First on the menu.

AI does not arrive where it is most capable. It arrives where its answers can be checked. Mathematics and programming were among the first fields that AI changed, and it is tempting to put that down to raw difficulty: models got clever enough for proofs and code before they got clever enough for anything else. That gets the order wrong. A proof either goes through or it does not. Code either passes its tests or it fails. A set of accounts reconciles or it does not. A memo on strategy, a hiring decision or a piece of advice to a client has no such test. Whether it was good may only become clear years later, and people may still disagree.

This essay argues that the difference matters more than it seems. Before anyone lets an AI do a piece of work, someone has to be able to tell whether the work is right. Where that is cheap, AI moves in early. Where it is expensive, AI waits, even when it is perfectly capable of producing the output. I took the idea to data on how a widely used AI assistant is actually used across eighteen thousand work tasks. The pattern is there, though more modest than the story suggests.

Five reasons a field gets there first

Ask why AI reached some kinds of work before others and five forces come up. The first is verifiability: whether an answer can be checked cheaply against a clear criterion. The second is whether the work is born digital. Text, code and spreadsheets are already in the form models read and write; a nurse's or an electrician's output is not, and there robotics is the constraint. The third is the feedback horizon: a task that resolves in seconds can be corrected in seconds, while one whose outcome appears in five years, like a strategy or a hire, cannot. The fourth is context: much of the value of some work lies in firm-specific data, relationships and unwritten rules that no model has seen. The fifth is accountability: where someone must sign, such as the auditor, the doctor or the lawyer, capability can arrive long before the human leaves the loop.

These are not independent. A long horizon is one reason an output is hard to check. Missing context is another. Accountability is what society does when checking is hard and errors are costly. Verifiability is the thread that runs through most of them, which is why it is the one I tried to measure.

Checking is part of installing AI

In an earlier research note I argued that the scarce input in an economy of cheap AI is not the model but the capacity to put it to work. Before an agent can do a task inside a firm, someone has to connect it to the firm's systems, redesign the workflow and, crucially, evaluate its output. That work is a fixed cost per task, done by skilled people. The model's prediction was that firms automate tasks not in order of how good the AI is at them, but in order of how much they save relative to what it costs to get the AI working on them.

Evaluation is where verifiability enters. If the output of a task can be checked mechanically, building the evaluation is cheap: write the tests, compare against the source, check the totals. If it can only be judged by an expert, every output needs an expert's time, and the saving from automating it shrinks. If it cannot be judged at all, the firm is taking a bet. So among tasks the AI could do equally well, the ones that are cheap to check should go first.

What the data show

To test this I needed two things that do not exist off the shelf. The first is a measure of how AI is actually used, task by task. The Anthropic Economic Index provides one: it maps conversations with Claude to the roughly eighteen thousand task statements in O*NET, the US Labor Department's catalogue of what each occupation does. The second is a measure of how cheaply each task's output can be checked. For that I rated every one of O*NET's 332 work activities on a five-point scale, from "no ground truth" (negotiating, counselling, setting an organisation's goals) to "checked mechanically" (processing invoices, reconciling accounts, writing code that must pass tests), and gave each task the rating of the activities it involves.

The raw pattern is in Figure 1. But on its own it proves little. Claude's users include many programmers and analysts, whose tasks are both verifiable and heavily exposed to language models, so the gradient could simply reflect who uses the tool. The cleaner comparison is within occupations: tasks done by the same people, in the same jobs, which differ in how cheaply their output can be checked. I also hold fixed a standard measure of how much a language model could technically speed up each task, so that verifiability is not just picking up capability.

Within the same occupation and at the same technical exposure, each step up the scale raises the probability that a task shows AI use by about two percentage points. Against an average of 15%, moving from the bottom of the scale to the top raises it by roughly half. The estimate barely moves when computer and mathematical occupations are dropped, or when the exposure measure is swapped for one rated by people rather than by a model. Exposure itself is a strong predictor, as it should be. What matters is that it does not absorb verifiability: being able to do a task and being trusted with it are different things.

The last row of Figure 2 is the honest caveat. My ratings vary by work activity, not by task, and many tasks share an activity. Allowing for that, the interval widens to include zero, just. The direction is clear and the size is meaningful, but this is a first look, not a settled result.

One of the five forces did not survive contact with the data. The feedback horizon, rated separately, adds nothing once verifiability is accounted for. The two are strongly correlated, and the data cannot tell them apart. In practice, a long horizon seems to matter mainly as one of the reasons an output is hard to check.

The frontier goes where the checks are

There is a second reason verifiable work goes first, and it may be the stronger one. AI developers improve their models by training them against signals of success. Where an answer can be checked automatically, they can generate those signals in unlimited quantity: millions of problems, each graded by a program. That is why progress in mathematics and code has been so fast. Developers also aim where the market is largest, and programming is both a large wage bill and the work that speeds up AI research itself.

Economists have a name for this: directed technical change. Innovation flows to where it pays. Applied to AI, it means the frontier of capability is not a tide rising evenly over all work. It is pulled towards tasks that are verifiable, digital and valuable. My data cannot separate this channel from the one about installation costs, and both point the same way: verifiability governs the order.

It also implies something hopeful for fields that feel left behind, and something worrying for fields that feel safe. Verifiability is not fixed. Fields climb the queue by building ways to check work. Mathematicians are formalising proofs in languages that a computer can verify. Software teams write test suites and evaluations before they write code. Engineers build simulators. Empirical economists publish replication packages that anyone can rerun. Each of these was built for its own reasons, and each makes the field easier for AI to enter. If you want to know which profession will change next, watch where the investment in verification is going.

I tried to see this in the data, comparing usage in early 2025 with usage in spring 2026. The share of AI use going to highly verifiable tasks fell, as the theory says it should once the easy ground is taken. But almost all of the change comes from programming leaving the chat product for dedicated coding tools that the later data do not cover. Outside programming, nothing moved. The test will have to wait for better data.

What follows

For firms, the lesson is to build the check before the agent. A task that cannot be evaluated cannot be safely delegated, to a machine or to anyone else. Much of the work of adopting AI is writing down what a good output looks like, and the firms that do it first will be able to automate first.

For workers, the map is uncomfortable in one place and useful in another. Work whose output is easy to check, and done on a screen, is where AI arrives first; holding a routine position there is the exposed spot. But the people who make a domain checkable, who write the tests, build the evaluations and define what correct means, hold one of the most valuable positions there is, because they decide when everyone else's work can be handed over.

For researchers, the next step is clear. Rate verifiability task by task, with several independent raters, and follow usage over time across products. If the order of adoption really runs from the checkable to the uncheckable, it should be visible in every dataset that tracks AI use, not just this one.

The usual question about AI and work is what it can do. The more useful one is what we can check. Capability tells you what is possible. Verifiability tells you what happens first.