Though chatbots have been growing in popularity in recent years, they are still somewhat limited in their capabilities. One of the things holding them back is the difficulty in programming chatbots with advanced machine learning and natural language processing that would enable them to communicate in a way that more closely resembles human speech. One of the reasons chatbot users cite for being dissatisfied is that chatbots often get stuck and need to be bailed out by live customer service agents.

Facebook hopes to address these issues by developing chatbot programs that can negotiate and reach compromises that are satisfactory to both customers and the company the chatbot represents.

The project

Facebook’s artificial intelligence branch, FAIR (Facebook Artificial Intelligence Research), are working on a project that creates an open-source code that any company could use to make their chatbots that operate on Facebook messenger better negotiators. In order to facilitate the necessary machine learning for the project, Facebook collected transcripts of thousands of negotiations between real people and then fed them into the program so that their chatbots, or “dialog agents” as Facebook refers to them, could then discern patterns of speech that led to effective outcome during negotiations. They further developed the accuracy of dialog agents by having them enter into negotiations with each other and taking note of what led to successful outcomes and what did not.

Finally they tested the dialog agents in conversations with real people. In these experiments conducted by FAIR, the majority of people didn’t even realize there were chatting with a bot and not a human. Their best dialog agent was found to be as effective as human negotiators achieving a better deal about 50% of the time.


The majority of today’s chatbots are rather limited only responding to simple questions or helping people perform relatively simple tasks. FAIR’s work on developing chatbots that can effectively negotiate in a way that’s indiscernible from human speech is a powerful first step towards the ability to develop more advanced chatbots. As machine learning and natural language processing research advances, we could eventually see chatbots that can effectively handle increasingly complex issues for customers. This will free up human resources and lessen wait times for customer service issues that can’t be handled by bots. For the time-being, chatbots will continue fielding the easy questions and routing the harder issues to live agents but that could change soon.