AI is also beginning to help managers peer into personal aspects of job performance that used to be left up to managers' instincts and observations -- for instance, attitudes toward the job. Veriato analyzes email and other messages, looking at words and phrases employees use. Then it scores those expressions for positive or negative sentiment. The system can set a sentiment baseline over time, and then calculate a daily score for each employee.
It can send an alert if a worker's use of certain language exceeds a threshold, or if it detects any change in tone or a shift in relation to a group of employees. The customer can evaluate the context in which the expression occurred -- including screenshots captured by the system -- to decide how to proceed. "If the tone of a typically happy person suddenly goes negative, that may be an alert that they're at risk of flight, insider threat or even just a productivity problem that needs remediation," says Veriato Chief Security Officer David Green.
KEEPING TOP PERFORMERS ON BOARD
Some AI aims to predict when employees may be winding down their career at the company -- and advises how to keep them on board.
Products from Entelo, International Business Machines Corp. and Workday, as well as Microsoft Corp.'s internal management system, look for patterns identified by researchers and their own software to predict when workers are likely to jump ship.
For instance, Workday's retention-risk analysis feature, which made its debut in April, bases its analysis on data from selected customers representing 100,000 individuals over 25 years, says Leighanne Levensaler, a senior vice president of corporate strategy at Workday. It tunes itself to a given customer, calculating a risk score for individual employees based on roughly 60 factors including job title, compensation, time off and time between promotions.
The software also suggests potential next steps in an employee's career path based on what other people in similar situations have done, so managers can move proactively to retain valuable workers. Ms. Levensaler says the retention-risk score is best thought of as one element of a broader picture, "a pattern we see that's instructive for you in your conversation, but you're still managing."
THE LIMITS OF AI
For all their promise, these systems raise a number of issues. Some are evident today, in the early stages of adoption, while others may take time to become clear.
Privacy is an obvious concern when tracking employees, particularly personal behavior. Systems that sort job candidates also raise questions. Entelo's may emphasize people with a large online footprint; SAP's might prefer those who best match characteristics of people who were hired in the past.
Entelo Chief Executive Jon Bischke acknowledges the possibility that the data set in his company's recruiting system is biased, but says it doesn't necessarily affect his customers. "Our area is hiring for highly skilled jobs," he says. "The vast majority of candidates [in that area] have a presence on the web."
Mr. Mueller of SAP says that, in practice, Resume Matcher reduces bias by highlighting a more diverse selection of candidates than managers otherwise would have considered. "Many recruiters were surprised when they saw the candidates, but when they looked deeper, they could see why the system selected them," he says. For instance, one manager testing the system was taken aback by the high ranking of candidates from China that he otherwise would have overlooked; he was unfamiliar with the top Chinese schools where they were educated.
Beyond that, the use of such tech in workplaces is new and not widely proven -- and in many cases it may not be easy to determine that a machine's insight was sharper than a human would have perceived. That's a concern when inaccuracy in an AI report -- painting someone as a poor performer, for example -- might set back an employee's career.
Forrester Research Inc. analysts David Johnson and J.P. Gownder voiced such concerns in a recent report. The authors argue that employers' ability to gather data about employees has outstripped managers' capacity to interpret it properly, opening the door to a variety of counterproductive practices.
Managers tend to pay attention to what they can measure, Mr. Johnson says -- hours spent in workplace apps, say, rather than quality of output. Focusing on individual performance may lead managers to overlook hindrances to productivity that are systemic.
"I don't want to cast negative light on these companies" selling data-driven management tools, Mr. Johnson says in an interview. "They don't have control over how people use their products. I'm just pointing out the risks."
Some management professionals share those worries. Kenny Mendes, who runs recruiting at a software startup that hasn't yet launched publicly, previously directed human resources at the online work-collaboration service Box Inc. (He is an adviser to Entelo.)
Mr. Mendes spent two years experimenting with ways to predict and maximize employee success using a statistical programming language and "lots of spreadsheets." The experience led him to believe the problem is too complex for the current generation of software.
The limitations of current approaches, he says, boil down to the difficulty of drawing valid conclusions from incomplete data.
For instance, measurements of employee performance at any given company are based on the set of people hired and lack information about candidates who were passed over -- or weren't even interviewed -- who may have, say, produced more in less time. Aggregating data from many customers, as some larger vendors including SAP and Workday do, can reduce bias, but the problem remains that different companies may not track the same variables in the same way, and subtle but important ones are likely to be missing.
Moreover, management systems can't account for conditions outside the office that may energize or depress individual employees at work -- especially personal conditions that can shift unpredictably. On top of that, human psychology is a wild card; if workers know their overseer is tracking hours on the job rather than output quality, they may spend an extra hour a day at the office simply chatting by the water cooler.
"Even the smartest people will make bad decisions with bad data, and I think we have a lot of bad data in this process," Mr. Mendes says.
He favors technology that helps managers "without disqualifying people." However, he believes the most effective personnel-management tools are references, work-product tests, and strong personal relationships between supervisors and their charges.
Mr. Greenwald is a reporter in The Wall Street Journal's San Francisco bureau. Email him at [email protected]