In February 2024, Air Canada lost a court case to its own chatbot.
The bot had invented a bereavement-fare discount that didn’t exist. A customer relied on it. The airline argued, with a straight face, that the chatbot was a separate legal entity responsible for its own words. The tribunal disagreed, and Air Canada paid.
Almost every public AI failure of the past three years has this same shape. A New York City government bot told employers they could take workers’ tips, which is illegal. A dealership bot was talked into “selling” a $76,000 truck for one dollar. Media outlets retracted dozens of AI-written articles. A major bank announced AI-driven layoffs, then reversed them when the union showed the workload hadn’t actually dropped.
Different industries, different vendors, different models. One common cause: in every case, nobody checked the output before it hit the real world.
The opposite is also true. The AI deployments that quietly work, such as coding assistants, claims processing, and customer-support tools, all share one trait: every output gets checked, automatically and cheaply, before it matters.
That single difference is the most important idea in the economics of AI right now. It has a name: verifiability.
What “verifiable” actually means
Some work is easy to check and hard to do. Some work is hard to check even when it’s easy to produce.
A completed Sudoku takes twenty minutes to solve and two seconds to check. A bank reconciliation either balances or it doesn’t. Code either passes the test suite or it doesn’t. An insurance claim either matches the policy terms or it doesn’t.
Now compare: a legal opinion. A treatment plan. A five-year strategy. A hiring decision. Producing a plausible version of any of these takes an AI seconds. Checking whether it’s actually right takes an expert, takes time, or, worse, takes years for reality to deliver the verdict.
AI researcher Jason Wei formalized this as Verifier’s Law: the easier a task is to verify, the faster AI will master it. Andrej Karpathy puts it even more simply: traditional software automated what we could specify; AI automates what we can verify.
Why? Because a cheap, reliable check does two jobs at once. It’s how the AI learns (train against the check, millions of times), and it’s how a business can trust it (catch errors before they become Air Canada moments). No check, no learning, no trust, no deployment.
So the question “what will AI automate?” becomes a different question: “what work can be cheaply checked?”
Let’s count.
The arithmetic, step by step
I spent the past months building this into a working model: nine synthetic firms, including a bank, an insurer, a hospital, a law firm, a manufacturer, and a trucking company, each rebuilt from its real cost structure and redesigned around AI agents, with every task admitted only if its output could be cheaply verified. The numbers below come from that model, and I’m happy to share it with anyone who wants to pressure-test them.
Take a 100-person operations department: claims processing, customer service, back-office finance, that kind of work. Walk through what happens when you rebuild it around AI agents whose every output is verified, with humans handling the exceptions.
Step 1: How much of the work can be checked cheaply? Go decision by decision through the workflows and about 60% of operational decisions can run inside verified loops, checked against policy rules, ledgers, records, or tests. The rest is judgment calls, relationships, exceptions, and anything with no clean way to confirm “correct.”
So: 100 units of work becomes 60 loopable.
Step 2: Can your systems actually support the check? This is the gate most analyses skip. A verified loop needs something to verify against: a trustworthy record. Try this test on any company. Pick something a department produced last week, and ask whether the firm can produce, within one day and without manual reconciliation, a machine-readable record proving it was done correctly. Most companies fail. Official statistics say only about 10% of US firms run AI in production at all, and the typical enterprise captures roughly 65% of its theoretical loop potential because its records simply aren’t clean enough.
60 × 0.65 gives 39.
Step 3: The exceptions come back. Automated flows don’t eliminate humans; they concentrate them. In real production systems, exception handling persistently eats 20 to 40% of the “automated” work. Call it roughly one unit in five returning.
39 × 0.78 gives about 30.
Step 4: Oversight isn’t free. Somebody writes the policies, audits the loops, and signs off. That governance layer costs 2 to 9% of transformed operating cost at launch, settling to roughly 1 to 5% at steady state. (Compute, the thing everyone worries about, runs 1 to 3% and never becomes the binding cost.) Net it out:
About 28 of the original 100 units of operational labor actually leaves.
Sixty percent coverage became twenty-eight percent displacement. That gap between coverage and displacement is where almost every AI forecast goes wrong, in both directions.
Step 5: Operations isn’t the whole company. A firm is also sales, clinical staff, engineers, drivers, relationships, physical work. One hospital illustrates it perfectly. Its billing and revenue-cycle department is about 56% loopable, the best score in the whole analysis. Its clinical care, half the payroll, is about 8% loopable, because a treatment decision has no cheap check before irreversible harm. Same building, opposite worlds.
Blend it all and the typical established firm lands around 1.24×, roughly 19% cheaper to run, against a measured reality today of 1.05 to 1.15×, because most firms are still early. Within the specific functions that get reorganized: 1.2 to 1.6×, which matches the best field studies we have.
What that means in dollars
Global labor compensation is about $58 trillion a year (global GDP of roughly $111T, labor’s share roughly 52%). The functions where verified loops apply this decade, meaning administrative, clerical, support, claims, software delivery, and the back offices of finance, insurance, and healthcare, cover roughly 20% of it: about $11.6 trillion in wages.
Where are we today? Be honest about the starting point. Official statistics show AI adoption is real but shallow. Most firms aren’t in production at all, adoption weighted by employment is around a third, and the majority of adopters use AI in three or fewer business functions, mostly to assist humans rather than to run verified loops. Multiply it through: $11.6T affected wages, times a 4 to 6% net gain rate, times the sliver of organizations actually capturing it, and the realized market right now is roughly $50 to 140 billion a year. Tens of billions. Not trillions.
Compare that with the spending. Industry forecasts put generative-AI spend at over $600 billion in 2025, but roughly 80% of it is hardware, with software a few tens of billions. In other words, the world is currently spending several times more building AI capacity than it is capturing in verified value from it. That gap is not a scandal. It’s a queue. Value shows up only when infrastructure gets converted into governed, verified loops, and most organizations haven’t started.
Now run the funnel forward through coverage, readiness, exceptions, and oversight, and phase it by how fast organizations actually become capable of capturing anything: roughly $120 to 240 billion a year by 2028, rising to $1 to 2 trillion a year by 2035.
A trillion-plus is enormous. But it’s worth being clear about what that number is not. It is not “AI does half of all work.” It’s a fifth of wages addressable, and a quarter-ish of that captured, after every honest deduction.
That is the map today. Now for the more important question.
Is the unverifiable 80% actually fixed?
Every number above assumes the wall between checkable and uncheckable work stays where it is. It won’t. “Unverifiable” isn’t one condition. Decompose any judgment-heavy domain and you find five different reasons work can’t be checked, with wildly different life expectancies:
Nobody records the outcome. The ground truth exists; there’s no sensor on it. Half-life: short. Cheap sensing is eating this layer alive.
Checking costs too much. A check exists but costs more than the work. Half-life: short to medium. Cost curves win.
The standard is unwritten. The rules exist only in experts’ heads. This one masquerades as “judgment” more than anything else, and history keeps unmasking it. Credit decisions were a banker’s intuition until scorecards decomposed them in the 1960s. Options pricing was trader feel until Black-Scholes. Aviation turned pilot judgment into checklists. Each time, judgment turned out to mean policy nobody had written down yet, and language models are unusually good at writing down unwritten rules.
The feedback is slow. Reality delivers the verdict in years (strategy, diets, careers). Erodible through proxies and simulation, but with a hard floor, described below.
Checking is impossible in principle. The genuine article. Rare.
Want proof the wall moves? Look at the domain everyone calls least verifiable: medicine. Closed-loop insulin delivery is in production right now, an autonomous system making dosing decisions in the highest-stakes category that exists, approved by regulators. What moved it across the wall wasn’t a smarter model. It was a sensor. Continuous glucose monitoring made the outcome observable faster than harm can accumulate.
That points to the general law hiding underneath all of this:
A decision becomes automatable when verification latency drops below harm latency.
The wall was never “medicine” or “law” or “judgment.” It’s a ratio: how fast you can observe the outcome versus how fast a mistake becomes irreversible. Every new sensor, every automated first-pass check, every written-down rubric attacks that ratio.
The future numbers
Price the drift and the map redraws. If verification technology keeps improving at anything like its recent rate, loop coverage of operational work climbs from about 60% toward 68 to 75% within a decade, adding $300 to 600 billion a year to the pool. And in the scenario where automated verification crosses reliability thresholds in even two major expert domains, moving legal review or medical first-reads from “expert-only” to “machine-checked with human sign-off,” the pool runs to $3 to 5 trillion a year, because that’s where the biggest wage pools sit, locked behind expensive verification.
The market is not a fixed number. It is a number with a derivative, and the derivative is verification technology.
What survives
Strip away the four eroding layers and a rock core remains: real, but far smaller than today’s “judgment premium” assumes. And as the contingent layer erodes, the judgment that survives doesn’t just survive. It concentrates and re-prices upward. Four things sit in the rock:
One-shot decisions with no rerun. You can’t A/B test the merger or replay the patient. Only protocols can be verified; the individual call cannot.
Bets that outlive the world. If the outcome arrives in seven years and the environment reshuffles in three, the ground truth expires in transit.
Questions of what’s worth wanting. Six good months, or eighteen medicated ones? No instrument can ever verify that answer, because there’s no fact out there to converge on, only a person’s own trade-off. Populations can be A/B tested; a person’s future self cannot. And that decision was never the doctor’s to automate in the first place. It’s the patient’s.
Independence. A system vouching for itself is a closed loop. The accounting profession has made exactly this argument about AI-assisted work, and it’s their home turf: an audit signature has value precisely because someone independent supplies it. Machines may one day do the checking, but the signature must always be someone else’s. AI can move the frontier. It cannot verify itself across it.
The question to ask about your own work
Not “is my work judgment?” That framing assumes judgment is one thing, permanently out of reach. It isn’t. Most of it is unwritten policy, missing sensors, and slow feedback wearing a trench coat.
The real question: which of the five reasons makes my work hard to check, and what is that reason’s half-life?
If your moat is an unwritten standard, assume it will be written. If it’s a missing sensor, assume it’s coming. If it’s the cost of checking, assume the cost curve wins. The durable ground is the rock: the one-shot calls, the long-horizon bets, the values conversations, and the independent signature. That’s the work that gets more valuable as everything around it becomes checkable.
Roughly two trillion dollars a year says the glacier is moving. Build on rock.
Methodology Appendix: How the Numbers Were Built
The model, the evidence, and the arithmetic behind “Unverifiability Has a Half-Life”
This appendix documents where every figure in the article comes from: the model that produced it, the evidence that disciplines it, the exact arithmetic, and the honest limits of both.
1. The approach in one paragraph
No organization has yet been rebuilt end to end around verified AI-agent loops with a published, audited before-and-after income statement. Every well-documented deployment bolts AI onto legacy processes. So the analysis combines two layers. An evidence layer draws on peer-reviewed field studies, tribunal rulings, official statistics, and documented rollbacks to establish what is measured. A mechanism layer, a simulation of nine synthetic firms, supplies what the evidence cannot reach: the cost structure of a firm actually reorganized around verified loops. Every claim carries a grade. Measured means it comes from Tier 1 or 2 evidence. Derived means it is a model output with evidence-bounded parameters. Assumed means exactly that, and is disclosed.
2. Definitions
A governed agentic loop is a unit of work in which AI agents execute tasks within explicit, human-authored policy; humans handle escalated exceptions and run an independent verification process that confirms outcomes against a system of record.
Work is admitted into a loop only if it passes the verifiability test: the result can be checked objectively, fast enough to catch failure before it becomes irreversible, at a cost below doing the work manually. No task in the model is classified as loop-run without naming how its result is checked and what the check costs.
3. The engine
Nine archetype firms span the economy: a property and casualty insurer, a contact-center outsourcer, a regional bank, a B2B software company, a law firm, an acute hospital, a discrete manufacturer, a general retailer, and a logistics carrier. Each is built in stages.
Stage 0. Construct the firm’s financial anatomy from public financial and occupational data: revenue, labor cost by function, non-labor operating cost, capital. Pass-through costs (claims payouts, cost of goods, purchased transport) are excluded from transformable cost.
Stage 1. Decompose each value-chain link into the outcomes it must produce and the decisions inside them, never into today’s process steps.
Stage 2. Apply the verifiability test to every decision. Each is classified loop-run, human judgment, or loop-running work (the new work that loops create). The loopable fraction of each function, lambda, is an output of this design, not an input.
Stage 3. An adversarial pass attacks every surviving loop from three directions. A Skeptic attacks verifiability: true exception rates, the cost of a missed error. A CFO attacks economics: compute, integration, governance, transition cost. An Operator attacks the seams: bursty exceptions, degraded modes, staffability. Loops that fail are killed and recorded. The pass is applied to every labor pool, including the governance function itself.
Stage 4. Re-derive the financials. Per function f:
L′f = (1 − λf) · Lf · αeff + λf · Lf · εf
αeff = 1 − (1 − αf) · κ
Cf = ρf · λf · Lf (compute)
Gf = [Gkernel + Gvar(t)] · λf · Lf (governance)
OpEx′ = Σf (L′f + Cf + Gf) + N′ + ΔK
Gain = OpEx / OpEx′
Where lambda is loop coverage, epsilon is the exception burden (work that comes back to humans), alpha is the assist factor on retained human work, kappa is maturity capture, rho is compute cost per unit of loop volume, and G is the governance layer with a permanent kernel plus a variable component that decays as policies stabilize.
Stage 5. Everything runs as a Monte Carlo over all parameters, with one critical design choice: a common-mode factor correlates the dominant uncertainty (whether cheap verification holds up across the firm at all) so it cannot be averaged away across functions.
4. Where each parameter comes from
Loop coverage, lambda (the 60%). Derived, then validated. The derivation was done by reasoning agents given only the verifiability test and the frozen firm anatomy, blind to all deployment evidence and to any target numbers, to prevent anchoring. Across the nine firms, the designed coverage of rules-based operational work clusters at 0.52 to 0.65, median 0.60, with a structural band of 0.48 to 0.68. The derivation was then scored against fourteen documented deployment outcomes (successes and rollbacks, from the Air Canada tribunal ruling to the McDonald’s drive-thru withdrawal). The verifiability test retrodicts twelve of fourteen in the right direction. This validates the ordering and structure of lambda, not its precise level, which is why the level carries a wide band.
Maturity capture, kappa (the 65%). A verified loop needs a trustworthy record to verify against. Maturity is graded by a concrete test: can the firm, within one business day and without manual reconciliation, produce a machine-readable record sufficient to verify last week’s output against explicit policy? Three grades follow. M1 (fails almost everywhere): kappa 0.25. M2 (core transactions covered, exceptions leave the system): kappa 0.65. M3 (complete, continuously verified record): kappa 0.92. The economy-wide mix, 35% M1, 50% M2, 15% M3, is calibrated to official statistics: roughly 10% of US firms using AI in production, about a third on an employment-weighted basis, and a majority of adopters using AI in three or fewer functions (US Census Bureau business trend surveys, 2025 to 2026). The article’s typical firm sits at M2, hence 0.65. This is the softest calibrated input in the model and it ranks first in every sensitivity analysis.
Exception burden, epsilon (the one in five). Bounded by field evidence. Production straight-through-processing programs in payments and claims persistently show exception handling consuming 20 to 40% of nominally automated flows. Review-inclusive cost analyses of frontier model output (GDPval, 2025) reduce naive time savings to 1.2 to 1.6 times once checking is priced. The model uses 0.12 to 0.30 by function; the article’s funnel uses 22%.
Assist factor, alpha. Bounded by the best field studies. A quasi-experimental rollout across 5,179 support agents measured 15% average productivity gains, concentrated at 34% among novices and near zero for the most experienced (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025). A randomized trial of 453 professionals found 40% faster writing on bounded tasks (Noy and Zhang, Science, 2023). A randomized trial of experienced developers on their own mature codebases measured a 19% slowdown against a predicted speedup (METR, 2025), which caps alpha for expert work. Assist is gated by kappa: a firm that cannot trust its records cannot reliably deploy copilots against them.
Governance (2 to 9% at launch, 1 to 5% at steady state). The governance function was put through the same verifiability test as everything else. Its monitoring majority is itself rules-based work against a record and partially loops. Its irreducible kernel, independent attestation and adversarial audit, cannot loop by construction: a system attesting to its own compliance is circular, a conclusion supported by peer-reviewed findings that AI judges exhibit self-preference bias, and by the EU AI Act’s legal requirement for human oversight of high-risk systems. Hence a cost that decays but never reaches zero, with a permanent kernel of 0.6 to 1.5% of transformed cost.
Compute (1 to 3%). Modeled generously per unit of decision volume and it still never binds, consistent with a measured price decline of more than 280-fold for constant-capability inference over roughly two years (Stanford AI Index, 2025).
5. The funnel arithmetic
The article’s 100-unit department walkthrough is the engine’s equation evaluated at typical (M2) maturity:
Coverage: 0.60
After maturity: 0.60 × 0.65 = 0.39
After exceptions: 0.39 × (1 - 0.22) = 0.30
After governance: 0.30 - 0.02 = 0.28
Coverage of 60% becomes net displacement of about 28% of operational labor. Blending all functions across each firm’s full anatomy, weighted by labor cost, gives the firm-level results: 1.15 to 1.39 times for typical firms by industry, 1.24 times economy-weighted, with a structural uncertainty band of 1.12 to 1.42. Immature (M1) states land at 1.05 to 1.15 times, which matches the measured present and is the source of that figure in the article.
The hospital example is a direct model output: its revenue-cycle function designs to lambda of roughly 0.56 (every output checks against payer rules and the medical record), while clinical care, about 52% of its labor, designs to roughly 0.08 (no cheap check before irreversible harm).
6. The market arithmetic
Wage base. Global GDP of roughly $111 trillion (IMF World Economic Outlook, 2025) times a labor income share of 52.3% (International Labour Organization, 2024) gives about $58 trillion in global labor compensation.
Addressable slice. The functions covered by the archetypes (administrative and clerical work alone is 12.2% of US employment per the Bureau of Labor Statistics, plus customer service, software delivery, and the administrative layers of finance, insurance, and healthcare) map to roughly 20% of the global wage pool: about $11.6 trillion.
Today’s realized value. $11.6T, times a mature net gain rate of 4 to 6% of addressed labor cost per year (the pool-weighted average of the archetype results), times current effective capture of 10 to 20% (from the adoption statistics above), gives roughly $50 to 140 billion per year.
The path. Capture is phased on the diffusion record of comparable technologies (cloud computing’s two-decade spend curve, e-commerce’s quarter-century to 16% of retail): 10 to 15% of the addressable pool captured within capable organizations at year three, 25 to 35% at year five, 55 to 70% at year ten. That yields roughly $120 to 240 billion per year by 2028 and $1 to 2 trillion per year by 2035. A second, fully independent method (2025 software and services spend of about $65 billion, compounded at 25 to 40% decaying to 15%, times a 2 to 4 times value multiplier) lands at $0.7 to 2.8 trillion, overlapping the first. The divergence between the methods is analyzed, not averaged.
The frontier scenario. If verification technology continues improving at its recent measured rate (agent task horizons doubling roughly every seven months, automated judges reaching human-level agreement rates), loop coverage of operational work drifts from 0.60 toward 0.68 to 0.75 over a decade, worth $300 to 600 billion per year. If automated verification crosses reliability thresholds in two or more expert-review domains, the pool runs to $3 to 5 trillion. These scenario figures are quarantined and never blended into the base case, because the drift rate is the most speculative quantity in the analysis.
Spend versus value. Industry forecasts put 2025 generative-AI spending above $600 billion, with roughly 80% in hardware and software in the tens of billions (Gartner, 2025). The comparison in the article divides realized value by that spend directly.
7. Evidence standards
Sources are tiered. Tier 1: peer-reviewed studies and randomized trials. Tier 2: tribunal rulings, regulatory texts, official statistics. Tier 3: document-supported journalism. Tier 4: named large-sample surveys. Tier 5: vendor and consultancy claims, directional only, never load-bearing. Field evidence is admitted into the model only as bounds on epsilon, alpha, governance, and the maturity mix. It is never allowed to inform the lambda derivation, which is where anchoring would corrupt the result; lambda is derived blind and validated afterward against the case record.
8. Uncertainty, honestly
Two bands are reported. The parametric band reflects the model’s chosen probability distributions and is authored precision: it can be made arbitrarily tight by choosing tight distributions, so it is disclosed but not trusted. The structural band adds the coarseness of the nine-archetype grid, alternative functional forms, and the calibration error of the maturity mix. The structural band is the honest one. For the base-case market figure it spans roughly 0.6 to 1.6 times the central estimate.
9. Limitations and how the analysis could be proven wrong
The case set that validates the mechanism is small and publicity-biased. The maturity mix rests on the best available official statistics, which measure adoption rather than record quality directly. The model prices states and a phased path between them, not the true speed of organizational change. Bolt-on field evidence can only lower-bound designed loops, and no audited, end-to-end reorganized firm yet exists to test the gap.
Specific observations would break specific claims. A production rules-based loop running far outside the 0.48 to 0.68 coverage band at twelve months breaks lambda. Closed-loop audit pass rates measured across a real sector diverging more than twofold from the assumed mix re-weights every market case. Audited multi-firm evidence of sustained organization-wide gains above 2 times within three years breaks the gain band upward; five years of median adopters capturing nothing breaks it downward. A regulator or insurer accepting fully machine attestation of loop compliance breaks the governance kernel. Official adoption statistics still below 12% of firms in production in 2028 mean the market path is too optimistic; verified positive returns on more than $500 billion of annual software and services spend by 2028 mean it is too conservative.
The full model, the per-firm parameter files with their evidence bounds, the blind derivations, and the case-set scoring exist as runnable artifacts and are available on request.