Opening

Reader, on the homepage of Insightful, a company that sells employee-monitoring software, there’s a customer testimonial worth pausing on. When they first rolled it out, productivity sat in the high 70% range. Now it’s up to 92%, and employees actually come ask what their score was that day. The testimonial even says work has become like a game.

But in a Wall Street Journal interview, the same company’s CEO said the range companies actually aim for is 60-80% of the workday. And the advice that article gave to office workers was this: if you’re going to artificially inflate your activity rate, be careful — hitting 90% gets you suspected of gaming the system.

Three numbers, and they don’t line up. A bragging-rights 92%. A “normal” range of 60-80%. A suspicion threshold at 90%. Let me give you the conclusion up front. The measurement system is reading its own theater back to itself as performance.

---

Is 92% Achievement, or Is It Acting?

First, we need to look at what this number actually counts. Activity rate¹ is the share of work hours spent on activity classified as “work.” The entire trick is in that phrase — “classified as work.” It’s not measuring what you actually produced. It’s measuring whether your cursor sat inside apps and sites your manager pre-designated as legitimate.

Which means there are only two ways to raise this number: do more work, or look like you’re working.

Interestingly, every single tip the Wall Street Journal compiled falls into the second category. The first piece of advice: fill your calendar meticulously. When you make a call or step into an in-person meeting, your messenger status flips to “away” — and the monitoring tool cross-checks whether there’s a calendar entry for that time slot to decide if the absence is legitimate. If your calendar is empty, it reads as slacking off. There’s also advice to use a physical mouse jiggler² rather than a software one, because security systems have started blocking the software kind.

Notice what all these tricks have in common? Not one of them is a way to do your job better. Every single one is a way to be read well by the machine.

Worth noting how big this market has already gotten. In its own materials, Insightful claims nearly 80% of major companies have adopted employee monitoring, and that more than 5,100 teams use its product. The product copy is telling, too: it promises full visibility without surveillance or keystroke logging, while in the same breath explaining that it tracks app and website usage, idle time, and even AI adoption — including unauthorized AI use. The word “surveillance” is the only thing missing; the list of things being counted just keeps growing.

This game, in fact, is old. It’s usually traced back to the “boss button,” built in the early 1980s by a developer named Roger Wagner — one keystroke, and the screen would switch to a spreadsheet the moment a boss approached. Wagner himself says it was a joke — not meant to shield lazy employees, but a jab at overly intrusive managers. 40 years later, that joke has become a survival skill.

What Meta Was Actually Counting Was Tokens

Up to this point, it’s just a guessing game. The problem is that this number started being used to make employment decisions.

On July 13, 26 current and former employees filed a lawsuit against Meta in federal court in Oakland. The backdrop is the roughly 8,000-person layoff—about 10% of the workforce—announced in May. The complaint describes the company deploying a “constellation of internal AI systems,” and claims that keystroke and activity monitoring data, algorithm-assisted performance rankings, and AI token³ usage dashboards were used to select targets. Meta’s position is that humans make the final call.

The sentence in the complaint I lingered on longest was a different one. It states that these scores and ratings are designed such that employees on protected leave, or those whose output dropped due to disability, cannot accumulate them. All 26 plaintiffs had taken medical, parental, or family leave, or had requested disability accommodations. The temporary restraining order request was denied on July 17, and a preliminary injunction hearing was held yesterday, August 24.

What matters in this case isn’t who wins or loses. It’s the fact that the company used AI usage as a metric when deciding who to cut.

The backdrop makes this clearer. Reuters reported in April 2026 that Meta had been logging employees’ keystrokes, mouse movements, and clicks for AI training purposes. The program reportedly ran afoul of European privacy regulations, and was reportedly halted after it emerged that other employees could access the conversations and records the system had collected. The layoff announcement came the following month. It’s a timeline that shows just how fast data collected for observation can flow toward other uses.

And data from workers themselves reveals the nature of this metric. In a survey of 1,000 U.S. workers by Visier, cited by the Wall Street Journal, 48% said they had inflated their own AI usage. The researchers’ interpretation was that this is behavior driven by the fear of appearing to fall behind within the organization.

Let’s lay these two facts over each other. The company counts AI usage, half the employees inflate that number, the company ranks people by the inflated numbers, and that ranking becomes a list. Anyone without the bandwidth to inflate gets pushed down. The clearest example: someone on maternity leave who simply never logged in at all.

There’s one more twist here. Lately, companies have started scrutinizing token costs, so using a lot can also read as waste. The zone of the “right answer” is narrow, and it keeps moving. In last July’s issue, I called the invisible time spent verifying AI output a “verification tax.” This time, what’s being billed isn’t verification—it’s time spent performing.

South Korea Already Has a Law on the Books

In the US, depending on the state, companies aren’t even obligated to disclose that monitoring is taking place. South Korea’s situation is a bit different.

Article 37-2 of South Korea’s Personal Information Protection Act (PIPA) sets out data subjects’ rights regarding automated decisions. The provision has four branches. If a fully automated system’s decision has a material impact on a person’s rights or obligations, that person can refuse the decision (Clause 1); they can demand an explanation of the decision (Clause 2); if such a demand is made, the business must take measures like human-involved reprocessing (Clause 3); and the criteria and procedures behind automated decisions must be disclosed in an easily accessible way (Clause 4).

Here’s the catch. The right to refuse under Clause 1 comes with a proviso. If consent was given, if it’s a statutory obligation, or if it’s necessary to fulfill a contract, refusal isn’t possible. The logic of “fulfilling an employment contract” leaves plenty of room to fall under this exception. Kwon Seok-hyun, a lawyer at Minbyun (Lawyers for a Democratic Society, a Korean civic legal advocacy group), points out that the blind spot in South Korea’s current personal information legal framework is that it treats “consent” as a master key that justifies any form of surveillance. It fails, he says, to account at all for the power imbalance between employer and employee.

But if you read the provision to the end, there’s something left standing. The right to demand an explanation under Clause 2 and the disclosure obligation under Clause 4 carry no such proviso. Refusal may be blocked, but asking what criteria you were scored against — and forcing the company to disclose those criteria — is a separate matter entirely. As far as I can tell, these two clauses are barely used in South Korea today.

Put into question form, it looks like this: which items in your work data feed into your evaluation, how are those items weighted, and does that calculation pause during leave or sick days? If a company is using automated judgment, these three things fall within the scope of what must be disclosed.

The situation on the ground isn’t quiet, either. A report on electronic workplace surveillance released in October 2025 by Jikjang Gapjil 119 (Workplace Gapjil 119, a South Korean labor rights watchdog group) found that commercial solutions capturing PC screens second-by-second and replaying them like video are already on the market. The report also catalogs what’s actually being collected at South Korean workplaces: internet usage logs during work hours, messenger and email records, CCTV footage, GPS-based location data, personal social media activity, and even detection logs of PC power state and mouse/keyboard activity. These items overlap almost exactly with what’s at issue in this lawsuit. The report includes cases where an employee who refused a recommended resignation had their movements confronted with CCTV footage, and where an employee who filed a complaint had personal messenger conversations recovered from their work PC and used against them. A note: this piece is not legal advice. If you’re facing an actual dispute, consult a professional first.

Oswarld’s Lens

I’ve designed dashboard KPIs many times while building GTM strategies, and metrics always age in the same order. At first, they’re for observation. You build them to see what’s happening. Next, they become reporting tools. They end up on slides shown to executives. Finally, they become evaluation tools.

A number that was accurate when it was for observation invariably improves the moment it becomes an evaluation metric. But reality stays the same. This is exactly what Goodhart’s Law⁴ describes.

That’s why I can’t read a figure like 92% as a success story. If that number shows up, I don’t see something to celebrate — I see a signal to start an audit. The vendor itself said the normal range is 60–80%.

AI usage, in particular, is about as bad a metric as they come. It’s an input, not an outcome; the cost of gaming it is essentially zero; and now there’s even a countervailing signal that using it too much is wasteful. Tying a metric like this to people’s employment is a design failure.

I’m not against monitoring tools themselves. They’re useful for spotting bottlenecks. But between measurement and evaluation, there must always be a layer of human judgment.

Closing

To sum up:

First, utilization rate and AI usage aren’t outcomes — they’re traces of behavior. The moment a trace gets tied to evaluation, the skill of leaving traces starts to substitute for actual skill. Second, what the Meta lawsuit revealed is that these traces are already being used in employment decisions, and that people who structurally cannot accumulate them are put at a disadvantage. Third, Korean workers already have a channel to demand explanation and disclosure of criteria. It’s just not being used.

If you’re a manager, I’d suggest opening your dashboard today and asking just one question. Is the easiest way to move this number up doing good work — or looking like you’re doing good work?

What number are you being evaluated on at your company right now? If you’ve ever done something that had nothing to do with actual performance just to protect that number, tell me in the comments what it was. Once enough examples come in, I’ll organize them by metric in a future issue.

---

💬 Tell me in the comments which number is running your day · 📨 If you have a colleague staring at a dashboard, pass this along

---

Found this useful? Share it

Copy linkXLinkedInThreadsKakaoTalk네이버 밴드

References & Further Reading

Primary sources

• Cordilia James, “How to Outsmart AI When It’s Tracking Your Workday”, The Wall Street Journal, 2026. ··· This is where this issue started. The strength of this piece is that it reports on both the makers of surveillance tools and the people who work around them.
• Sanders et al. v. Meta Platforms, Inc., U.S. District Court for the Northern District of California (Oakland), filed July 13, 2026. ··· The key part of this 71-page complaint is where it addresses the non-cumulative nature of AI scores. Regardless of the outcome, it’s a well-organized document on the structural blind spots of automated evaluation.
• Personal Information Protection Act, Article 37-2 (Data Subject’s Rights Regarding Automated Decisions), Korea Law Information Center. ··· I’d recommend checking the proviso in Paragraph 1 against the differences in Paragraphs 2 and 4 yourself. Reading the clause takes about 3 minutes.

Background

• Jikjang Gapjil 119 (Workplace Gapjil 119, a South Korean labor rights watchdog group), Policy Report on the State of Electronic Workplace Surveillance and Legal Reform, October 12, 2025. ··· This one is full of concrete cases from Korean workplaces. Reading only foreign coverage, it’s easy to feel this is someone else’s problem — this report makes clear it isn’t.
• Discussion of Charles Goodhart’s indicator theory, in Marilyn Strathern, “Improving Ratings: Audit in the British University System”, European Review, 1997. ··· The line we commonly cite as “Goodhart’s Law” was actually coined in this paper.

Related past issues worth reading

• What the Luddites Actually Smashed Wasn’t Machines ··· This issue covered the concept of a “verification tax.” It connects to this issue’s discussion of “time spent performing.”
• Thousands of IBM Consultants Are Taking a Test ··· This covered how AI usage becomes a signal of qualification within an organization. This issue is about what happens when that signal turns into a performance metric.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

각주

1. Utilization rate: The share of work hours classified as work-related activity. It measures where time went, not what was produced — a different concept from output. ↩

2. Mouse jiggler: Software or a USB device that automatically moves the mouse cursor to make it look like someone is at their desk. It’s used to prevent a status from switching to “away.” ↩

3. Token: The smallest unit an AI model uses to process text. Since it’s the basis for measuring usage and cost, the number of tokens consumed is treated as a proxy for how much AI was used. ↩

4. Goodhart’s Law: The principle that once a measure becomes a target, it ceases to be a good measure. It originated from British economist Charles Goodhart’s commentary on monetary policy. ↩