A manager who feels their team but cannot count it is flying blind; a manager who counts their team but cannot feel it is reading the wrong instruments. People analytics is the discipline that sits between those two failures, using workforce data to answer questions you would otherwise answer by gut alone: who is likely to leave, where hiring is breaking down, which managers grow people and which quietly burn them out.

The quick version

  • People analytics is applying data and analysis to decisions about hiring, developing and keeping people, the same evidence-based instinct that finance and marketing have had for years, pointed at the workforce.
  • Workforce metrics are the individual numbers it runs on: turnover and retention, time-to-fill, engagement, internal mobility, cost-per-hire, and the like.
  • The best-known proof it works is Google's Project Oxygen, which set out to show managers don't matter, and found the opposite, in the data.
  • The biggest trap is Goodhart's Law: the instant a measure becomes a target people are paid to hit, it stops measuring reality and starts measuring how clever they are at gaming it.

The idea in depth

People analytics is not new technology so much as an old idea finally reaching HR: that decisions about people deserve the same evidence other functions take for granted. The clearest early statement came from Thomas Davenport, Jeanne Harris and Jeremy Shapiro in "Competing on Talent Analytics" (Harvard Business Review, October 2010). They mapped how leading firms move up a ladder, from basic descriptive metrics that report the organisation's health, through analysis that connects people data to business performance, to predictive modelling that forecasts future headcount and flight risk. Their point was not "collect more data." It was that the companies pulling ahead were the ones connecting their people numbers to outcomes the business already cared about.

So the move is to start from a decision, not a dashboard. Before you stand up a single chart, name the question you can't currently answer well, "why do good people leave in their second year?", "is our hiring slow because of us or the market?", and let that question decide which metric earns its place. A dashboard built question-first stays small and gets used; a dashboard built data-first sprawls, impresses no one, and quietly rots.

What the data actually found: Google's Project Oxygen

The most cited demonstration of people analytics is a study that set out to fail. In 2008 Google's People Operations team launched Project Oxygen with a deliberately provocative hypothesis: that managers don't matter, and that the quality of a manager has no real effect on a team. Mining performance reviews, employee surveys and double-blind interviews, the statisticians found the opposite, managers mattered a great deal, and went further, distilling what the best ones actually did into a short list of behaviours, led by being a good coach and empowering the team rather than micromanaging it. Google published the approach openly through its re:Work "Following the data" guide.

What makes Oxygen worth copying is not the eight behaviours, they will sound familiar to anyone who has read a management book, but the method. Google didn't assert that coaching matters; it found that it did, in its own data, against a hypothesis that it wouldn't. So the move for a normal team without Google's data-science bench is smaller but the same in spirit: pick one belief you hold about your people ("our best hires come from referrals", "people quit because of pay"), then go and check it against the records you already keep. The discipline is letting the data argue back.

flowchart LR
  A(["A real people decision
(who'll leave? why slow to hire?)"]) --> B(["Pick the metric
that answers it"])
  B --> C(["Check it against
data you already have"])
  C --> D{"Does it confirm
or challenge your gut?"}
  D -->|"Confirms"| E(["Act with more confidence"])
  D -->|"Challenges"| F(["Investigate, your
gut may be wrong"])
People analytics in one loop, start from a decision, let the evidence argue back. Leaders Loop

Engagement is the metric most often connected to hard outcomes, and here the evidence is unusually deep. Gallup's Q12 meta-analysis (11th edition) pools 736 studies across 90 countries and 183,806 business units, and reports that units in the top quartile of engagement out-earn the bottom quartile on profitability by a median of 23%, with markedly lower turnover, a gap of around 51% in low-turnover organisations and 21% in high-turnover ones. Two cautions before you cite that at a town hall: it is correlational, so engagement and performance reinforce each other rather than one simply causing the other; and it is Gallup's own instrument, so weight it as strong, consistent evidence rather than the last word.

The move that follows is to treat engagement as a diagnostic, not a leaderboard. A falling score is a prompt to ask why in the next team conversation, not a number to defend, and certainly not one to chase for its own sake. Which brings us to the trap that swallows naïve metric programmes whole.

Why targets backfire: Goodhart's Law

The single most useful idea in this whole field is a warning. The economist Charles Goodhart observed in 1975 that statistical regularities collapse once you lean on them for control; the anthropologist Marilyn Strathern later sharpened it, in a 1997 paper on accountability in education, into the line everyone now quotes: "When a measure becomes a target, it ceases to be a good measure" (see Goodhart's Law). The mechanism is simple and human. The instant people are rewarded for moving a number, they optimise the number, not the thing the number was standing in for.

When a measure becomes a target, it ceases to be a good measure.

The workforce is full of casualties. Make time-to-fill a recruiter's headline target and roles get filled fast with weaker candidates who churn within the year. Tie managers' bonuses to engagement scores and you teach them to lobby for high marks, not to lead better. Reward low attrition absolutely and a manager will cling to a poor performer rather than show them the door. So the move is to hold metrics in pairs that pull against each other, time-to-fill and quality-of-hire, attrition and regretted attrition (the leavers you wanted to keep), engagement and what people say in exit interviews. A single metric, weaponised, will be gamed. A balanced pair is far harder to fake.

An honest limitation. People analytics inherits a problem the harder sciences mostly don't: small numbers and confounded ones. A team of nine is a sample of nine, and a one-person departure swings a percentage wildly; "performance" and "potential" are partly opinions wearing the costume of data, and any model trained on past promotion decisions will faithfully reproduce yesterday's biases. None of this means don't measure. It means measure with humility, pair every number with the human question behind it, treat small-team figures as conversation-starters rather than verdicts, and never let a model make the call a manager should own. The data is a flashlight, not a judge.

A worked example

Take a 200-person software company, call it Meridian, where the CEO is worried about turnover. (Illustrative figures throughout; this is a teaching example, not real data.) The headline number looks fine: 14% annual attrition, bang on the industry rumour-mill average. The board nods and moves on. The head of people doesn't.

She does what Project Oxygen modelled, she lets the data argue back, and segments the single number instead of averaging into comfort. Splitting attrition by tenure and team, a different picture appears: company-wide 14% hides a brutal 31% among engineers in their first two years, while everyone past year three is rock-steady. The problem isn't "retention." It's early retention, in one function. Now she pairs the metric against its partner. Of those early leavers, exit interviews say the same thing, onboarding was thin and the first manager was absent. This is regretted attrition: the people the company most wanted to keep.

flowchart TD
  A(["Headline: 14% attrition
'looks normal' (illustrative)"]) --> B(["Segment by tenure & team"])
  B --> C(["Finds: 31% among engineers
in their first 2 years"])
  C --> D(["Pair with exit-interview data:
thin onboarding, absent manager"])
  D --> E(["Act on the cause:
fix onboarding & manager support"])
  E --> F(["Re-measure that segment,
not the blended average"])
The same number tells two different stories, the average hides the problem the segment reveals. Leaders Loop

The fix that follows is specific, not a "retention initiative": a structured first-90-days onboarding plan for new engineers, and coaching for the one manager whose hires keep leaving. And, crucially, she re-measures that segment, not the blended company average that looked fine all along. Had Meridian simply set "reduce attrition to 12%" as a target, the easiest path would have been to hire people unlikely to leave for dull reasons, not to fix the broken onboarding. The number would have improved while the actual problem rotted. Goodhart, quietly, in a quarterly report.

Frequently asked questions

Do I need a data team to do people analytics?

No. The bench Google brought to Project Oxygen is the exception, not the entry fee. Most useful people analytics is a careful question, a spreadsheet of records you already keep (joiners, leavers, tenure, hiring dates), and the discipline to segment rather than average. Start there. Buy tooling only once a manual analysis has proven a question is worth answering repeatedly.

Which workforce metrics actually matter?

The ones tied to a decision you're trying to make. As a durable starter set: regretted attrition (not just raw turnover), time-to-fill paired with quality-of-hire, engagement read as a diagnostic, and internal mobility (are people growing here, or only leaving?). Resist the urge to track everything, a dashboard nobody acts on is just expensive decoration.

Isn't reducing people to numbers a bit dehumanising?

It is, if the number becomes the point. The guard against it is to treat every metric as a prompt for a human conversation rather than a substitute for one. A falling engagement score should send a manager into a 1:1 to ask why, not into a spreadsheet to defend the figure. Used well, analytics surfaces the people a busy organisation would otherwise overlook, the quiet flight risk, the manager whose team keeps leaving.

How do I stop my metrics being gamed?

Assume they will be, then design against it. Pair every metric with a counter-metric that moves the opposite way when someone games the first (time-to-fill against quality-of-hire). Keep the highest-stakes measures off individual bonus formulas, where the incentive to game is strongest. And review numbers with the people they describe, so context travels with the figure.

What's the difference between HR metrics and people analytics?

HR metrics are the raw counts, headcount, attrition rate, cost-per-hire, that report what happened. People analytics is what you do with them: connecting them to business outcomes, segmenting them to find the real story, and sometimes modelling them to anticipate what's coming. Metrics are the ingredients; analytics is the cooking.

Related in the Toolkit

People analytics is only as good as the talent system it measures, the data on how you recruit and assess talent feeds time-to-fill and quality-of-hire, while the early-attrition story almost always traces back to onboarding and ramp, the single highest-leverage place a metric tends to point.

Where to go next