Coaching and Surveillance Come in the Same Tool
Visible Work: Making hidden workplace behaviors visible can improve coaching, but only when people can act on findings.
Design Tests: Recording and analytics become defensible when recipients, employee control, and data retention are decided before deployment.
JPMorgan Pilot: JPMorgan’s monitoring report exposes hours analysts cannot control, potentially redirecting pressure toward inaccurate timesheets.
Data Risk: Employee monitoring creates lasting security and reuse risks when companies collect more information than necessary.
Better Support: AI can strengthen workplace guidance when leaders use employee questions to identify weak support, not punish individuals.
A manager at Cognite sat down with a readout of his own hiring interviews and found that 90% of his questions had been closed-ended. Another learned he had talked for 85% of a conversation he was supposed to be spending on listening.
Neither of them became aware of this as part of a disciplinary process or critique. They were being shown something they had no way of seeing on their own, and according to Brad Wilkins, the vice president of people who runs the coaching scripts that produce those readouts, most of them fixed it.
"Most of them want to be better, they just never had the mirror," Wilkins told us.
Create a Free Account to Keep Reading—and Keep Leading Smarter
Unlock this piece and join a community of forward-thinking leaders discovering tools, playbooks, and insights for thriving in the age of AI.
Have an account? Log In
That is a good outcome, and it arrived through a process that captured every word two people said to each other in a private conversation. The recording happened first, the coaching came after, and the sequence matters more than it looks like it does.
Wilkins is describing something real. He talks about AI making formerly invisible work visible, be it the calendar that has drifted away from what a manager said their priorities were, the onboarding that is going sideways in week three, or the high performer who is beginning to slide.
Those things were always happening. They were just happening below the resolution of anyone's attention, and now they aren't.
Measurement Always Comes With a Workaround
Lynn Parramore, a researcher who studies workplace measurement, traces the lineage back to Samuel Bentham, who designed a factory in which a single overseer could observe every worker from a central point. His brother Jeremy borrowed the design for a prison and gave it the name that stuck. The point of the panopticon was never the watching itself, it was that workers would internalize the possibility of being watched and regulate themselves.
Frederick Taylor brought a stopwatch onto the shop floor a century later and broke labor into timed components. It worked, Parramore says, "up to a point." The point where it stopped working was the moment workers found the seam. They reset the clock when he left the room.
A hundred years on, remote workers jiggle a mouse to fake activity, or prop a pineapple in front of a laptop camera to pass as a person at a desk.
People just have an innate resistance to invasive, intrusive monitoring, and that’s not going to change. That’s just human nature.
That pattern holds today. Measurement comes with something the organization genuinely wants, whether that's fewer factory floor accidents or a manager who finally learns how he sounds in a room, and the resistance it produces gets treated as a compliance problem.
Parramore points out that the answer to that question is often decided by geography rather than by principle. In much of the EU, works councils have consultation rights over monitoring systems before they go in, so the conversation about what else the data could do happens in advance and involves the people being measured.
The American default is closer to a property question. The company owns the laptop, so the company can see what happens on it.
Three Questions Decide What the Capability Becomes
Recording, transcription and behavioral analytics are one capability. What that capability turns into is settled by the answers to three design questions, and my argument here is that those decisions do all of the work that intent is usually credited with.
- Who sees the output?
- Is the person being measured able act on what they see?
- Does the record persist past the moment it was used?
Run Greg Russell's operation at Cover Genius through those. Russell, the vice president of talent, records every phone screen and panel interview with consent, and AI transcribes and summarizes them.
He uses the output to coach interviewers on question quality, on talk-time ratio, on whether they are actually selling the company to a candidate who has three other offers. He calls it unprecedented visibility into interviewer performance, and for most talent acquisition teams it is genuinely a new capability rather than an upgrade to an old one.
He also notes something easy to miss, which is that his interviewers are more present in the conversation because they aren't scribbling notes while someone talks.
The output goes to the interviewer and to Russell. The interviewer can change how they interview next week and the guardrail Russell built is the one that matters most, which is that humans write their own assessment and their own score, and the AI never auto-rejects a candidate.
Russell's Leadership and Values interview carries veto power, and he reads the transcript on any "no." He was mildly surprised he had to say the part about no auto-rejection out loud.
I don't think he should have been surprised. The reason it needs saying is that nothing in the technology enforces it. It holds because he decided it would.
Wilkins passes the same test, and his three-tier taxonomy of where AI leads, where it assists, and where it gets no vote at all is the same instinct expressed as structure rather than as a rule.
Consent is doing real work in Russell's case. Candidates and interviewers know the recording is happening. That is not a small feature of the design, it is the thing that makes the rest of it defensible, and it is the first thing that tends to get dropped when a tool scales from one team to a whole company.
JPMorgan Passes and Fails
In March, the Financial Times reported that JPMorgan had begun a pilot comparing the hours junior bankers report on their time sheets against activity recorded by the bank's IT systems, drawn from video calls, desktop keystrokes and scheduled meetings.
Bankers get a weekly report showing the two numbers side by side. The bank said the tool is "about awareness" rather than enforcement and that it won't be used in performance evaluation, and it compared the reports to the weekly screen time summary on a phone.
The framing is wellbeing, because the underlying problem is real. Junior banking has run on 100-hour weeks for decades, and after the deaths of two young bankers in 2024 and 2025 drew industry-wide scrutiny, JPMorgan created a role overseeing junior banker wellbeing and capped the workweek at 80 hours.
This is the good-faith version. Nobody is trying to catch anyone. The cap didn't hold on its own and visibility is what's left to reach for, which is the same instinct behind showing a manager his talk-time ratio.
Run the pilot through the three questions. The banker gets the report, which is more than most monitored workers get.
The second one is where it comes apart. A manager who learns he asked closed-ended questions can ask open ones tomorrow. A first-year analyst who learns she worked 91 hours cannot decide to work 60, because the hours are set by deal flow and by the managing director who sends the comments at 11 p.m.
The report tells her something she already knew about a condition she does not control, and the only variable she can actually move is what she writes on the timesheet.
Reporting on the pilot noted that juniors had previously been encouraged to underreport hours to stay inside internal caps, which means the measurement lands on the one behavior the employee has any power over, and it lands there whether or not anyone intended it to.
The third test is the one that needs some attention. That data now exists inside the bank's systems, generated continuously, and the commitment that it will not be used in evaluation is a policy commitment rather than an architectural one. Policies are revised by people who did not write them.
Intent Does Not Survive Contact With Infrastructure
So what happens to collected data when nobody has decided how long it lives.
WorkComposer is an employee monitoring tool used by more than 200,000 people that logs activity and captures periodic screenshots.
In February 2025, researchers at Cybernews found more than 21 million of those screenshots sitting in an improperly configured Amazon S3 bucket, open to anyone who knew where to look. The images included emails, internal chats, confidential business documents and login pages displaying usernames, passwords and API keys.
Researchers notified the company the day after they found it, but the bucket was not closed until April 1.
Six weeks. Nobody at any of those companies decided to publish their employees' workdays. They decided to measure them, and the publishing was done by a storage configuration.
That is the part of this that no statement of intent can reach. Whatever you collect, you are now responsible for keeping, and you are responsible for it under conditions you don't fully control, through vendor changes and acquisitions and staff turnover and every incentive that will ever exist to find a second use for data you already have.
Parramore raises a case that shows what the second use looks like. Meta employees, she notes, have found themselves monitored in ways that feed the training of systems they suspect are being built to replace them. The collection outlived its stated purpose, and the people generating the data were the last to be consulted about the new one.
When the system produces an output the subject cannot contest, the subject does not stop having judgment. The judgment just goes underground, and the organization loses the information along with the trust.
This is the word I have been avoiding, and it is time to use it. Surveillance is not a category of technology. It is what monitoring becomes when the person being monitored can see the output and do nothing with it, or cannot see it at all, and when the record outlasts the reason it was made.
A Useful Design Decision
Analiese Brown, the chief people officer at Campminder, built a career-levels project in Claude, connected to the company's job descriptions and its performance philosophy. Employees use it to ask questions about their own progression, the sort of question people historically asked a manager if they trusted the manager.
Her team can see the questions people ask and the responses the system generates. That is the same arrangement as everything else I’ve covered here, one group watching what another group does inside a tool. What she does with it is different.
She treats it as a safety net, watching for where the answers are thin or where people keep circling the same uncertainty, because that is where human guidance is needed and isn't reaching. Her framing is that automation doesn't replace human guidance, it protects it by showing where it's needed.
Brown is watching the quality of the support rather than the performance of the person, and what the system produces is not a record but a prompt for someone in HR to go find the employee who keeps asking the same question three different ways.
Procurement is Answering On Your Behalf
These three questions are being answered right now inside most companies that have bought a notetaker, a transcription tool or an analytics layer in the past 18 months, and in a lot of those companies they are being answered by a renewal date, a default retention setting and a vendor's onboarding checklist.
That is not a failure of ethics. It is a failure of sequencing. The tool arrives, the pilot goes well, the coaching value is real and visible in the first month, and the retention policy is whatever the vendor shipped. By the time anyone asks how long the transcripts live and who else can pull them, there are 14 months of them.
Russell, Wilkins and Brown all made their decisions before deployment, and none of them needed a governance framework to do it. They needed to have decided what the visibility was for.
