine takes this description of the feedback, analyzes it, and fixes the underlying artificial intelligence models, and updates suggestions improving our customer’s KPIs! This approach provides unparalleled power to managers by empowering them to take control of the content of suggestions, leading to better behaviors and compliance by the agents. In a test with a customer, this approach has improved their sales conversion by 34%!

In this blog post we would dive into what this is and how it works.

Status Quo: Currently there are two main categories of suggestions models:

  1. Rule-based models – With such models we explicitly write the rules of dialog. For example, one could write it as – “Recommend the gold plan if the customer expressed a need for high speed”. The problem with such an approach to model really complex interactions is that with real conversations, the number of such rules blows up very quickly, making it impractical. Even after writing many rules, they only cover a small portion of possible dialogs, and moreover, those suggested dialogs sound very robotic.
  2. Deep models – Here we provide a powerful machine learning model a lot of data and it automatically figures out how to generate a response. Such a model is easier to create since it takes away the work of writing many rules.The problem with such a model is that it can only be as good as the data it is trained on, and usually in contact centers not all agents speak perfectly every time. This leads to lower quality of suggestions.

Recently, the deep models have become very popular leading to an improvement in the quality of suggestions. Such models are “black-box” meaning it is very hard to control them or explain why they are making their decisions. This leads to a situation where it gets hard to improve them even when we know they are not ideal.

Cresta’s Answer: We want to provide our customers complete control and power to tailor their conversations so that they are always ideal for every situation. In fact, this was highly requested by our customers. Towards this goal, we have invented an unprecedented technology where our expertise engine directly takes in the feedback from our customers about what should be improved, and how we can improve the conversations so that with that feedback, the conversations resemble the ideal one.

The breakthrough with Suggestions Director is that customers or experts don’t need to write rules like in the rule-based model. We instead take in feedback in only plain English explaining what is wrong with the agent’s messages, how to fix them, and an example of an alternate good response. The types of feedback we can take are very broad, ranging from improper phrasing, improper grammar, incorrect pricing, not following the call-flow, etc. Given this feedback, we make a powerful Deep Model to understand what that feedback means and correct itself to follow the new guidelines. The new updated suggestions are not just a copy of the provided example good response but they are attuned to what is ideal depending on the context of the conversation. Moreover, managers only have to give feedback on a single occurrence and our expertise engine automatically figures out all the situations where that feedback is applicable.

The best part about Suggestions Director is that it delivers great results for our customers. In an A/B test with a customer, this technology led to a staggering 34% increase in sales conversion! This happened because our expertise engine took in the feedback from the contact center managers, who had the best ideas and insights and figured out the best thing to say at crucial moments in the chats. Agents also loved the better suggestions, resulting in 40% higher suggestions usage than before. Agents also said that the suggestions were becoming “more human”, have better timing, and are more accurate.

Driving Results – Suggestions Director amplifies and spreads expertise throughout your contact center organization through higher quality, accurate suggestions. This leads to quicker training of agents and better behavior during key sales moments leading to much higher KPIs like revenue per chat or sales conversion. Suggestions Director makes suggestions not just explainable but also controllable by our customers. If you can explain in words what should be improved with the current agent messaging then our expertise engine can learn from it and provide that insight to all the agents through better suggestions!

Currently, this feature is in limited availability, please contact the Cresta team if you are interested in implementing Suggestions Director for your organization.