The Performance of "Belief"

The problem with AI at work is not just bad outputs. It's that too many organizations now require a performance of belief.

Not a result. Not proof. Belief. The willingness to say, out loud, in the right meeting, that AI is transforming your work — whether or not it actually is. And once an organization starts rewarding that performance, it stops being able to hear the one sentence it needs most: this is not working.

The coordination problem nobody wants to name

Cory Doctorow, writing about the consultant Nikhil Suresh, put a name to something a lot of us have felt but couldn't quite say. Suresh spent a year and a half talking to hundreds of executives and their subordinates about what AI was actually doing for their businesses. His finding: in a year and a half, zero enterprise AI projects he witnessed succeeded. Zero.

But that's not the disturbing part. The disturbing part is why the story keeps going anyway.

Suresh describes it as a coordination problem around honesty. If a leader admits the gains aren't there, they defect from a story every one of their peers has publicly endorsed. They get fired by embarrassed colleagues who've now been implicitly called liars, and replaced by someone who will recite the script. If everyone could tell the truth at once, there might be hope. But nobody can coordinate that moment. So everyone keeps performing.

The result is a system where just about everyone — boards, executives, employees, vendors, consultants — has an incentive to overstate what AI is delivering. The head of AI at a billion-dollar firm told Suresh his own job was "totally fraudulent," but also "the only promotion pathway remaining." That's not a technology problem. That's an honesty problem wearing a technology costume.

AI washing and the token leaderboard

Watch what happens to the people at the bottom of that structure.

Employees who voice honest, informed objections get passed over or laid off. So they adapt. They "AI wash" their work — do the job the way they've always done it, then say Claude did it. Some go further, writing circular processes where one chatbot prompts another and back again, burning tokens for no reason except to climb a corporate "token leaderboard." The point isn't output. The point is to look like a believer.

Think about how strange that is. People are spending real money and real time manufacturing evidence of enthusiasm for a tool, because enthusiasm is what gets measured. Usage becomes a loyalty oath. Consumption becomes a performance review. And nobody is asking whether any of it made the work better, because asking is the dangerous part.

This is where the meter comes in. Token burn feels free until it doesn't. We saw the sticker-shock pattern with Ford — the moment scaled usage stops being an abstraction and starts arriving as an invoice. The cruel twist is that by the time the bill lands, the organization is already culturally committed to pretending the value was there all along. You can't easily walk back a triumph you announced to the board.

The economics underneath the story

Zoom out and the same performance is running at industry scale.

A Nikkei investigation reported by Futurism found that five US tech giants — Alphabet, Microsoft, Amazon, Meta, and Oracle — are carrying roughly $1.65 trillion in off-balance-sheet debt tied to their AI infrastructure buildout. That's more than the $1.35 trillion they reported officially. Meta alone accounts for around $420 billion of it. Observers have started making the obvious comparison: Enron, which collapsed in 2001 for propping itself up with debt hidden behind special purpose vehicles until the house of cards fell.

I'm not predicting a collapse. I'm pointing at a pattern. The visible story is inevitability and growth. The less visible story is enormous obligation, structured to look healthier than it is. That's the macro version of the token leaderboard — spending arranged to signal confidence, with the accounting doing the work that results are supposed to do.

When the performance of belief runs at every level, from the analyst faking token usage to the balance sheet hiding the real number, you don't have a strategy. You have a shared refusal to look.

What a commitment decree actually signals

I've watched this movie before, and I want to be specific about it.

Years ago at Western Union, a new CEO arrived and asked all of us to sign a giant public commitment — a decree, laid out in the cafeteria, that we would get behind the new direction. It was theater. Sign here, in front of everyone, and demonstrate your belief. Something in it read as authoritarian to me, so I slipped out a side door and never signed. It felt less like alignment and more like a loyalty test dressed up as culture.

Twelve months later, the company was in trouble with the FTC for misleading investors. I had already left for IBM.

I'm not claiming the unsigned page predicted the fraud. But the two things came from the same soil. An organization that needs you to perform belief in a cafeteria is usually an organization that has already stopped rewarding people for telling the truth. The decree and the FTC problem were the same instinct at two different scales: manage the story, suppress the doubt, and hope the reality catches up before anyone notices the gap.

It rarely does.

Honesty debt

I've written before about trust debt — the compounding cost of shipping experiences your audience tolerates but doesn't believe. There's a sibling to it, and AI is running up the balance fast. Call it honesty debt: the accumulating cost of every moment an organization made it safer to lie than to tell the truth.

Honesty debt is quiet at first. A pilot that can't fail because no one defined success. A dashboard everyone praises and no one trusts. A leader who mandates AI in the core work of skilled professionals the way no serious CEO would mandate a surgical technique without the surgeons' agreement. Each of these is a small loan against your ability to see straight. And like all debt, it comes due at the worst possible moment — usually when the meter arrives, or the returns don't, or a regulator starts asking questions.

This is the whole point of a Human-Driven Operating Model. HDOM isn't a framework for being nice about technology. It's a framework for keeping an organization honest — for building the conditions where someone can say "this tool helps here and nowhere else," or "this pilot has no proof," or "this bill is bigger than the value," and keep their job for saying it. The health of your operating model is not measured by adoption. It's measured by whether the truth can still travel upward without being punished on the way.

Strip that out and you're left with a company that can no longer tell the difference between working and looking like it's working. Which brings me to the part that should keep leaders up at night.

We spend a lot of energy worrying about AI hallucinations — the model confidently making things up. But the model's hallucinations are legible. You can catch them, correct them, route around them. The dangerous ones are the hallucinations the organization tells itself: that the pilot succeeded, that the usage means value, that everyone signing the decree actually believes it.

The most dangerous hallucination in enterprise AI may not be the model's. It may be the organization's.