Ministry of AI · Dispatch from 2047
Measuring AI Economic Value: We Publish the Error
Editorial note. The Ministry of AI is a work of disciplined foresight: it describes the year 2047 in the present tense, and treats our own era as history. The institutions are imagined. The economics, the evidence and the historical parallels are real and sourced.
Every quarter, the Ministry publishes a number it knows is wrong.
It appears in the second column of the Machine Yield Account statement, immediately to the right of the estimate, and it is wider than most people expect. In professional services it is embarrassingly wide. My youngest son, who is twelve and has never known an institution that pretends otherwise, assumes this is simply how numbers are written down: the value, and then how much of the value is guesswork. He was surprised, when I told him, that the agencies of my working life mostly published the first column alone.
I helped build the systems that made that column hard to fill. In the 2020s I designed measurement inside companies — attribution models, incrementality tests, dashboards that told executives what their spending had bought. My job was to turn absences into numbers. When we switched on the first coordinated model systems, I watched a familiar thing happen: the work continued, the output rose, and the line where a person used to sit produced no signal at all. There was nothing to measure because nothing had been billed. A cost that is never incurred leaves no invoice.
That is the whole difficulty, and it deserves to be stated plainly before any institutional claim rests on it.
The value that shows up as an absence
Wages were the great accidental measuring instrument of the industrial era. If a firm wanted judgment, it bought judgment, and the price appeared in the accounts of both parties. Payroll was a receipt for value moving from a company to a person. When the judgment moved into a model, the value did not disappear; the receipt did.
Economists saw this coming and could not resolve it. The best statement of the problem was written in 2017, when capability claims were already loud and measured productivity growth was stubbornly weak: the two facts could be reconciled by implementation lags, by mismeasurement, or by both, and the paper that named the clash of expectations and statistics refused to pretend the data settled it. Ten years later, careful field studies were still finding results far smaller than the rhetoric — one of the most rigorous, a 2025 study of early generative AI adoption across thousands of workplaces, reported earnings and hours effects statistically indistinguishable from zero. That was not evidence that nothing was happening. It was evidence that the instruments were pointed at the wrong quantity.
Meanwhile the estimates of exposure — how much of a given occupation a model could touch — arrived as ranges, not points. The 2023 exposure study that became the reference reported bands depending on whether you counted a model alone or a model plus the software built around it, and the two numbers were far apart. Statistical agencies were candid about the same gap from the other direction: national accounts were built for goods and priced transactions, and the dedicated digital economy accounts existed precisely because the standard measures were missing things. A central-bank assessment of AI’s likely labour-market effect laid out the mechanism cleanly — task automation converts labour cost into margin — while conceding the magnitude was unknown.
So the honest position, in my century, was this: everyone could see the boundary moving and nobody could price the movement. What we did with that admission was the mistake. We waited. Precision was treated as a precondition for action, which meant the transfer of an inheritance was postponed until the accounting caught up, and the accounting never does.
The instrument the Ministry chose to build
The Machine Yield Account measures produced value at the point of deployment — the work done by autonomous systems, in the place where it is done, not the profit that survives at the end of a corporate structure. That choice was deliberate and costly. Profit is a residue: it can be shifted, deferred, capitalised, relocated. Yield is noisy, but it sits where the humans and the machines actually are, and it cannot be booked in a jurisdiction where nothing runs.
The Ministry took the harder number and paid for it with uncertainty. Three rules make that affordable.
The band, not the point. Every quarterly yield estimate is published with an interval, per sector. Nobody is asked to believe a single figure.
The conservative floor. Where the interval is genuinely wide, the assessment is struck at the low end. Uncertainty resolves against the collecting institution, never against the firm. A company is never charged on a number the meter cannot defend in a hearing.
The revision register. When better data arrives — and it always does, one to three years late — prior quarters are restated in public, with the reason attached. Under-collection is recovered forward, spread across future schedules. It is never clawed back from households that have already spent a dividend.
| Approach | What is measured | Error behaviour | Failure mode |
|---|---|---|---|
| Corporate profit tax | Residual income after accounting choices | Appears precise; error hidden in structuring | Base moves offshore; nothing is captured |
| Headcount-based levy | Jobs formally eliminated | Looks exact; misses absorbed capacity | Firms restructure without layoffs and pay nothing |
| Voluntary firm disclosure | What the firm chooses to report | Unquantified, unaudited | Reports arrive optimistic and unfalsifiable |
| Machine Yield Account | Output of autonomous systems at deployment | Published band, floor-assessed, revised openly | Wide bands in judgment-heavy sectors; systematic under-collection |
Read that last cell again, because it is the confession this dispatch is built on. The design under-collects. It was built to under-collect. Every quarter the Ministry leaves money on the table in exchange for the right to be believed the following quarter.
Why the admission was the load-bearing part
The temptation, in year one, was to publish a clean figure. Clean figures win a news cycle. But a body that claims precision has exactly one revision before its opponents own the story permanently, and there was never a version of this instrument that would not need revising.
This was not a novel insight; it was borrowed. Forecasting institutions had published their own accuracy records for decades — error statistics for professional forecasts were open, tabulated and unflattering long before anyone metered a model — and the forecasts survived precisely because nobody had promised more than they delivered. The Ministry copied the posture wholesale. The first statement carried the band on the front page, above the estimate rather than in a footnote, and the second column has never been removed.
I watched the effect on my own family. My mother, nineteen years an insurance claims adjuster, spent her career being told by systems that a number was final. When her role was absorbed in the thirties by a model trained on her own decisions, the first thing she read about the Ministry was not the size of the dividend but the size of the doubt. She trusted it because it flinched. She was fifty-nine when the first payment cleared. She is seventy-three now, and on Tuesdays she sits, unpaid, with families arguing with the health service — reading their files the way she used to read claims, on the side of the person this time.
That is what the second column bought. Not accuracy. Standing.
What it still gets wrong
Three failures are worth naming, since a dispatch that only admires its own institution is advertising.
The bands are widest exactly where the inheritance claim is strongest. Professional judgment — the compressed reasoning of millions of documented human decisions — is the hardest yield to price, so the sectors most obviously built on our collective record are the ones assessed most conservatively. The moral claim and the measurement precision run in opposite directions.
The floor rule is quietly regressive across borders. Yield is metered where deployment is booked, and the countries with the thinnest statistical capacity get the widest bands and therefore the lowest assessments. Under-collection is not distributed evenly; it concentrates where measurement is poorest, which is where the dividend was needed most.
And the revision register has never fully answered a simple question my eldest asked me last year, at twenty, with the flat scepticism of someone who did not live through the transition: if you know the number is low, why not simply raise it? Because the moment the Ministry sets an assessment it cannot defend line by line, it becomes a taker rather than a meter, and the argument collapses back into politics. Capacity is not contribution, and work that must be done to survive is not chosen work — but a claim without a defensible meter is only a sentiment with a bank account.
So we publish the error. Every quarter, in the second column, wider than anyone would like. It is the least impressive number the Ministry produces and the reason the rest of them are still believed.
FAQ
What does measuring AI economic value actually mean? Estimating the output an autonomous system produces that a paid human would otherwise have produced, and what that output is worth once booked. It is measured at deployment — the only place the human-to-machine boundary is visible — not at the level of corporate profit or vendor spending.
Why is it so hard? Because the value appears as an absence. A cost never incurred leaves no invoice, and quality shifts at the same time as volume. The productivity debates of the 2020s were, in large part, an argument about instruments rather than about technology.
What is the conservative-floor rule? Where the estimate band is wide, the assessment is struck at its low end. Uncertainty resolves against the Ministry, never against the firm. The predictable consequence is systematic under-collection, accepted as the price of enforceability.
What is the revision register? A public restatement of prior quarters when better data arrives, with reasons attached. Shortfalls are recovered forward through future schedules; nothing is ever reclaimed from households that have already received a dividend.
Why not tax profits instead? Profit is a residue shaped by accounting choices and easily relocated. Yield is noisier but sits where the work happens. The Ministry chose the harder number deliberately, and pays for it in error bars.
What is the biggest unsolved problem? That the bands are widest in judgment-heavy sectors and in economies with the weakest statistical capacity — precisely where the inheritance is largest and the need greatest. The meter is most conservative where it should be most confident.