The Contractor Nobody Signed Off On

There was a time, not long ago, when routing work to an outside contractor without disclosing it would have been a serious conversation with HR. Maybe a written warning. The work belonged to you; the undisclosed help didn’t.

Then AI arrived. The same organizations that would have demanded disclosure for a human collaborator actively encouraged the undisclosed one. Use it. Use more. There are no limits. Token consumption wasn’t a cost center — it was a productivity signal.

Tokenmaxxing and the culture it built

The word for what followed is tokenmaxxing — optimizing for token consumption rather than business outcomes. Meta reportedly built internal leaderboards ranking employees by how much AI they used. The message was clear: more is more. If you weren’t saturating the model, you weren’t working hard enough.

This wasn’t malicious. It was a reasonable early assumption: the tools were demonstrably useful, the cost was abstracted away from the employee, and measuring adoption was easier than measuring value. So companies measured adoption.

The problem with measuring the wrong thing is that you get very good at producing the wrong thing.

The bill arrived

By 2026, Uber had burned through its entire annual AI tools budget in four months. Microsoft quietly began canceling the majority of its internal Claude Code licenses, steering engineers back to GitHub Copilot — a cheaper, less capable tool it already owned. JPMorgan published a note titled “AI Token Costs are Eating Internet Profits Alive.”

What happened wasn’t that the tools stopped working. They worked. The problem was that tools working at scale, without governance, produce invoices that surprise CFOs.

The correction was swift. Providers shifted to stricter usage-based billing. Companies imposed limits. Engineers were told to route simple tasks to cheaper models and think before firing off a prompt. The enterprise swung from “use everything” to “justify every token.” That swing is healthy. It’s also overdue.

The governance question nobody asked

Here’s what’s strange: if you had told your manager in 2022 that you were routing work to an external consultant on a per-task basis — no contract, no disclosure, no budget line — you would have been stopped. The organization would have asked the questions it asks of every outside collaborator: Who owns the output? Who’s liable when it’s wrong? Who’s accountable for what it touched?

AI got a pass on all three. Not because the questions didn’t apply, but because the tool was new and enthusiasm outpaced oversight.

Notice which question finally got asked. Not IP. Not liability. Not accountability. Cost — because cost is the one that arrives as an invoice, and an invoice is hard to ignore. Finance found the line item before legal found the exposure. The reckoning everyone is calling an AI cost correction is really just the first of the contractor questions getting answered, and it happened to be the cheapest one to notice.

The other three are still open. We’ve started rationing the tokens. We have not, in most organizations, decided who’s accountable when the unsanctioned contractor ships something wrong.

The restraint being applied now isn’t a rejection of AI. It’s the first installment of the governance that should have accompanied the adoption — and the smallest one.

The bigger bet

At the individual token level, the math is being scrutinized. At the macro level, the math is harder to parse.

Goldman Sachs estimates $7.6 trillion in cumulative AI infrastructure CapEx between 2026 and 2031. IBM’s CEO has been blunt that the math doesn’t work: a 100-gigawatt buildout runs to roughly $8 trillion at today’s costs, requiring something like $800 billion in annual profit just to service it. His conclusion — “there’s no way this is going to pay off at today’s infrastructure costs” — isn’t that scale fixes the unit economics. It’s that scale makes the hole bigger.

I don’t know if that’s right. I hope the economics improve — not because I want the bet validated, but because the underlying capability is genuinely valuable. There are real problems this technology can solve. But there’s a difference between a technology that’s useful and an investment thesis that pays. The former is already true. The latter is still being decided.

The scrutiny at the employee level is a signal. When the same scrutiny reaches the capital allocation level — when it moves from tokens to terawatts — that will be the more important conversation.

The principle that holds

Measure what matters, not what’s easy to measure. The tokenmaxxing era confused activity for output. The correction is an invitation to ask the harder question: what is this actually producing, and for whom?

The contractor nobody signed off on is now on the org chart. Whether it earns its line item is still being negotiated.

---

Related Reading

• Don’t Automate the Rube Goldberg Machine — Eliminate before you automate. The same principle applied to workflow.
• The Cost of Every Yes — Every unchecked yes is a hidden no somewhere else.
• It’s Not What You Can Do — It’s What You Can Get Done — Why AI-as-amplifier doesn’t eliminate the choice problem.
• Why AI Is Different: Every Prior Tool Took Sides — The larger arc this economic reckoning sits inside.