When we tell an operations team that an agent will observe their work for a few weeks to learn how the process actually runs, somebody in the room stiffens, and that reaction is not a soft-skills problem to be managed around. It is a correct read of a real body of evidence. Workplace monitoring has been studied for decades, and the finding that keeps replicating is not that monitoring is uniformly bad but that it produces the opposite of what the people who buy it want, under specific conditions that are worth naming precisely. An observation engagement that ignores those conditions earns the resistance it gets, while one designed against them is supported by the same evidence. The difference is in how the data is used, not in how the deployment is described.
What the research actually found
The clearest recent statement comes from a research program led by Chase Thiel. Thiel, Julena Bonner, John Bush, David Welsh, and Niharika Garud reported in Harvard Business Review in 2022 that monitored employees were substantially more likely to take unapproved breaks, disregard instructions, and work at a deliberately slow pace, and that when a second, experimental study gave people an opportunity to cheat, the ones told they were under surveillance took it more often. Their explanation is mechanical rather than moral. Monitoring shifts the felt locus of responsibility, so when a person believes someone else is watching and accountable for catching problems, their own sense of ownership over rule-following weakens, and the behavior the monitoring was meant to suppress goes up instead of down. Thiel returned to the question in 2024 with Shawn McClean, Jaron Harvey, and Nick Prince, and the follow-up located the damage more precisely: when information from monitoring was used for control purposes such as performance review, employees were more likely to engage in counterproductive behavior, time theft, inattentiveness, cyberloafing, and tardiness among them, while the same information used for feedback left employees trusting their supervisors and performing better in their jobs.
This is not a fringe result. A 2024 survey of 633 workers in Cyberpsychology found that privacy concerns about electronic performance monitoring produced a sizable reactance effect, one that lowered both attitudes toward the monitoring and job satisfaction, and its authors connect that reactance to resistance behaviors like cyberloafing and quiet quitting. The same survey found that trust in the employer was more strongly correlated with positive attitudes toward the monitoring than reactance was with negative ones, which places the outcome in the relationship the tool lands in rather than in the tool itself. Buyers will not learn any of this from the people selling the trackers: a 2024 content analysis in Human Behavior and Emerging Technologies examined 15 vendors of employee monitoring software and found their webpages under-represent the risks of the tools relative to the benefits. The survey and the content analysis measure different things, reactance in one case and vendor marketing in the other, and neither cuts against Thiel's result that the damage follows what the data is used for rather than the capture itself, which is the variable a buyer gets to choose.
The variable that changes the result
Read those papers closely and the damaging cases share a structure. The data is tied to identified individuals, it feeds performance review or discipline, it runs indefinitely, and the person being measured is given no account of what is collected or why. Each of those is a choice, and each one is what a productivity tracker, by design, does, since the tracker exists to score the individual, which is precisely the configuration the research blames.
A diagnostic observation built to redesign a process inverts every one of those variables. It runs for a bounded window rather than as a permanent fixture, and it aggregates at the level of the workflow, asking how a month-end close or an accounts-payable queue moves through the organization, not how fast a named person clicks. Its output is a redesign, fed forward into the work, rather than a review fed back onto a record. None of this is a softer phrasing of the same activity. It is a different activity that happens to record at the same desk. The mechanism Thiel's team describes, responsibility shifting to the watcher, only fires when the watching is positioned as judgment of the person, and process discovery positioned as the opposite does not meet that condition.
Architecture is half the answer, behavior is the other half
How we build the capture matters, and we have written separately about what a capture agent records and what it throws away, because aggregation and redaction are real engineering rather than reassuring copy. Architecture alone, though, does not earn trust. A system that technically cannot score individuals still reads as surveillance if it arrives unannounced and unexplained, because the felt experience of being watched, rather than the schema behind it, is what the research measures. The behavioral commitments have to be visible to the people whose work is observed, which means they belong in the change-management sequence and not only in the system design.
The sequence we run before anything is captured
We name the purpose before deployment, in plain terms and to the team directly rather than over their heads: the agent is here to learn how the process runs so it can be rebuilt, and that is the entire reason it exists. We show people what is and is not captured, concretely, because abstract assurance is what surveillance also offers and specificity is what distinguishes a diagnostic from it. We exclude individual performance scoring contractually, so that the commitment survives a change of sponsor or a quarter where someone is tempted to repurpose the data, which is the moment trust would otherwise break. And we report findings back to the people whose work was observed, so the observation visibly feeds redesign that changes their day rather than disappearing into a review they never see. That last step addresses what the research says matters, returning ownership to the people who hold it rather than leaving it with the watcher.
A reasonable counter, and why it doesn't hold
The honest objection is that this is a distinction without a difference, that workers cannot verify our intentions and will read any observation as surveillance regardless of how it is framed, so the framing is theater. There is something to it, because trust is earned through behavior over time and not granted on the strength of a promise made at kickoff. But the framing is exactly what lets workers read the behavior while it accrues. The research does not say workers resist all observation; it says they resist observation whose data is used to control and discipline them, and that distinction is one they make accurately when they are given the information to make it. The contractual exclusion of performance scoring is verifiable, the bounded window ends on a date, and the findings come back to the room they came from, which is more than a productivity tracker ever offers because the tracker's whole point is the thing we have removed. Respecting the concern means meeting its specific terms, which is why we treat it as the design problem it is rather than the morale problem it is too often mistaken for. The same care is what separates a serious discovery phase from a faster one, and it is bounded on the other side by the monitoring laws that govern any observation agent, which set the floor that the behavioral commitments build on top of.