Beyond finding the best, most engaging messages, the chatbot also learns to increase the length of the conversation by selecting the best message sequence, also finding the best point in the dialog to include the call-to-action button.
For example, if one message such as “How does it work?” was consistently getting a good response, but was repeatedly followed by the customer ignoring the next message then the sales agent would detect that “How does it work?” may be a conversation killer – it is seeking to optimize for length of the dialog, and button taps on the call-to-action.
The sales agent uses the reinforcement method of machine learning to explore and adapt to its environment (customer interaction) – this allows for winning strategies to be repeated while new ones can also be continually tested. Since customer preference can be unpredictable this can help find a balance between exploring new ways to entice customers vs. exploiting previously found successful strategies.
This “convergence” point is when the optimal behavior of the agent has been achieved, though at any time the “explore more” strategy can be turned up.
The agent holds 16 persuasive statements about the product or service divided into 2 categories: Attribute and Non-Attribute.
See this example of the selling points to provide in order to load Rally.
1) Product Attribute persuasive statements (for high prior knowledge customers), contain specific details for example:
“The i7-6700 incorporates the HD Graphics 530 IGP operating at 350 MHz and a turbo frequency of 1.15 GHz. This chip supports up to 64 GiB of dual-channel DDR4-2133 memory.”
(in response to the prompt: “What are the specs?”)
2) Non-Attribute persuasive statements (for low prior knowledge customers), contain details about the general benefits or how it will make them feel, for example:
“You might think it's the number of cores, but for the most part desktop Core i5 processors have four cores, just like i7s.”
(in response to the prompt: “Is i7 faster than i5”)
These statements can be written in the form shown above or using an argumentation framework with common arguments or “barriers to purchase” (in place of question prompts) being refuted by counter-arguments such as “Is the CPU slow?” followed by a persuasive argument #1 above – the 16 persuasive statements can also be a mix of question/answer and argument/counterargument – the agent will find the best messages.
At the start, the agent will try to determine if the customer is high prior knowledge or not.
By dividing the persuasive statements into groups with the maximum conversation length of 4-5 messages there are tens of thousands of possible dialog trees.
This number includes messages with or without the call-to-action. It may be determined that a customer segment prefers to get right to the action and taps “Sign me up” or “Order Now” in message #2, or perhaps they need a little more nurturing.