Good bosses have an uncanny ability to sense when employees are unhappy and work with them to fix problems in the office before it’s too late. At VMware in Silicon Valley, they let the machines figure it out.
VMware has been testing a new prediction technology from Workday, which makes software for human resources departments. The system delivers notifications about when employees might be getting ready to quit, and allows managers to intervene before it's too late. It looks for trends within employee activity, when promotions were last handed out, regional factors, changes in the industry and other data to make its predictions. The recommendations can improve over time as employers train the system.
"We've had some great results to date with the data,” Amy Gannaway, VMware’s senior director for worldwide human resources information systems, said at a Workday conference in September. The tool gave VMware "a very high percentage" of accurate predictions for which employees would leave the company, she said.
Workday can do this because the technology underpinning it is based on machine learning—a bag of advanced statistical techniques that lets companies lay out complex problems, spot patterns and come up with predictions. Machine learning has been available in one form or another for decades, but its commercial uses have traditionally been the exclusive domain of the richest, data-stuffed companies like Google, Facebook, Microsoft and Netflix. Finally, younger tech companies, including Airbnb, Box and Workday, are able to hop on the predictive-cloud bandwagon.
"In the past five years, because of the steady advance of computers and storage and so on, everybody has sufficient data," says Alexander Gray, the co-founder of Skytree, which provides machine-learning tools to businesses. “Now, we all have the same magic that Google has."
Box, a file storage provider that's expected to go public next year, acquired a startup last year named DLoop to help develop predictive capabilities for its applications. DLoop's technology can analyze the contents and attributes of a document, and automatically tag them. Box has been testing this with the company's own legal department to figure out which files may contain confidential information that would require extra precautions with how they're stored and shared. The experiment looks at about 100 documents a day. And it works, says Sanam Saaber, Box’s senior commercial counsel. "We have enough confidence in this product to go live on a day-to-day basis,” she says.
Similar to the Workday system, Box’s legal team had to train the DLoop software to improve the recommendations. It got so good at classifying documents that it reached a 100 percent accuracy rate, according to Saaber. "It's actually scary how quickly it's learned,” she says.
Box plans to eventually package and sell the technology to customers, says Aaron Levie, the company’s chief executive officer. "As more and more data goes into our platform, we can produce important and useful insights to customers around how their data is being shared, who can be seeing what kinds of content, where are potential security anomalies, or how can we dramatically accelerate a business process by connecting the dots for you,” Levie says.
Computer predictions aren’t just about making office life a little more pleasant. Airbnb uses a variation of these algorithms to predict which renters and guests would be the best fit. The room-rental site says the technology has improved matches by 4 percent. Airbnb is currently developing a system to look at the photos of homes uploaded to the site and figure out how “attractive” they are to customers. “We are trying to promote listings with more attractive images,” says Maxim Charkov, the search lead at Airbnb.
Eventually, Airbnb may offer a digital interior designer that predicts ways to enhance listings and spruce up homes to increase bookings. “Maybe you should improve your images; maybe you should provide these amenities that are popular,” Charkov says. “We want to bring the insights back to the host."
Workday’s ability to identify which employees have one foot out the door also originated from an acquisition. The company bought Identified in February after getting a demonstration of the startup’s data-powered crystal ball. Mohammad Sabah, who was the chief data officer at Identified, had joined the startup from Netflix and Facebook. Sabah had discovered that the same techniques Netflix uses to recommend movies with the familiarity you'd expect from a clued-in friend could be used to determine what’s going on inside of a company. "The domain is so different, but the techniques and the algorithms and the tools are general,” Sabah says.
By combining company data on employee hiring, promotions, relocations, compensation, employee satisfaction surveys, managerial decisions and job cuts with public data sets like the standard of living in the region and workforce demand for certain skills, Workday can spot patterns. Businesses can input decades of historical staff data into Workday to inform and customize the system’s recommendations. In one case, Workday analyzed more than 1 million data points for 100,000 employees across 25 years to come up with employment suggestions. To train the software, companies must look back on worker-retention predictions and give the software an electronic pat on the head for ones it got it right and a virtual swat with a newspaper for those it got wrong. The system learns over time how each company works and, like an experienced HR employee, develops a gut feeling for which people the company needs to keep a closer eye on.
Bob Pasker, a technologist who helped define how the computer industry architects the modern cloud, expects these kinds of machine-learning applications to soon become ubiquitous. Their potential to improve corporate efficiency means no CEO will want to operate without them. "This is what the computer industry does," Pasker says. "We take hard problems that have been solved by scientists and turn them into tools for regular human beings."
(Updates with scale of Workday software in the 12th paragraph. An earlier version of this story corrected Sabah's title at Identified.)