People analytics has a split personality. One version helps organizations spot burnout risk before it becomes attrition, find pay inequities before lawyers do, and learn which managers actually develop people. The other version counts keystrokes, scores "engagement" from email metadata, and tells your boss how long your camera was off.
Same tools. Same data, often. The difference is intent, design, and one question most analytics programs never ask out loud.
Data employees do not trust becomes data employees learn to game.
The Trust Test
Here's the filter for every metric you collect: would you be comfortable explaining to employees, in plain language, exactly what you track and how it's used? If the honest answer requires euphemism — "workforce optimization signals" — you already know which side of the line you're on.
This isn't just ethics; it's engineering. Surveillance metrics decay on contact with awareness. The moment people know activity is scored, activity gets performed: mouse jigglers, strategic message timing, meetings attended for the attendance log. You end up with precise measurements of theater.
Principles That Keep You on the Right Side
1. Aggregate by default, individual by exception
Most legitimate questions are population questions: is this department burning out? Are promotion rates equitable? These need team-level patterns, not personal dossiers. Reserve individual-level data for narrow, declared purposes — payroll, compliance, safety — and wall it off from managerial curiosity.
2. Measure outcomes, not activity
Activity metrics (hours online, messages sent) are cheap to collect and nearly meaningless — they measure presence, not contribution, and they punish your most efficient people. If you can't connect a metric to an outcome someone actually cares about, it's surveillance with a dashboard.
3. Give the data back
The strongest trust signal: employees see what you see. If you analyze meeting load, show people their own. If you survey engagement, publish team results and what will change. Analytics done to people breeds resistance; analytics done for them builds the data quality everything else depends on.
4. Put a human between the model and the decision
Any algorithm that influences hiring, pay, or performance needs a named human owner who can explain it, override it, and answer for it. "The model flagged you" is not a sentence anyone should hear in their career.
Start With Questions, Not Data
The failed analytics programs collect everything and hunt for insight; the good ones start with three business questions worth answering — why does one division lose women at twice the rate? what do our best managers do differently? — and collect only what those require. The discipline of the question is also the discipline of the boundary.
Done this way, people analytics becomes something rare: a data practice employees actively want, because it visibly works for them. That endorsement is worth more than any dashboard.