| CPC G06Q 20/203 (2013.01) [G06N 20/00 (2019.01); G06Q 20/202 (2013.01); G06Q 30/0601 (2013.01)] | 20 Claims |

|
1. A method comprising:
training a software agent comprising a machine learning algorithm using conversation data that includes hundreds of thousands of conversations;
after training the software agent, deploying the software agent on a server;
receiving, by the software agent, an utterance from a customer, the software agent trained to engage in a conversation with the customer to take an order;
creating, via the software agent, a cart associated with the order;
converting, via the software agent, the utterance to text;
providing, to an encoder of a natural language processing pipeline, as structured data:
an order context including an interaction history between the software agent and the customer; and
a conversation state;
predicting, by the encoder, based on the text and the structured data, an utterance vector;
determining, based at least in part on the utterance vector, that the software agent is untrained to respond to the text;
establishing a connection between:
the server hosting the software agent; and
a point-of-sale device that is associated with a human agent;
receiving, from the human agent, a modification to order data displayed by the point-of-sale device;
resuming, by the software agent and based at least in part on the modification to the order data, the conversation with the customer, wherein the human agent does not directly interact with the customer during the conversation between the software agent and the customer; and
after determining, by the software agent, that the customer has completed the order, retraining the software agent based at least in part on:
the conversation between the software agent and the customer;
the text;
a content of the cart; and the modification made by the human agent,
the retraining causing an improvement in an accuracy of the software agent in order taking.
|