What people analytics can't see — and why that's a feature

The default design instinct for any People analytics product is to maximize visibility. More dashboards, more drill-downs, more cross-references between calendar metadata, email volumes, Slack patterns, performance ratings, and survey scores. The pitch deck shows a heat map of which teams are at risk and which managers are coaching effectively. Procurement leans in.

We took the opposite design move. The pitch is what we deliberately decided not to see.

The two kinds of visibility a People function actually needs

Strip out the dashboard mockups and the work of People analytics reduces to two distinct questions. Is the program working at the organizational level? And is there a specific person I need to talk to?

These look like questions of the same kind. They are not. The first is an aggregate question that needs an aggregate signal. The second is a relational question that needs a relationship. When we conflate them, we end up building tools that produce a worse answer to both.

A heat map by team is an attempt to answer question one with a denser version of the same signal. A drill-down to individual scores is an attempt to answer question two with analytics instead of a conversation. Both moves degrade what they were trying to solve. The aggregate signal degrades because employees notice the drill-down exists and start producing performative inputs. The individual question degrades because by the time it appears on a dashboard, the actual conversation is months late and now has to begin with "the system flagged you," which is the worst possible opening.

What we chose not to instrument

We made three structural decisions early that govern what an ORG_ADMIN account on Nela can see.

First, no individual content visibility. Reflections, wins, challenges, 1:1 agendas, outcomes, and AI suggestions are owner-only at the database — enforced by Postgres Row-Level Security, not just hidden in the UI. There is no admin-level read-through, no service-role bypass, no export of any individual employee's content. This is covered by integration tests that fail closed if any code path tries to bypass it.

Second, no sentiment analysis. We do not run NLP over reflection text to score mood, surface keywords, or generate manager-facing summaries. Even if we did, the employee would know we did, the act of recording would shift to performative writing, and the signal we extracted would be the artifact of our extraction. The cleanest solution is to not build the extractor.

Third, an aggregation threshold. Team-level engagement counts — seat activity, reflection cadence, open-loop closure rate — are visible to ORG_ADMIN only when a team has five or more active seats. Below that threshold, counts return null at the database layer. This is not a UI nicety. It is a structural design choice to prevent the inference move where "12 of 15 logged a reflection" silently becomes "we know who the three are."

These three decisions look, in product terms, like deliberate visibility loss. In design terms, they are how we make the visibility we do have worth reading.

What's left to see, and why it's enough

What ORG_ADMIN does see is a small, durable set of behavioral signals at the aggregate level.

Seat activity tells you whether the program is being used at all. If 80% of seats logged at least one capture this week, you have an actively-adopted tool. If 30% did, you have a procurement decision and a non-adoption problem.

Reflection cadence tells you whether the depth-of-engagement habit is sticking. Active capture is a low bar. Weekly reflection — the habit that actually produces the structured-reflection effect documented in the field — is a higher bar. Teams clustered at the weekly-reflection cadence are getting the mechanism Di Stefano and colleagues identified (an 18% performance improvement, in their field experiment). Teams that capture but never reflect are not.

Open-loop closure rate tells you whether the things employees raise are getting resolved. An open loop is something the employee flagged in a 1:1 agenda that has not yet been closed with an outcome. A team with a healthy closure rate has functional 1:1 conversations. A team where loops pile up has a manager problem you can address — without ever reading the loops themselves.

Last-active-at tells you who is still using the seat. Aggregate at the team level. We surface individual last_active_at only because seat administration requires knowing which seats are dormant for deactivation purposes. It is metadata about presence, not content.

That is the entire visibility model. Four aggregate signals plus seat metadata.

What you give up, in honest terms

You give up the ability to read what any specific employee said. You give up sentiment analysis across the organization. You give up the engagement scoring model that ranks teams against each other. You give up the executive dashboard that lets the CHRO drill from organization to function to team to individual.

In exchange you get a signal that does not degrade over time, because the underlying behavior it measures is honest. The thing that makes the signal honest is exactly the thing you gave up — the read access that would have made the employees produce performative inputs.

This is the trade. It is a real trade. It is not a free lunch. The argument for it is that the visibility you gave up was costing you more than the visibility it provided, because the act of measuring at that depth was degrading the underlying state. The remaining signal is smaller and more reliable.

Mechanism and product

The economics literature is clear that intrusive monitoring erodes intrinsic motivation (Frey, Falk and Kosfeld, Ravid et al.). The autonomy lens from Self-Determination Theory says the same thing from the motivation side. The aggregate-only design we built is one of the few People-analytics products that has internalized this constraint structurally instead of paying lip service to it.

Whether the resulting signal is actionable enough for your function is what the pilot measures. We are not selling you a richer dashboard. We are selling you a leaner one that you can trust. The bet is well-founded; the test is yours to run.

How Nela Helps

Use Nela to log your wins, track your challenges, and build a private 1:1 agenda from your own evidence for your next conversation. Your data is owner-only at the database — enforced by Postgres Row-Level Security, not just hidden in the UI — and only you can read it back through the app. Request pilot access.

Further reading

• Frey, B. S. (1993). "Does Monitoring Increase Work Effort?" Economic Inquiry.
• Ravid, D. M., Tomczak, D. L., White, J. C., & Behrend, T. S. (2020). "EPM 20/20." Journal of Management.
• Di Stefano, G., Gino, F., Pisano, G. P., & Staats, B. R. (2014). "Learning by Thinking." Harvard Business School Working Paper 14-093.
• Deci, E. L., & Ryan, R. M. (2000). "Self-Determination Theory." American Psychologist.