AI Has an Accounting Problem

A balance sheet can’t tell you the quality of a kitchen

On paper, the best restaurant in town is an ordinary business. You buy the knives a great chef uses, high-end ovens, the same fish off the same boat. On the ledger, you park it on the balance sheet.

What you can’t park are thirty years in the chef’s hands, the knowing when the sauce will break, a beat before it does.

The equipment is an asset, an easy thing you can buy, own, and carry on the books. The chef is a cost, and the value she adds shows up in the quality of the food.

When the numbers tighten, the move looks obvious. Keep the gear. Let the chef go. Backfill with someone cheaper. The books show a saving, and the same kitchen, with the same equipment, turns out food no one crosses the street for.

Think about each company as that kitchen, now with methods autonomously embedded in the operation. The line runs more smoothly. Models handle repetitive tasks that change or remove people’s roles. The core gets hollowed out.

The pattern reflects the economic trajectory when the numbers drive the bus.

What Accounting Can and Can’t See

Consider the accounting rationale behind the AI boom. Companies continue to plow capital into models and wrappers, feeding them into racks of GPUs that burn cash and throw off horsepower in equal measure.

Investment lands on the balance sheet like a stamp of arrival. The number that says, look, we’re becoming AI-ready. Labor moves the other way, off the books. The layoffs get a new label. “AI transformation,” whether AI had anything to do with them or not.

This quarter, reality reached the CFO’s desk.

The layoffs cut too deep. Companies gutted functions on the assumption AI could hold them, and it couldn’t. Hire-backs entered the picture.

Realization landed that models and tokens meter like electricity, a mouth requiring constant feeding.

Transformation and L&D teams sense a new job they didn’t know they had: helping teams aim token spend at programs that move the business.

All land on the cost line, expensed as it happens.

Under IFRS’s IAS 38, a company doesn’t control what its skilled people or its training actually produce — the benefit lives in them, not in the firm. No control, no asset. The most valuable thing on the payroll never makes the balance sheet. The rule reads the same in a data center as in a kitchen.

The rule never had to account for agent workers. It frames how to book them — as licensed software. What it can’t book is that they work. Agents learn, they accumulate, they get better on the job. The cost is legible. The labor isn’t.

Tangled incentives and new accounting mechanics keep the endgame out of view.

A spreadsheet can prove the spending dropped. It can’t show the work improved. The swap logs a saving either way — better output, worse output, the number looks the same.

The Wrong Scores

Right now the leading indicator inside companies is adoption. Inside and out to the Street, the AI uptake curve reads as progress. But a higher AI utilization number can’t tell whether anyone’s doing better work, or the right work at all.

Behind the adoption-at-all-cost sits a productivity bubble in plain sight. On any given AI project, training and literacy get 3 to 6 cents of every dollar. Zoom out, and it gets worse, not better: American companies spend about $103 billion a year teaching their people — against a buildout that runs $765 billion in 2026 alone and climbs toward $1.6 trillion by 2031. The machines get an order of magnitude more than the people who are supposed to run them.

The sharper number is the direction. As AI spend surged, formal training shrank — from 47 hours a year per employee down to 40. Capital floods toward AI components bought in bulk and starves what can’t be minted in a model: human capability. People build it the slow way, a step at a time, running on a different clock than the machines.

I see it up close. A company burns through a billion tokens and ships nothing that adds up to anything better. Someone walks us through the adoption dashboard, every line up and to the right, and calls it traction.

Now the token bill comes under a harder light. How the tokens were spent, and to what end, doesn’t matter. What’s missing is output per dollar of fuel. MPG for AI, and whether people are ready to raise it.

AI is not an advantage

Models get cheaper and better every quarter. The upgrades reach everyone the same day. Competitors get the same jump, at the same price. That’s the definition of a commodity, the price of entry.

So what then?

Orchestration becomes differentiation. Nuanced orchestration.

Anything repeatable, the next model absorbs. Link planning, processes, and information flows with a frontier model, and every competitor can link theirs the same way. The transmission of work increasingly embeds in the commodity model.

No model can orchestrate the reality a business demands. AI gains that reflect your data, your rules, your people, your structure, your goals, with the politics and regulations you have to navigate. The friction Silicon Valley writes off as waste is what a model can’t copy and a startup won’t track. That friction is a moat. The harder it is, the deeper it runs.

Why can’t someone just buy this, or copy it? Because lots of what runs companies isn’t written down. Sixty years ago, in The Tacit Dimension, Michael Polanyi put it plainly. We know more than we can tell. Models feed on what’s recorded, and they capture more of it every year. The part that was never recorded is the part that stays in the vault.

What holds is what resists the page. The sense that something’s about to break before any dashboard flags it. The judgment call that only makes sense inside your own reality, and changes the moment you try to explain it. It passes hand to hand, over years. No one can rent that.

Which brings us back to the ledger. What makes companies competitive is what accounting files as a cost. The edge sits in the expense column, mistaken for overhead. The know-how and judgment. It walks out the door before the executive suite notices it’s gone.

The Rebuild Is the Race

The last paradigm shift of equal weight happened a century ago, and it cost forty years of potential. Economist Paul David noted that electric motors arrived in American factories during the 1880s. The productivity gains didn’t arrive until the 1920s. The motors worked. The problem was literally in the building.

The factory ran on a single steam engine. A long driveshaft ran the length of the building. Each machine had to sit close enough to connect to the shaft with a belt. That shaft dictated how the plant ran. Where a machine could fit, how work flowed between tasks, the size of the floor all revolved around that one line of spinning steel.

When electric motors arrived, factories pulled the steam engine, bolted a big electric motor in its place, and ran the same driveshaft off it. Everything else stayed exactly where it was.

New power, old structure. And for twenty years, almost nothing improved.

The gains came when plants laid the floor out around the work. The rearrangement was the revolution. It wasn’t possible in the old plants. So, it had to wait for a new generation of buildings. Each one was run by people who never worked in the old way.

AI grants no such patience. Leaders are on the hook for the same turn in years, not decades. Enter the restructuring charge.

Companies do it all the time. When the shape of the business no longer fits where it’s going, CFOs take the charge and rebuild. It’s common practice to pay for the layoffs, and the Street rewards it.

The same instrument that funds the demolition could fund the building. The books will bankroll subtraction at any scale. Addition has to fight for a line.

Without intentionality, addition loses the way good things do, never by decision, only by drift.

It goes one sensible cut at a time, each right on the books, until the night the plates come back and the food is just food, the same as anywhere, and no one at the table can say when it changed.

It doesn’t go that way everywhere. Somewhere a leader is reading the ledger the other way. She takes the charge the others spend tearing things out, and spends it on the build. Keeps the cooks. Cuts the menu to the few plates worth crossing the street for, and puts the whole house behind them.

Sets the line so it moves as one thing instead of ten.

None of it books as an asset. All of it is why people still make the trip to see you.

The oven, anyone can buy anytime. The kitchen is a hard, necessary thing to rebuild. That’s the restructure. Not the cutting. The choosing.

She’ll get the edge for it. But that isn’t what moves her.

She still knows what the table is for, and demands the food taste better.