The Wall Street Journal reported that Google is working on an algorithm to predict which of its employees is likely to leave. Putting aside for the moment whether there is some creepy aspect to the machine anticipating personal career choices, there are two aspects to this that I find interesting.
First, exactly how and why did this make it into the news? I’m not questioning whether it’s newsworthy (it is), I’m wondering about the exact path that this took from fact to front page. Was this program broadly announced at Google, and a random employee alerted the media? That would be one typical path. Was it an unannounced program, and someone in the HR department leaked it to the press? Or perhaps the HR folks proactively fed this to the press. If one of the latter two, what’s the motivation? Does it make Google look smart as a company? Does it comfort employees to know that Big Brother is watching and cares whether or not you leave? I don’t think any answer to any of these questions is necessarily a bad thing, I’m just wondering what the answers are.
Second – and more answerish this time than questioney – it’s really interesting to consider the reported inputs to the algorithm. According to the WSJ, “data from employee reviews and promotion and pay histories” comprise the formula. At first glance, this might seem logical: you would hope that reviews and compensation have some correlation to job performance and satisfaction.
Ah but then – you’ve been hoping that all your life, haven’t you? And haven’t you found that the content of employee reviews is often radically disconnected from actual performance, that the “wrong” person often gets promoted, that compensation has at best a tenuous connection to performance and satisfaction?
So the data set might not be as logical as it first seems; fortunately the magic of the sample size means that this isn’t a logic problem, but simply a correlation exercise. There is certainly some correlation between the content of that data and the incidences of employee departure, or Google wouldn’t be trying this.
But I’d bet that the correlation is much weaker than you’d think, and much weaker than Google would like. Any manager knows that reviews and compensation are blunt tools, and these are likely to be trailing indicators rather than the leading indicators of dissatisfaction that you want for real predictive power.
I worked at a company that somewhat famously had a “Love Machine” – a simple tool for employees to express appreciation for each other. Basically, as an employee, if a colleague did something that you appreciated, you could send them a point of Love, along with a one-line comment as reason for the gift. It was just a neat little way to say Thank You, and people really appreciated both giving and getting Love. And as very minor benefit, periodically the accrued Love points would be paid out to employees as a cash bonus. (Distribute that on a Friday and watch people really spread the love around.)
When this was reported externally, some people misunderstood the corporate purpose of the Love Machine. Which is to say, some people thought it had no purpose at all, except as some gross hippy dippy vibe. Those people not only misunderstood the effect on morale, but also vastly underappreciated the value of the data from the Love Machine.
Think about it: as a manager, is it valuable to know which of your employees is appreciated, and which is not? The Love Machine is certainly not the sole source of information on this, but the data gives at least some general information on both specific and aggregate cases. You can see not only who is appreciated, but who gives appreciation. You can see which departments are praised and which toil without recognition. You can see who works across departmental boundaries and who only interacts upwards or downwards in their own reporting silos.
All this is phenomenally powerful data that gives factual indication of matters that you might otherwise pay significant amounts to guess about through expensive organizational consultants. I was convinced back then and I still am today that if HR departments and managers understood the power of this data, every company would have a Love Machine. Most would call it something else, but that part is just internal marketing.
Back to Google: as far as I know, they don’t have a Love Machine. But that doesn’t mean that they are feeding the best data they have into their Who’s Leaving? algorithm. Nearly every company has somewhat similar data in their email traffic, and certainly in a big data set, email traffic would work quite well. Now, I’m not saying that the email content should be analyzed – even though many companies actually do that, I find it unacceptably creepy. I just mean the fact of communication. Every manager knows that communication is a powerful indicator (and motivator) of job satisfaction.
Who’s talking to who? Who’s on how many email lists? What is the response rate and timing and length? Is communication cross-functional, up and down and across reporting levels, inside and outside the company? All of that information is available in company email – I guess you could call it the email graph.
I would bet almost anything that this kind of data from email traffic provides a more powerful indicator of who’s leaving the company than the data from reviews and compensation. In fact, I’d bet that the email graph has better correlation to actual job performance than performance reviews!
Of course, the combination of those data sets could be even more powerful, and it’s entirely possible that Google is already doing this, though that wasn’t reported in the Journal piece.