Can employee monitoring data predict turnover?

To a degree, yes — working-pattern changes do precede departures often enough to be statistically real. The difficulty is that it is probabilistic and the base rate is low, so most people a model flags are not leaving, and acting on individual flags produces more harm than the prediction is worth.

What are the behavioural signs someone is about to resign?

Narrowing scope to assigned work, declining participation in optional and forward-looking activity, reduced collaboration breadth, disengagement from planning for future periods, and in some cases a sharp drop in after-hours work following a long period of excess. None is individually conclusive.

Is it legal to use monitoring data to predict resignations?

It depends on jurisdiction and purpose. If you collected activity data for capacity planning, repurposing it for attrition modelling is a new purpose requiring its own lawful basis and notice in the EU, UK, Canada and India. Automated inference about individuals also engages additional obligations under GDPR.

Should managers be told who is a flight risk?

Generally no. Telling a manager an individual is a flight risk changes how that person is treated — often reducing their opportunities, which makes departure more likely rather than less. The prediction becomes self-fulfilling, and the employee never learns why their work changed.

What is the defensible way to use attrition data?

At aggregate level, to find the conditions that produce turnover rather than the people experiencing it. Which teams, roles, managers, workload patterns or tenure bands lose people, and what the leading indicators are across a population. That drives structural fixes and does not require singling anyone out.

Can You Predict Who's About to Quit?

Attrition prediction mostly works, which is the problem. What the signals are, why acting on individuals backfires, and the aggregate use that's defensible.

Attrition prediction is one of the few people-analytics claims that broadly holds up. Working patterns do change before people resign, and with enough history a model will beat chance at identifying who.

This is precisely why it needs handling carefully, and why the interesting question is not whether it works but what you are allowed to do with an answer.

The signals that actually precede departure

They are more subtle than the folklore suggests. The classic tells — updating LinkedIn, dressing differently, taking odd afternoons off — are not available in activity data and were never reliable anyway.

What shows up in working patterns is a narrowing of engagement with the future.

Scope contracts to assigned work. The person who used to pick up adjacent problems, comment on other teams' designs or fix things nobody asked them to fix stops doing that. The committed work still gets done, which is why output metrics often show nothing.

Participation in forward-looking activity declines. Planning sessions for next quarter, hiring loops, roadmap discussions, anything whose payoff lands after a notice period. This is the most consistent signal and the most intuitive once stated — people disengage from futures they do not expect to be part of.

Collaboration breadth narrows. Fewer distinct people and systems touched, more time in a smaller set of tools.

And occasionally, an abrupt reduction in after-hours work after a sustained period of excess. Read naively this looks like improvement. It frequently means someone has stopped trying to fix a situation and has decided to leave instead. The burnout signals piece covers the preceding phase.

None of these is conclusive alone. All of them are also consistent with a person who has sensibly set better boundaries, or whose role changed, or who is dealing with something at home.

Why the base rate ruins individual prediction

Here is the arithmetic that most vendor demos omit.

Suppose annual voluntary attrition is 15%, so in any quarter roughly 4% of your people leave. Suppose your model is good — it correctly flags 70% of leavers and wrongly flags only 10% of stayers. Those are strong numbers for this kind of problem.

In a thousand-person company, that quarter: about 40 leavers, of whom you catch 28. And about 960 stayers, of whom you wrongly flag 96.

So of 124 flagged people, 28 are actually leaving. Roughly four in five people your good model flags are not going anywhere. You now hold a list where the majority are false positives, and every action you take on that list is mostly being taken on people who were staying.

This is not a modelling problem to be solved with a better algorithm. It is what low base rates do, and it is why individual-level attrition flags are a bad instrument almost regardless of model quality.

The compliance position

Repurposing is the issue. If you deployed activity monitoring for capacity planning and then build an attrition model on it, you have a new purpose, and in most modern privacy regimes that requires its own lawful basis and its own notice rather than inheriting the old ones. Our UK/ICO and GDPR guides cover purpose limitation, and the DPDP and Canadian positions land in the same place.

Automated inference about individuals raises the bar further in the EU and UK, particularly where it influences decisions affecting them — and reduced access to opportunities is such a decision, whether or not anyone wrote it down as one.

The practical consequence: aggregate attrition analysis is comparatively straightforward to justify. Individual flight-risk scoring is not, and in several jurisdictions it would be hard to defend at all for a purpose as soft as retention management.

The version that is both useful and defensible

Flip the unit of analysis from the person to the condition.

Instead of asking who is likely to leave, ask what is producing departures. Which teams lose people disproportionately. Which managers. Which tenure bands — the six-month cliff and the two-year cliff have different causes. Which workload patterns precede resignations across a population. Whether the people leaving are the ones carrying the most concentrated load, which is a capacity planning finding rather than an HR one.

This is more useful anyway, because it produces interventions you can actually make. "Engineers on the platform team resign at twice the rate of the rest of engineering, and platform engineers carry the highest after-hours load in the company" is an actionable sentence. "Priya has a 0.62 flight risk score" is not — there is no honest action that follows from it that you would be willing to explain to her.

And the aggregate version needs no individual flags, so it sidesteps both the base-rate problem and most of the compliance exposure.

If you want to do something about an individual

Talk to them. This is not a quip.

The information a manager needs about whether someone is thinking of leaving is available through a conversation that is already supposed to be happening, and unlike a model, a conversation can address the cause. It also respects the person enough to let them answer.

The honest version of that conversation is not "are you planning to leave" — nobody answers that truthfully. It is closer to asking what they want to be doing in a year, and whether the current path gets them there. Teams where that question is routine do not need attrition models, and teams where it would be strange to ask have a problem no model will fix.

Where ProdView fits

ProdView produces the aggregate working-pattern data this analysis needs: focus time, meeting load, after-hours and weekend activity, workload distribution across teams, read as team-level trends. That supports the condition-level question — which teams and which load patterns correlate with turnover.

When not to pick us: you want per-person flight-risk scoring. We do not provide it and will not build it. The base-rate arithmetic makes it a poor instrument, the self-fulfilling dynamic makes it harmful, and the compliance position in most of our customers' jurisdictions makes it difficult to defend. If a vendor is offering you individual attrition scores, ask them what action they expect a manager to take on a flag that is wrong four times out of five.

Try before you commit

Three seats are free forever. If you suspect load concentration is driving your attrition, the distribution data will show you where — and that is a fixable finding. The ROI calculator will put a number on the replacement cost you are avoiding.

General information as of October 2026, not legal advice. Automated inference about employees carries jurisdiction-specific obligations; take advice before deploying any model that influences decisions about individuals.