An engineer at OpenAI recently processed 210 billion tokens in a single week – enough text to fill Wikipedia 33 times over. At Anthropic, one Claude Code user ran up a bill exceeding $150,000 in a month. These are not cautionary tales, apparently they’re bragging rights.
Welcome to tokenmaxxing – the latest way for knowledge workers to perform productivity without necessarily proving it.
Kevin Roose’s recent New York Times piece, “More! More! More! Tech Workers Max Out Their A.I. Use,” and the subsequent Hard Fork podcast discussion, describe a growing culture inside tech companies where employees compete on internal leaderboards to demonstrate how many AI tokens they consume. Shopify’s CEO Tobi Lutke has made AI usage a baseline expectation, folded into performance reviews. Nvidia’s Jensen Huang has floated giving engineers a “token budget” alongside their salary.
The framing is aspirational – these are power users, early adopters, the future of work made flesh. But Roose asks the question that should be nagging at every manager reading along: is this genuine productivity, or “a very expensive new form of looking busy?”
For those of us who have spent years in remote teams, this question is not new. It is depressingly familiar.
The jacket on the back of the chair
Before there were token leaderboards, before there were distributed teams, there was the jacket on the back of the chair.
You remember that person. You know that person. They arrived at the office conspicuously early, draped their coat over their seat so everyone could see they were “in,” then disappeared for an extended coffee run. The performance was the point – being seen to be present mattered more than what you actually produced while you were there. When the boss turned up, they’d be there, animatedly shouting into their phone, or typing intently, putting the rest of the team to shame for their sheer level of busyness.
When work went remote, the behaviour adapted. Sending Slack messages at 6am. Scheduling emails to land at 11pm. Commenting on threads while on annual leave just to show you’re still in the chat. Keeping your status dot green during lunch. Checking out shared documents just to leave a timestamp, back in the days where everything had to be checked in and out. The tools changed; the anxiety underneath them did not.
Digital presenteeism is remarkably well-documented at this point. A 2025 Forbes survey found that 58% of employees admit to frequently pretending to work. Harvard Business Review and workplace research consistently shows that employees focused on appearing busy complete significantly fewer meaningful tasks – with some studies finding up to a quarter less productive output per week.
More than half of knowledge workers report feeling pressure to demonstrate they are online during specific hours regardless of whether those hours are when they do their best work. The root cause is consistent: when organisations cannot or will not measure outcomes, they default to measuring visibility. And visibility is easy to game.
Productivity paranoia never went away
Microsoft’s landmark 2022 survey of 20,000 people across 11 countries put a name to the dynamic: productivity paranoia. While 87% of employees said they were productive, only 12% of leaders were confident that was true. Nearly half of hybrid managers admitted they struggled to trust their teams.
The response, predictably, was not better outcome measurement. It was surveillance. Almost half of the business leaders surveyed had installed monitoring software on employee devices. The message was clear – if we cannot see you working, we will watch your screen instead. Because we don’t trust you.
This paranoia has only intensified in the years since. A 2026 survey found that 69% of companies now measure return-to-office compliance, up from 45% in 2024. A third take disciplinary action for non-compliance. Badge tracking, performance-linked attendance scores, promotion penalties for working from home – the infrastructure of distrust is expanding.
And it is not working. Ninety-nine percent of companies with RTO mandates report declining employee engagement. Eighty percent admit to losing talent over the policies. The surveillance arms race creates exactly the behaviours it claims to prevent – people focus their energy on being seen rather than being effective. A new product category emerged on Amazon, in response to demand – people were searching for physical ‘mouse jigglers’ to keep their Slack light green and their activity apparent.
Europe, to its credit, has largely held a more pragmatic line. European banks average 3.4 required office days per week compared to 4.2 in North America. Over 40% of European job seekers say they would reject a role without remote options. The EU AI Act imposes specific obligations around algorithmic management and worker monitoring. But the cultural pressure is global, and the anxiety travels across borders faster than legislation can contain it.
Tokens are just the latest wrong metric
What makes tokenmaxxing particularly instructive is how perfectly it mirrors every previous failed attempt to measure knowledge work through inputs rather than outcomes.
Lines of code. Emails sent. Meetings attended. Hours logged. Documents created. Commits pushed. Each of these has, at some point, been treated as a meaningful proxy for productivity. Each turned out to measure effort – or the appearance of effort – rather than value. As a writer who straddled the twin disciplines of content marketing and copywriting, it was often challenging to explain why the shorter, tighter version of the same message might actually cost them more to produce (while I cringed at the relentless adoption of the term ‘content’ anyway, which felt like something delivered in a vat and slapped on with a trowel to smooth over cracks in messaging strategy. How much, for how many words?)
Tokens consumed is the same category of mistake, with one crucial difference: it is expensive. When someone games their Slack status or pads a blog post, the cost is negligible. When someone burns through $150,000 worth of AI tokens in a month to climb an internal leaderboard, that is a direct hit to the balance sheet.
As FinOps specialist Kevin Prokopetz noted in response to Roose’s article, unmanaged AI tool adoption creates massive token burn with “zero visibility into actual ROI.” Nate Patel put it more bluntly: “If it doesn’t tie to shipped outcomes or saved time, it’s just a burn.” Ouch!
The parallel to remote work surveillance is obvious. Monitoring whether someone is at their desk – physical or virtual – tells you nothing about whether they are solving the right problems, and creating value for the organisation. I once hired a developer on Upwork to do a job for me, and he turned on their monitoring software by default – so it sent me screenshots of his desktop every 10 minutes, images of lines of code that meant nothing whatsoever to me. All I wanted was the final deliverable, and I didn’t care if it took him 5 minutes or 5 days work to produce it by the agreed deadline.
Just as those screenshots held no intrinsic value for me, organisations monitoring how many tokens users consume tells you nothing about whether the AI output advanced a business objective. In both cases, the metric is a comfort blanket for managers who have not done the harder work of defining what “done well” actually looks like.
The AI washing connection
This is happening against a backdrop that makes the stakes considerably higher. Companies are simultaneously cutting staff while citing AI capabilities that do not yet exist at the scale claimed.
A Harvard Business Review survey of over 1,000 executives found that 60% had made workforce reductions in anticipation of future AI capabilities, while only 2% had made significant cuts based on actual AI implementation. Sam Altman himself has acknowledged that “there’s some AI washing where people are blaming AI for layoffs that they would otherwise do.”
The combined message employees are receiving is deeply contradictory. Use more AI. Use it visibly. Your job may depend on how enthusiastically you adopt these tools. But also, the company may cite those same tools as the reason your role no longer exists.
In this environment, tokenmaxxing is not irrational, it’s simply a survival strategy. If your employer is building leaderboards that track AI usage, and your employer is also making redundancies justified by AI adoption, then maximising your visible AI consumption is a perfectly logical response to the incentives on offer.
That makes it understandable, but it doesn’t make it productive.
What should actually be measured
The fundamental problem has not changed since Peter Drucker first wrote about knowledge worker productivity decades ago. There is no universally accepted way to measure it, and organisations keep reaching for whatever is easiest to count rather than whatever is most meaningful.
For remote teams – and increasingly, for AI-augmented teams of any configuration – the answer has always been the same. Measure outcomes.
Not tokens consumed, but problems solved. Not hours logged, but deliverables shipped. Not AI usage frequency, but whether the AI-assisted output was faster, better, cheaper, or some combination of the three compared to the alternative.
This requires managers to do something much harder than reading a leaderboard. Instead they need to define clear objectives, agree on what success looks like, and then trust their teams to get there – checking in on progress and removing obstacles (like token limits) rather than monitoring keystrokes or token counts.
The companies that have figured this out – and there are plenty of them, many of them fully remote – consistently outperform on retention, engagement, and actual productivity. The research is not ambiguous on this point.
The remote work lesson the AI era needs to learn
Remote workers have been having this argument for the better part of a decade. We know what happens when you measure presence instead of performance. We know what happens when you install surveillance tools instead of building trust. We know that presenteeism – digital or otherwise – actively reduces the quality of work while consuming enormous emotional energy.
Tokenmaxxing is the AI-era version of the green Slack dot. It looks like engagement. It feels like progress. It might even correlate loosely with genuine productivity in some cases. But as a metric, it measures the wrong thing – and worse, it incentivises the wrong behaviours.
The companies building token leaderboards today will learn what the companies installing employee monitoring software learned before them, and what the companies obsessing over office badge swipes learned before that. Measuring inputs does not produce better outputs. It produces better performances.
If your organisation is deploying AI tools – and it should be – the question worth asking is not “how much are people using them?” It is “what are people achieving with them?” The answer to the first question is easy to obtain and nearly meaningless. The answer to the second requires thought, context, and honest conversation about what matters.
That has always been the hard part. AI has not changed it. AI has just made the cost of getting it wrong considerably more visible – $150,000 per month visible, in some cases.
The jacket is on the back of the chair. The Slack dot is green. The token count is climbing. The screen is filling up with lines of code or content… And the real work – defining, measuring, and rewarding genuine outcomes, growing the bottom line – remains undone.
Sources:
- Kevin Roose, “More! More! More! Tech Workers Max Out Their A.I. Use,” The New York Times, March 2026
- Microsoft Work Trend Index 2022: “Productivity Paranoia” – survey of 20,000 people across 11 countries
- MyPerfectResume, “The Great Compliance: RTO Mandates and Employee Monitoring 2026”
- Harvard Business Review, “Companies Are Laying Off Workers Because of AI’s Potential, Not Its Performance,” January 2026
- Forbes/Robinson 2025 survey on performative work behaviours
- Shopify internal memo, Tobi Lutke, April 2025
- Sam Altman on AI washing, Fortune, February 2026
