Signal vs Noise: The AI Cost Is Now Bigger Than the Employee It Replaced
The headline. The context it left out. The move.
Signal vs. Noise | TheMoneyZoo.com
The Headline
“Some Companies Are Finding That AI Costs More Than the Employees It Replaced”
Analyst Jack Gold has documented companies where token costs are exceeding the cost of the employee within a month or two. Uber’s CTO admitted the company maxed out its full-year AI budget four months into 2026. An Nvidia VP acknowledged that AI compute now costs more than the employees using it. The phenomenon has a name: “tokenmaxxing” — the usage binge that happens when companies deploy agentic AI without governing it.
The narrative building around these numbers: AI isn’t actually cheaper than humans. The cost case that justified cutting headcount was wrong. The replacement is failing.
That narrative is incomplete. And for anyone making career decisions based on it — in either direction — the incompleteness matters.
The Context They Left Out
The “AI costs more than employees” headline is accurate for specific companies in specific circumstances. What it doesn’t tell you is why.
Meta built a real-time dashboard called “Claudeonomics.” It ranked 85,000+ employees by token consumption. Live. In real time. The signal it sent to every person on that leaderboard was unambiguous: usage is how we measure you. So usage exploded. Jensen Huang told engineers they should burn tokens worth half their salary annually. That’s a directive to consume, not a directive to produce. The tokens got burned. The useful output didn’t necessarily follow.
This is Deming’s most famous observation dressed in modern infrastructure. When you manage by visible figures, you get more of the visible figure. Token usage is visible. Useful output is not. Meta measured usage. They got usage. Their CTO has since walked it back — nobody should use AI tools just for the sake of using them. That walkback is the organization admitting it managed by the wrong metric.
Agentic AI consumes up to 1,000x more tokens than standard usage. The mechanics: an agent re-feeds its entire work history and codebase every time it decides on its next action. One developer running an ungoverned agent on a large codebase for a week can generate a token bill that looks, on a spreadsheet, like a full-time salary. That’s not AI being expensive. That’s AI being ungoverned.
Anthropic’s own enterprise data tells a different story. The average Claude Code engineer costs approximately $13 per day. The blowups that are generating the headlines are the ungoverned tail — the outliers, not the mean. The companies reporting AI costs exceeding employee costs are the companies that deployed without budgets, metrics, or ROI thresholds. They are not a representative sample of what governed AI deployment actually costs.
The Real Problem
The companies generating these headlines didn’t have an AI problem. They had a management problem that AI made expensive.
Token usage became a proxy for productivity because it was measurable. It was visible. It showed up in a dashboard. So it got optimized — not toward useful output, but toward the metric. This is the same failure that produces sales teams that maximize call volume instead of closed deals, or customer service organizations that optimize handle time instead of resolution. The metric and the outcome diverged, and the metric won.
The organizations that cut junior employees to fund AI deployments and then ran those deployments without governance didn’t save money. They moved the waste from a salary line to a token line and lost the institutional knowledge, the training pipeline, and the junior-to-senior development pathway in the process.
A junior employee costs a salary and produces work plus a future senior. An ungoverned AI agent costs an unpredictable bill and produces work plus nothing. The firms that traded one for the other managed to buy the worst of both.
The right comparison was never token bill versus employee salary. The right comparison is cost per unit of governed output. A junior who produces ten useful deliverables a week at $1,200 weekly cost is $120 per deliverable. An AI agent that produces fifteen deliverables a week at $2,000 in token costs — but requires a domain expert to verify and correct six of them — is producing nine net useful deliverables at $222 each, plus the cost of the domain expert’s verification time. The math only works when the denominator is honest.
The Move
The discipline that applies to every cost center now applies to AI spend. Budgets. Metrics. ROI thresholds. Usage limits. The companies that deployed AI without these controls got what ungoverned cost centers always produce: costs that scale without corresponding value. The emerging rule of thumb is already clarifying: AI-augmented workers need to be roughly 2x as productive to justify the combined spend. That’s not a particularly high bar. It is a bar, which is more than most organizations set when they deployed.
The human who governs AI output is the differentiator. Not the human who burns the most tokens. Not the one who runs the most agents. The one who understands the domain deeply enough to direct the AI toward useful work, verify that the output is actually correct, and catch the six errors in fifteen deliverables before they compound into something expensive. That is the “You Can’t Outsource the Audit” argument applied to AI spend: the verification layer requires domain expertise that no token budget can replace.
The tokenmaxxing story is a warning about intention, not about AI. The organizations that deployed AI with clear output goals, governed usage against budgets, and maintained the human expertise to evaluate what the AI produced are not in the headlines. They are running $13/day engineers and getting real productivity returns. The story of ungoverned AI spending is not evidence that AI doesn’t work. It’s evidence that deploying any resource without intention produces waste. This particular waste is just more visible because it shows up as a line item on a cloud bill.
The Scot Free Take
The executives who sat in offsites and said “AI will free our people to work on higher-impact items” — and then didn’t define what those items were — are now looking at token bills that exceed the salary of the person they let go to fund them.
That outcome was predictable. It was predicted here. In the absence of intention, reclaimed capacity doesn’t flow toward value. It flows toward waste. The only difference between wasted human capacity and wasted token budget is that token waste shows up on an invoice.
What this means for your career: the tokenmaxxing story does not change the direction of AI’s impact on the labor market. It doesn’t reverse the compression of entry-level generalist roles. It doesn’t bring back the jobs that were eliminated. What it does is sharpen the argument about what human value actually looks like in an AI-augmented organization.
It’s not the person who uses AI the most. It’s the person who uses it best. Governed. Directed. Verified. Producing output that the organization can actually use without someone downstream fixing six of the fifteen deliverables.
Deep domain expertise is the governance layer. It’s the thing that makes AI output usable rather than voluminous. It’s the thing that no token budget can replicate.
Career Warfare is the honest name for what’s happening right now. The companies spending more on tokens than they saved on salaries are losing it. The workers who built depth while others optimized for visibility are winning it.
The map is the same as it’s always been here. Build depth. Build credentials. Build the capability to verify and direct what the AI produces. That’s not a defensive posture. It’s the only one that compounds.
— Scot Free
Related: There Will Be Blood: What AI Change Management Is Actually Doing to the Workforce → | You Can’t Outsource the Audit → | AI Isn’t Taking Your Job. It’s Killing the One You Were Supposed to Get First. →
