Amazon is testing two AI-powered systems as customer support agents, one of which will automatically field customer calls without human intervention and the other which will help human service agents respond more quickly to requests. Amazon applied-science manager Jared Kramer said the AI agents rely on machine learning and refer requests they can’t handle to human agents. Amazon also debuted its cashierless “Go” technology in a large grocery store, in advance of possibly licensing the system to other retailers.

VentureBeat reports that Amazon’s new AI agents differ from the company’s “old flow chart system that specified responses to particular inputs.” The company said that its “customer-facing system [is] using a template ranker — where an AI model chooses among hand-authored response templates — that allows it to control the automated agent’s vocabulary.”

It will soon test a “generative model that crafts responses to replies on the fly.” Templates are “general forms of sentences, with variables for things like product names, dates, delivery times, and prices.” Because the agent has been pretrained on a dataset of interactions and “seen many responses that don’t fit its templates, it’s learned over time several general principles for ranking arbitrary sentences.”

Amazon researchers trained “separate versions of each model for … return refund status requests and order cancellations.”

In randomized trials, the new AI agents “significantly outperformed” the old rule-based systems. The trials combined “whether the automated agent successfully completes a transaction (without referring it to a customer service representative) and whether the customer contacts customer service a second time within 24 hours.”

Gartner predicted that chatbots will “power 85 percent of all customer service interactions by the year 2020.”

The Wall Street Journal reports that Amazon first opened Go convenience stores in 2018 but, said Amazon vice president of physical retail and technology Dilip Kumar, “improvements in camera technology and its use of algorithms have allowed it to build a 10,400-square-foot market … about five times bigger than the largest existing Go store.”

“There’s no real upper bound,” said Kumar. “It could be five times as big. It could be 10 times as big.”

Sources said that Amazon has “recently been in talks with potential partners and is targeting retail options including convenience stores and shops in airports and sports arenas.” Among potential revenue models, Amazon has mentioned “a fixed licensing fee or a revenue-sharing agreement.” Sources also reported that, “Amazon has considered offering to install and set up equipment to enable its cashierless technology and market it as a service.”

Gartner technology analyst Joanne Joliet noted that, “retailers could use the technology for a variety of purposes, including to track inventory or customer habits, that may stop short of having a full-scale cashierless store.” She added that, “Amazon would likely have to promise to set up a firewall that prevents its other business units from accessing potential partner data for them to feel comfortable.”

According to eMarketer analyst Andrew Lipsman, “Amazon’s Go stores are unlikely to be profitable at the moment, but the investment could pay off through licensing agreements or other means.” Among other companies exploring similar technology are Berkeley, California-based startup Grabango Co. and retail giant Tesco.