The First AI Workplace Crisis Isn’t Job Loss.

For three years, the dominant fear surrounding artificial intelligence was simple. AI will replace workers.

Executives issued warnings. Investors funded automation dreams. Economists debated displacement curves. Employees quietly updated their resumes. The narrative was everywhere, and it was loud.

But in 2026, something unexpected is happening.

The jobs apocalypse has not arrived yet. The productivity revolution has not materialized. And the enterprises that spent billions deploying AI are increasingly facing a different, quieter, and arguably more embarrassing problem.

They cannot prove any of it is working.

Despite record AI investment, global enterprise AI spending surpassed a whopping $200 billion in 2025, but productivity gains remain stubbornly difficult to measure. Many AI pilots have failed to scale. ROI justifications are growing thinner.

And in boardrooms from Sydney to San Francisco, finance teams are starting to ask uncomfortable questions that nobody prepared answers for.

The first major AI workplace crisis is not about replacing humans. It is about something far more mundane. It is about measurement.

The Great Adoption Race

To understand how we got here, we have to go back to 2023.

When Chat GPT crossed 100 million users in two months, something shifted inside corporate leadership. The fear of missing out became institutional (Enterprise FOMO). Boards wanted AI strategies. Investors wanted AI narratives.

Executives who hadn’t yet deployed AI tools were quietly questioned about whether they understood the moment. The message that cascaded through organizations was, in its simplest form: use AI.

And so companies deployed. Copilots. Assistants. Agents. Enterprise AI platforms. Microsoft embedded AI into Office. Google embedded it into Workspace. Salesforce and ServiceNow embedded it everywhere else. The procurement wave was real, and it was fast.

But here is what got lost in the speed.

The question was never, ‘what value does this create?’

The question was, ‘are we using it?’

Adoption became the goal. Not outcomes. Not profitability. Not productivity. Adoption. The KPI for the AI era, in its early form, was essentially: participation.

Tokenmaxxing: When Usage Becomes the Target

A natural consequence followed. If adoption is the goal, then someone needs to measure adoption. And the easiest thing to measure is activity.

Across major enterprises, reports began emerging of AI usage being tracked, discussed in performance contexts, and incorporated into expectations. Shopify’s CEO declared AI fluency a core competency. Amazon signaled that employees should demonstrate AI use in their workflows.

Meta began incorporating AI engagement expectations into management discussions. Uber and Microsoft followed with language about AI becoming a baseline skill, not an optional one.

None of this is inherently wrong.

Encouraging employees to learn and use new tools is reasonable. Technology adoption requires cultural push. Nobody got good at spreadsheets by choice alone.

But something subtler started happening alongside the encouragement.

AI activity became a proxy for productivity.

Not because leaders explicitly decided this. But because activity is visible and outcomes are not. Usage can be dashboarded. Value cannot be easily dashboarded. And when organizations are under pressure to justify billion-dollar AI investments, dashboardable metrics become deeply attractive.

This is the environment that produced a term most people outside of tech circles had never heard until recently.

Tokenmaxxing: The behavior refers to the act of maximizing AI usage, generating long, verbose, AI-produced outputs not because they are more useful, but because they are more visible. More tokens. More activity. More proof that someone is engaging with the technology.

Tokenmaxxing is not interesting as a Silicon Valley curiosity. It is interesting as a leading indicator.

Because it tells you that employees are reading the incentive environment correctly. They know what is being measured. And they are optimizing for it.

Goodhart’s Law Is Older Than AI

In 1975, British economist Charles Goodhart made an observation that has since become one of the most replicated findings in organizational behavior.

When a measure becomes a target, it ceases to be a good measure.

The history of that principle playing out in workplaces is long and depressing.

Software engineers rewarded for lines of code wrote longer code. Call center workers measured on call duration shortened calls without resolving problems.

Sales teams tracked on activity metrics filled pipelines with low-quality leads. Marketing teams evaluated on website traffic bought it. Email cultures that rewarded responsiveness produced endless shallow replies.

In every case, the measurement was reasonable in isolation.

More code, more calls handled, more leads, more traffic, more replies. These are not absurd things to want.

The problem is that measuring activity invites optimization of activity, not outcomes. And AI metrics are running the same experiment.

More AI usage does not equal more business value. That relationship has not been established. In fact, **Reuters reported** a specific version of this problem at scale: enterprises watching token consumption and inference costs rise significantly while struggling to point to proportional business gains.

The report discussed the emergence of AI-generated “work slop” high-volume, low-value AI output that consumes compute and clutters workflows without improving outcomes.

Goodhart would not have been surprised.

The Productivity Surge That Hasn’t Arrived

Here is where the AI story gets genuinely complicated.

The case for AI displacing human labor at scale depends on a productivity shock. If AI is automating meaningful portions of knowledge work, that transformation should eventually show up in economic data. Productivity growth should accelerate. Output per worker should rise. Labor demand in certain categories should soften.

Some of this has happened at the task level. AI tools demonstrably improve the speed of certain narrow activities like drafting, summarizing, coding specific functions, generating variations of content. These are real gains.

But at the macroeconomic level, the productivity surge has been difficult to find.

**Oxford Economics** has argued that if AI were already replacing labor at scale, productivity growth should be accelerating. Yet aggregate productivity data has not shown the kind of broad-based surge that many AI disruption narratives would predict.

In a January 2026 research briefing, the firm concluded that companies “don’t appear to be replacing workers with AI on a significant scale” and noted that AI’s impact on labor markets remains uneven and difficult to detect in macroeconomic data.

Even in a follow-up analysis published in May 2026, Oxford Economics described AI’s effects on productivity and employment as “mostly muted” so far, despite growing adoption and investment.

This creates a puzzle.

Enterprises are spending more on AI. Employees are using more AI. Token consumption is rising. Enterprise AI budgets are growing. And yet the measurable productivity impact that would justify all of it remains elusive at scale.

Three explanations are commonly offered.

The first is lag, that productivity effects from major technology transitions take years to materialize, and AI is simply early. This is historically plausible. Electrification took decades to show up in productivity data. Computers did too.

The second is distribution, that productivity gains are real but concentrated in specific companies, workflows, and use cases, not yet visible in aggregate figures.

The third is measurement, that we are improving the wrong things and measuring the wrong proxies, and the actual gains either do not exist yet or are hidden beneath adoption theater.

All three may be simultaneously true. But the third one is the uncomfortable one.

The CFO Conversation

Finance has entered the building.

In the first phase of enterprise AI, CFOs were largely absent from the conversation. The technology was framed as strategic and forward-looking, a category where short-term ROI questions felt like missing the point.

That phase appears to be Cessation

Reuters documented the growing scrutiny of AI-related infrastructure costs at major enterprises, particularly around rising token consumption driven by increasingly complex prompts and heavier AI workflows. The CBA (Commonwealth Bank of Australia) case illustrated how enterprise AI costs can scale in ways that outpace business value realization.

The questions CFOs are now asking are not philosophical.

What are we actually spending on AI?

Which specific workflows have measurably improved?

What is the cost per outcome, not cost per token?

Where did we expect ROI and where did we find it?

These are reasonable questions. There are also questions that many AI projects are not yet structured to answer, because adoption, not outcome measurement was the original brief.

Why This Doesn’t Mean AI Is Failing

It is worth being direct about what this argument is not. This is not a case that AI is useless, overhyped, or doomed.

Every major technology wave has passed through an accountability phase. The internet produced the dot-com way before producing Amazon, Google, and the modern digital economy. Cloud computing produced years of skepticism about ROI before becoming indispensable infrastructure. Mobile was dismissed as a toy before it restructured entire industries.

AI will almost certainly create enormous value. The task-level evidence is real. The trajectory of model capability is real. The transformation of specific industries is already visible in healthcare, legal research, software development, and logistics.

“The current problem is not that AI cannot create value. The current problem is that enterprises often cannot distinguish between creating value and performing value.”

The current problem is not that AI cannot create value.

The current problem is that enterprises often cannot distinguish between creating value and performing value.

Experimentation is not productivity. Adoption is not profitable. Activity is not an outcome.

Organizations that confuse these categories do not fail because the technology failed them. They fail because the measurement system failed them first.

The Crisis the Headlines Missed

Getting back to where this began.

For three years, the conversation about AI and the workplace centered on one question.

Will AI replace humans?

That question has not been answered yet. The honest answer in 2026 is: not at scale, not yet, and not in the way most predictions suggested.

But a different question has quietly become urgent.

Can organizations tell the difference between AI activity and AI value?

Because the evidence of 2025 and 2026 suggests many cannot.

When enterprises track adoption rates while ignoring whether adoption created returns, that is a measurement failure.

When rising token consumption and rising AI budgets coexist with stalled productivity growth, that inflated the sector and caused failures..

The first major AI workplace crisis will not be mass unemployment. It will be something far more recognizable.

Organizations measuring the wrong things, incentivizing the wrong behaviors, spending on the wrong metrics, and discovering too late that the productivity revolution they announced to investors and boards was partly adoption theater.

The Only Question That Actually Matters Now

Enterprise AI is entering what will probably be called its accountability phase.

The question is no longer whether to use AI.

The real question is whether organizations can accurately measure the value AI creates, where that value appears, and how much of it is genuine rather than simply a reflection of increased AI activity.

It requires defining outcomes before deployment, not after. It requires measuring baselines. It requires resisting the pull of visible metrics when invisible ones matter more. It requires asking, every time someone shows you a usage dashboard, what outcome does this usage correspond to?

The biggest obstacle to AI realizing its potential may not be the models. It may not be the infrastructure. It may not even be change management or employee resistance.

It may be the oldest problem in organizational life. The tendency to measure what is easy. And to forget what actually matters.

If this piece resonated, consider sharing it with a colleague building an AI strategy. The conversation about AI accountability is just beginning.