The Verification Premium
Why AI conquered software first, and what the cost of checking its work predicts about every other profession
This is a companion to The Coordination Premium, which makes the broader case that value in the AI era migrates from producing cognition to governing it. This piece follows one strand of that argument, the cost of verification, into a prediction about the professions.
There is a puzzle in the pattern of AI adoption that the usual explanations do not solve.
The models are, by now, broadly capable across a wide range of knowledge work. They draft legal arguments, read medical images, build financial models, write marketing copy, and produce software, and on narrow benchmarks they often match a competent junior professional. Yet adoption is wildly uneven, and the unevenness does not track capability. By early 2026, coding tools had become one of the largest categories of enterprise AI spending. AI use was widespread in professional services too, with 74 per cent of surveyed professionals in law, tax, and audit reporting use several times a week. But use is not delegation. Reaching for a tool several times a week is not the same as handing it the work, and whole classes of consequential work remained stuck at the pilot or assistant stage. Within a single firm, one task gets automated in a quarter and the tasks beside it resist for years.
If capability were the constraint, adoption would follow capability. It does not. So capability alone is not the binding constraint.
This essay proposes what is. The claim is simple enough to state in a sentence, and it predicts a surprising amount: AI is adopted fastest not where models perform best, but where their output is cheapest to check. I called this the verification-cost hypothesis in the companion essay on the economics of the firm; here I want to follow it into the professions, because, once you have it, the uneven map of AI adoption stops looking random and starts looking almost lawful.
One refinement before we begin, because the single word “verification” hides more than it shows. Checking an output has two economically distinct dimensions, what it costs and how long it takes, and both interact with a third thing that is not a property of verification at all but of acting on a wrong answer: the cost of the error itself. A meme and a mortgage approval sit at opposite corners of this spectrum. The generated meme is trivial to check and costless to get wrong, so adoption is instant; the generated approval is expensive to check and expensive to get wrong, so adoption crawls. When these align, the prediction is easy. The interesting cases are where they diverge. Consider latency. Some outputs cannot be verified at the moment they are produced at any price, because the ground truth arrives only with time. A venture investment, a hiring decision, a five-year strategy: the model’s output is not just costly to check, it is unverifiable until the future reveals whether it was right. So the sharper statement is that adoption is governed by the cost and the latency of trustworthy verification, weighed against the cost of being wrong. Where the verdict is cheap and immediate, AI pours in. Where it is expensive, or deferred, or where the price of error is severe, the human stays.
The thing that did not get cheaper
Let’s start with what AI actually changed. It drove the cost of producing plausible cognitive output toward zero. A draft, a function, or an analysis: all now essentially free, and available in unlimited quantity.
Producing the work, though, is only the first half of getting value from it. The second half is trusting it. Before an organisation acts on a piece of work, someone has to establish that the work is correct, or correct enough for the stakes involved. That act of establishing trust is verification, and verification is a cost like any other. For most of economic history it was a modest one, because the same skilled human who produced a piece of work also certified it along the way. A competent analyst who builds a model knows where the bodies are buried in it. This is Michael Polanyi’s point that we know more than we can tell: much of the checking a professional does is tacit, folded invisibly into the act of production, never written down because it never had to be. Production and verification came bundled in the same person, and you paid for them together without ever seeing a separate line for the second.
AI unbundles them. It produces the work but does not certify it, and worse, it produces output optimised for plausibility rather than correctness. The result is fluent, confident, well-formatted, and sometimes wrong in ways that are invisible on the surface. The tacit certification that used to ride along free with skilled production is exactly what the model does not supply. So verification, which used to be bundled, becomes a separate and often expensive line item. The cost of generation collapsed. The cost of verification did not, and in some domains it rose, because catching a confident, plausible error takes more work than catching an obvious one.
Economics has a name for what has broken here. Michael Spence won a Nobel for showing how markets function when quality is hidden: they lean on a signal, some observable proxy that reliably separates the good from the bad because the bad cannot cheaply fake it. A degree signalled ability because it was costly to obtain; polished, fluent work signalled competence because sloppy workers could not easily produce it. The signal worked because it was expensive to counterfeit. AI destroys fluency as a reliable separating signal. When anyone can generate confident, well-structured output for nothing, polish stops separating the careful from the careless, and the equilibrium that rested on presentation collapses. Other signals endure, reputation, provenance, track record, credentials, but the cheap one, the surface quality that let managers and clients triage a document on sight, is gone, and the market is forced back onto the expensive thing it used to stand in for: direct verification of the substance. Workslop, the confident and useless AI output that clogs an inbox, is not merely annoying. It is a signalling failure with a balance-sheet cost, and the cost is the verification that now has to be done by hand.
This is why measuring AI by the output it produces is a category error. Output is not the product. Trusted output is the product, and the price of trust now governs everything.
Why software went first
Look at software with this lens and its head start stops being mysterious.
Code is not easy for AI because it is simple. Much of it is fiendishly complex. Code is unusually easy to deploy with AI because it comes wrapped in the deepest stack of cheap, automatic verification humans have ever built around any category of knowledge work. Consider what checks a piece of generated code before a human ever reads it. A compiler rejects what does not parse. A type system catches whole families of error. Linters flag suspect patterns. Unit and integration tests confirm behaviour against intent. Continuous integration runs the whole battery on every change. And the runtime supplies another observable verdict: whether the program behaves as expected under real conditions.
None of these layers is perfect, and none certifies that the software does what the business actually needed. That final judgment still waits for user-acceptance testing, integration, and the slow verdict of production. But a large share of errors is caught early, automatically, at almost no marginal cost, long before an expensive human enters the loop. Decades of tooling built an assembly line of graduated, mostly free checks. Software is the one domain where verification was industrialised before AI arrived.
Now hold that against a legal opinion, a strategy memo, or a clinical assessment, and the real structural difference comes into view. It is not that code has checks and law does not. It is that code has a graduated hierarchy of checks, each cheap, each catching a different class of error, arranged so that almost nothing reaches the expensive human until the cheap layers have done their work. An error in generated code meets the compiler in milliseconds, the type checker a moment later, the unit test on save, integration on commit, production monitoring on deploy, and the user only at the end. Six nested loops, ordered from cheapest and fastest to most expensive and slowest, so that the cost of catching an error rises only as the cheap catches are exhausted.
Most professions have no such hierarchy. They jump in a single step from AI output to expert review. There is no compiler for a legal argument, no unit test for a diagnosis, no integration suite for a strategy memo. The first check is the last check, and it is the most expensive one available: a qualified human reading the whole thing with judgment. Every error, from a trivial slip to a catastrophic misjudgment, is caught by that expert or not caught at all. The model may be perfectly able to produce the draft. But the cost of establishing that the draft is trustworthy is high, because there is no cheap layer beneath the expensive one to absorb the easy cases first. The effective cost of a trusted legal opinion stays high even as the cost of a plausible one falls to zero.
That hierarchy, not the mere presence of tests, is what matters. A profession’s readiness for AI is roughly the depth of the verification stack it can put between a generated output and the human who must answer for it. Software had the deepest stack in the economy before AI arrived. Most other knowledge professions begin with a far shallower one.
Same models. Same capability. Radically different verification structure, and therefore radically different adoption.
A sceptic will object that software went first for reasons that have nothing to do with verification. Code had version control, APIs, cloud infrastructure, and an open-source corpus the size of a small library for models to learn from; perhaps that, not the compiler, is the whole story. The objection is fair, and those factors surely helped. But they cannot explain the pattern that matters most, which is the split inside a single firm. The same company, with the same repositories and the same cloud, adopts AI eagerly for inner-loop coding and resists it for architecture and requirements. The same hospital, with the same imaging systems, deploys AI aggressively in lesion detection while keeping the accountable read with the radiologist. Infrastructure is shared across those tasks; it cannot account for why one is automated and the one beside it is not. Verification cost differs between them, and it tracks the split cleanly. When a variable explains the variation that your rival explanation holds constant, it is doing real work.
The pattern holds inside professions, not just between them
If the hypothesis is right, it should predict not only which professions adopt AI first but which tasks within a profession fall first. It does, and the professional-services data from 2026 shows it with unusual clarity.
Accounting and audit are the clearest case, and they are moving fastest. Industry analysis of professional-services adoption finds that audit and accounting firms see the fastest first-year return on AI of any category, because their work is highly structured, repeatable, and increasingly fixed-fee priced. That finding is the verification-cost hypothesis stated in the language of a trade report. Structured and repeatable is another way of saying cheap to check. And the task-level split within the profession confirms it: Thomson Reuters’ survey data shows AI use concentrated in tax research and return preparation, the top two use cases at 77 and 63 per cent, with advisory rising more slowly behind them. Reconciliation and routine preparation expose much of their error surface to structured checks: totals must tie, fields must reconcile, source documents can be matched, and exceptions can be surfaced automatically. The verification is close to automatic, so AI floods in. Whether an aggressive revenue recognition is defensible is a judgment an expert must certify, and that judgment resists.
Law shows the same fault line. The fastest adoption is in document review and due diligence, where the output is a set of flags a lawyer can spot-check against source documents and errors surface quickly; contract review and discovery are moving in tandem, and legal-tech spending has grown accordingly. But strategic advocacy, the judgment call about how to advise a client under real uncertainty, stays stubbornly human, because verifying it means an expert re-deriving the whole reasoning. Notably, the share of lawyers calling AI a major threat to the profession rose to 50 per cent in 2026, up from 36 per cent a year earlier.
Medicine offers the cleanest demonstration of all, because it is where the framework’s logic is written into regulation. In 2016 Geoffrey Hinton told an audience to stop training radiologists, on the view that deep learning would make them obsolete within five years. It is now 2026, radiology has more FDA-cleared AI algorithms than any other specialty by a wide margin, roughly three-quarters of every AI-enabled medical device the agency has authorised, and demand for radiologists has held up rather than collapsed. The prediction failed in the most instructive way possible. What the tools actually do is flag, triage, and measure: the AI finds the candidate lesion, prioritises the urgent scan, drafts the preliminary read. Then the radiologist verifies and signs. The verification step was never eliminated, because the cost of an undetected error is a missed cancer and the accountability for it must rest on a licensed human. Precision matters here, because in high-stakes domains verification is not purely epistemic. It bundles three things a professional does at once, judging whether the output is probably correct, judging whether it complies with institutional policy, and being the identifiable person who bears responsibility if it is wrong. AI can increasingly assist with the first. The second and third are why the human stays. AI took the cheaply-checkable pattern-detection and left the expensive, accountable verification where it had always sat. The reimbursement system confirms the logic: regulators have been reluctant to create separate payment codes for most of these tools because the radiologist is already paid to find those findings and remains the accountable reader. Pathology is following the identical path, the algorithm surfacing suspect cells for a pathologist to confirm. This is not AI replacing the professional. It is AI colonising the cheaply-verified layer and leaving the expensive one human, which is the whole thesis in a single specialty.
Even within software, the fractal holds. Adoption is deepest in the inner loop, where the compiler and the test suite stand guard, and thinnest at the outer loop of architecture and requirements, where verification once again means senior human judgment and nothing automatic substitutes for it. The frontier of cheap AI tracks the frontier of cheap verification at every scale you look: between industries, between professions, and between the tasks inside a single job.
One case looks like a counter-example, and it is where the second variable earns its place. Marketing copy, ideation, first-draft brainstorming: these have adopted AI as fast as coding, yet their output is not cheap to verify in any rigorous sense. There is no compiler for a tagline, no test suite for a campaign concept. Why the speed, then? Because the cost of being wrong is trivial. A weak tagline is discarded at a glance and costs nothing; you do not need cheap verification when you do not need verification at all. This is the error-cost variable doing its work, and it sharpens the framework rather than breaking it. Where the cost of error is negligible, AI floods in regardless of verification cost, because nothing has to be checked. Where the cost of error is high, verification cost becomes binding, and adoption waits on it. The two variables are not redundant. They carve the map into quadrants: cheap-to-check and cheap-to-be-wrong adopts instantly; expensive-to-check and expensive-to-be-wrong crawls; and the interesting professions live in the corners where the two diverge, which is why a radiologist and a copywriter, both handed capable models, meet them so differently.
What the hypothesis predicts
A framework earns its keep by making predictions that could be wrong. Here are five the hypothesis commits to.
The next professions to fall will not be the ones where models improve most, but the ones where someone builds a cheap verification layer where none existed, and where the cost of error is low enough to permit it. The binding constraint is not on the model side, which improves on its own and reaches everyone at similar cost. It is on the verification side, which is specific, buildable, and ownable. Whoever manufactures cheap verification for a domain opens that domain to automation, and captures the value of having done it. This reframes the entire enterprise-AI problem: stop asking which model is best and start asking whose checking is cheapest.
Regulated and licensed work will be the durable holdout, and not for the reason usually given. The standard story is that regulation protects these jobs by fiat. A deeper reason is that much regulated work has expensive, accountable verification built into its legal definition. A licensed professional signing off is not a bureaucratic formality; it is society insisting that certain outputs be verified by an accountable human, because the cost of an undetected error is intolerable. This is why psychiatry, fiduciary wealth management, statutory audit sign-off, and the practice of law survive: expensive verification is the point, not an accident. Some licensing regimes undoubtedly also protect incumbents; but at their most defensible, they institutionalise costly human verification where society has judged the cost of error intolerable. The professions proving most resistant to AI are not the hardest ones but the ones built on licensing, high-trust judgment, and accountable sign-off. A psychiatrist’s prescribing decision resists not because a model cannot suggest the prescription, but because the liability and the trust must rest on a named, accountable human.
High-trust and high-stakes relationships resist for the same structural reason. When the cost of being wrong is a ruptured relationship or a catastrophic outcome, the standard of verification rises and the required verifier is a human who can be held responsible. Complex enterprise sales, senior advisory work, and crisis medicine stay human even as the research and drafting beneath them are automated. AI can draft the proposal perfectly well. What it cannot do is get anyone to act on it, because no one will until a trusted human has put their name to it, and that act is the job.
Professions will hollow from the bottom, and this is the uncomfortable one. Where AI takes the cheaply-verified junior tasks, the entry-level tier compresses first, because that is where verification was cheapest and production most routine. The evidence is already arriving. Stanford’s Digital Economy Lab, tracking payroll records for millions of workers, found that 22-to-25-year-olds in the most AI-exposed occupations saw a relative employment decline of around 16 per cent after generative AI spread, while older workers in the same jobs held steady or grew, and the declines concentrated in occupations where AI automates rather than augments. That is the hollowing, measured: the adjustment falls on the young, in the tasks cheapest to check, exactly where the framework says it should.
Michael Kremer’s O-ring theory explains why this is worse than it looks, and why the survivors get paid more. Kremer modelled production as a chain of complementary tasks in which a single failure ruins the entire output, the way one faulty O-ring destroyed the Challenger. In such a process the value of every task depends on the reliability of the others, so the ability to avoid failure at a critical step commands a disproportionate return, and high-skill workers who can guarantee those steps get matched together and paid accordingly. Verification is the O-ring of AI-augmented work. Cheap, abundant generation does not help if one unverified step blows up the outcome, which is why a hundred-step agentic workflow that is 99 per cent reliable per step, assuming the errors are independent, succeeds only about a third of the time. The person who can guarantee the critical step, the senior who verifies, is the O-ring, and O-ring logic predicts their compensation rises even as the juniors beneath them are automated away.
But this is the trap. The rungs of the ladder, the years of doing verifiable grunt work under supervision, are among the first things AI absorbs, and that grunt work was how juniors became the seniors who can guarantee the critical step. The profession automates the training of the very verifiers its O-ring depends on. Brookings has gone as far as proposing a medical-residency model for white-collar work to rebuild the pipeline the automation is dissolving. The profession that automates its juniors most efficiently may find, a decade on, that it has stopped manufacturing seniors.
Which brings the last and sharpest prediction, the one most easily proven wrong: the next wave of large enterprise-AI companies will not be built on better models. The undifferentiated copilot wave is reaching its limits, and the models beneath it are commoditising. The companies that reach real scale from here will be the ones that build the verification layer for an industry that does not yet have one, the compiler-equivalent for law or credit or clinical work, and sell trust in a domain where trust was previously manual, expensive, and slow. If, five years from now, most of the new enterprise value created above the model layer still accrues to generic copilots rather than to domain-specific verification infrastructure, this essay was wrong. I do not think it will be.
Verification as an asset
The strategic turn in all this is that verification cost is not a fixed property of a domain. It can be lowered, deliberately, by building the missing checks, and whoever builds them owns something valuable.
The instrument for lowering it has a name: the evaluation suite, or eval. The word undersells it, because most people hear “eval” and think of a benchmark, a score for whether the model answered correctly. An enterprise eval asks a different and much harder question. Not “is this output correct?” but “would my firm have approved this decision?” Those are not the same thing. A credit model can be correct in the abstract and still violate your bank’s risk appetite; a contract can be legally sound and still breach your firm’s negotiating policy. The eval that matters encodes institutional judgment, not general accuracy, which is why it is manufactured verifiability rather than mere testing: it builds, for a domain that never had a compiler, an inner layer that checks output against what this organisation counts as good. It catches the cheap errors instantly and reserves scarce expert attention for the genuinely hard cases. It does not abolish the expensive outer layer of human judgment any more than unit tests abolished user-acceptance testing. It thins the traffic that reaches it, and by thinning it, drags the domain across the adoption frontier.
There is real contract theory underneath this. Bengt Holmstrom’s informativeness principle, another Nobel result, says that in any principal-agent relationship you should condition your judgment of the agent on every signal that carries information about whether they did the right thing, and on no signal that does not. An eval suite applies a similar logic in software. The AI is the agent, the firm is the principal, and the eval gathers the signals that are informative about whether the agent acted as intended and filters out those that are not, cheaply enough to act on. Seen this way, the enterprise problem of managing AI is not novel at all. It is the oldest problem in the economics of organisations, getting useful work from an agent whose output you cannot fully observe, and evals are simply the newest instrument for solving it.
This matters commercially, and it turns on a distinction the word “verification” tends to blur. There is cheap verification, and there is the infrastructure that makes verification systematic, repeatable, and proprietary, and the second is the real asset. Anyone can spot-check an output once. What compounds is owning the accumulated machinery, the eval suites, the encoded policies, the exception histories, the institutional memory of every past correction, that lets a firm trust a commodity model’s output at scale and at speed. Models are commoditising. Any capability a frontier lab offers you this quarter reaches your competitor soon after, on similar terms, and raw model improvements diffuse faster than almost any input in the economy. Verification infrastructure does not diffuse that way, because the highest-value verification is specific to the firm that built it. What counts as an acceptable credit decision at your bank, a safe adjustment on your production line, a compliant contract under your regulator and your risk committee: these are not general facts a model can supply. They are yours. The machinery that encodes them is a private asset a competitor cannot readily purchase or reproduce, and a model cannot infer from the open internet.
Which points at the durable strategy for anyone deploying AI in a serious domain. Do not try to own the intelligence; you cannot, and it is becoming a commodity. Own the verification. Build the machinery that lets you trust a commodity model’s output faster and more cheaply than your competitors can trust theirs. In a world of abundant, interchangeable cognition, the firm that can most cheaply establish which of that cognition to believe holds the advantage.
Rent the intelligence. Own the judgment.
The map redrawn
The verification-cost hypothesis will not survive contact with every case, and it was never meant to travel alone. Error cost rides beside it, as the meme and the mortgage showed at the outset; regulation, liability, data readiness, and workflow integration all pull on adoption too, and a careful account holds them apart rather than collapsing them into one variable. The honest compression is that adoption speed rises with model capability and falls as verification becomes more expensive or more delayed, above all where the consequences of error are severe. Where all of those are low, AI is already everywhere. Where they stack up together, as in medicine or credit or law’s advisory core, it will arrive slowly and under supervision, whatever the benchmarks say. It is a hypothesis, not a law, and it should be held as one.
But as a first-order map of where AI is going, it is hard to beat, because it explains what capability-based accounts cannot: why equally able models are adopted at wildly different rates across the knowledge economy. The answer is that we were watching the wrong cost. We saw the price of producing intelligence fall and assumed value would follow it down into every profession at the same pace. Value did not follow production. It followed verification, and verification did not fall evenly.
So when you want to know how fast AI will change a given job, do not start by asking how good the models are at it. They are probably better than you think, and improving on their own. Start with a different question: when this work is done, how expensive is it, and how long does it take, to know whether it was done right? The answer to that, more than any benchmark, tells you what happens next.
For two centuries the scarce input in knowledge work was the production of competent cognitive output. AI is making that output abundant. What stays scarce is trustworthy judgment, and the winners of the next era will not be those who generate the most cognition. They will be those who can prove which cognition deserves to be believed.