Paresh David

OAKLAND, Calif. (Reuters) – Microsoft Corp LinkedIn saw an 8% increase in subscription revenues after it equipped its sales staff with artificial intelligence software. This software not only helps predict cancellations but also shows how the company got there.

Introduced last July, the system will be detailed in a LinkedIn blog entry on Wednesday. This is a significant breakthrough for AI in showing its “work” in a useful way.

Although AI scientists are not able to predict all business outcomes accurately, they have discovered that the AI might need another algorithm to help them.

The emerging field of “Explainable AI,” or XAI, has spurred big investment in Silicon Valley as startups and cloud giants compete to make opaque software more understandable and has stoked discussion in Washington and Brussels where regulators want to ensure automated decision-making is done fairly and transparently.

Artificial Intelligence technology (AI) can lead to societal biases, such as ones around culture, race and gender. AI scientists see explanations as an essential part in mitigating these problematic outcomes.

The Federal Trade Commission and U.S. consumer protection regulators have been warning for the past two years about the possibility that AI which isn’t explainable might be examined. Next year, the EU could approve The Artificial Intelligence Act. This comprehensive law requires that all users are able to make automated predictions.

Proponents of explainable AI say it has helped increase the effectiveness of AI’s application in fields such as healthcare and sales. Google (NASDAQ: Cloud) offers explainable AI services. These include telling clients who are trying to improve their AI systems what pixels they should use and when the best training examples will be available.

Critics say that explanations for why AI predicts what they do are not reliable because AI technology isn’t good enough to interpret them.

LinkedIn and the other developers of explainable AI recognize that every step in this process — analyzing predictions, creating explanations, verifying their accuracy, and making them useful for users — still needs improvement.

After two years of testing and errors in an application that was relatively low stakes, LinkedIn claims its technology is now practical. This is evident in the 8.8% increase of renewal bookings for the current fiscal year, an improvement on the normally expected growth. LinkedIn didn’t specify the dollar value of this benefit, but it described it as substantial.

LinkedIn used to rely on the intuition of its salespeople and some automated alerts regarding clients’ acceptance of their services.

The AI now handles analysis and research quickly. It is called CrystalCandle on LinkedIn and it allows salespeople to spot unnoticed trends. 

LinkedIn claims that explanation-based recommendation has grown to over 5,000 sales staff, spanning recruitment, marketing, and education.

It has provided insights that have helped salespeople navigate the conversations with potential customers. It’s also helped new salespeople dive in right away,” said Parvez Ahammad, LinkedIn’s director of machine learning and head of data science applied research.

To EXPLAIN or Not to EXPLAIN

LinkedIn provided the first predictions that were not explained in 2020. An accuracy score of about 80% indicates how likely it is that clients will renew soon.

The salespeople weren’t completely sold. Salespeople were not fully convinced by the team that sold LinkedIn’s Talent Solutions recruitment and software for hiring. They were unsure how to adjust their strategy especially since there was no chance of clients renewing.

In July last year, the team began to receive an auto-generated paragraph about the key factors which influenced their score.

AI, for example, determined that customer is likely to upgrade due to its growth of 240 workers in the past year. Candidates had also become more responsive by 146% over the month.

An index which measures the success of a client with LinkedIn has risen 25% over the three-month period.

Lekha Doshi is LinkedIn’s vice-president of global operations. She said that, based upon the reasons, sales representatives direct clients to services, support, and training that enhance their experience while keeping them spending.

However, some AI specialists question the necessity of providing explanations. Researchers say they could cause harm by creating a false sense security or prompting designers to make sacrifices in order to improve predictions.

Fei-Fei Li co-director Stanford University’s Institute for Human-Centered Artificial Intelligence said products like Tylenol or Google Maps have an inner working that is not well understood. These cases have been the subject of rigorous monitoring and testing, which has dispelled any doubts as to their effectiveness.

Daniel Roy, an associate professor of Statistics at University of Toronto, stated that AI systems could also be considered fair, even though individual decisions may not be easily understood.

LinkedIn states that it’s impossible to evaluate an algorithm’s integrity without knowing its thoughts.

Additionally, it maintains that CrystalCandle tools could also be used to aid AI users in other areas. AI could help doctors predict who is most at-risk of developing a disease. People could also be informed why AI recommends that they not be granted credit cards.

Been Kim from Google, an AI researcher, stated, “The hope is that explanations show whether a system aligns to concepts and values one would like to promote.”

She stated that interpretability is ultimately about facilitating a conversation between humans and machines. We need this if we are to allow human-machine cooperation.