AI Monitoring Cuts Costs But Can Drive Away Your Best Workers

As artificial intelligence makes employee monitoring cheaper and easier to deploy, companies face a counterintuitive problem: the falling cost of surveillance may be encouraging them to use far too much of it.

New research involving German bakeries, airport retailers, and other businesses reveals that monitoring systems impose two distinct types of costs. The first is obvious—hardware, software, staff time, and management attention. The second is hidden but often larger: the behavioral changes monitoring triggers in the people being watched.

When controls backfire

A large German bakery chain discovered this the hard way when investigating rising turnover among skilled bakers and store managers. The culprit turned out to be the company's quality control system, which required workers to complete more than 20 daily operational checklists. Experienced bakers interpreted these checklists as signals that management didn't trust their professional judgment.

Researchers ran a randomized controlled trial, removing two of the most-disliked checklists from half the chain's stores while keeping them in place at the others. The results were striking: sales in stores without the checklists rose 2.7 percent, and attrition among trained workers and managers fell by more than 20 percent. Quality didn't suffer—the problems the checklists were designed to catch never materialized. Customer reviews even improved, praising faster service.

But the story has a twist. Among less experienced staff, removing the checklists increased attrition by 20 percent. Newer workers benefited from the structure the checklists provided. The same control that insulted a master baker served as a lifeline for a novice.

The airport that chose not to monitor

Linus Arauz, who operates retail locations across multiple airport programs, faced a different version of the same question. Some of his locations are unstaffed—travelers take items and are expected to pay without supervision. The obvious solution would be comprehensive monitoring through cameras and sensors.

After analyzing the numbers, Arauz decided against it. The cost of time, overhead, and logistics required to count inventory across stores significantly exceeded the actual losses from theft. Control would cost more than it would save.

Why it matters

As AI drives down the direct cost of monitoring, companies risk falling into what researchers call "the control trap"—deploying surveillance simply because it's cheap and available, without calculating whether it actually delivers value. When monitoring costs almost nothing, managers stop asking whether watching pays. But cheap monitoring doesn't eliminate the hidden costs; it only makes them easier to ignore.

The trap has already caught major tech companies. Amazon shut down an internal token leaderboard after employees gamed the system to climb rankings, driving up computing costs without increasing productivity. Meta suspended its Model Capability Initiative, which logged keystrokes and screenshots, after internal data became more widely accessible than intended.

Matching controls to workers

The bakery research points toward a more nuanced approach: adapt monitoring intensity to worker qualifications. Dutch teaching hospitals have practiced this since 2006, allowing medical residents to perform tasks without supervision once they demonstrate mastery, with oversight falling away one documented skill at a time.

Companies should audit existing monitoring systems to identify legacy controls that no longer justify their costs. Every piece of surveillance should earn its keep by delivering measurable value that exceeds both direct and behavioral costs. And critically, falling costs aren't a reason to monitor more—they're a reason to choose more carefully.

The bakery didn't eliminate all its checklists after the experiment. It kept one. That kind of disciplined thinking about control, rather than reflexive deployment of cheap AI monitoring, will separate thriving organizations from those trapped by their own surveillance systems.

These findings were first reported by Guido Friebel, a professor of human resources at Goethe University in Frankfurt, in Harvard Business Review.