Your Job Is 25 Tasks. New Data Prices Each One Separately.

A software engineer who advises colleagues on which technology to adopt is paid more than one who diagnoses why the system went down. Same job title. Same employer. Different pay.

That gap has now been measured. ADP Research and the Stanford Digital Economy Lab linked more than 5 million job postings from ADP client employers to more than 9 million workers in payroll records, then asked something labor statistics almost never answer: what is each individual task inside a job actually worth? [Fact]

The answer breaks one of the most repeated pieces of AI-era career advice.

What the study actually did

The work was published on July 29, 2026, by Nela Richardson of ADP Research and Andrew Wang of the Stanford Digital Economy Lab. [Fact] The analytic sample narrows sharply from that headline linkage: roughly 7,000 workers in selected IT jobs between 2019 and 2025, producing more than 20,000 worker-year observations across more than 600 employers. [Fact]

Each posting was mapped to O*NET Intermediate Work Activity definitions — 25 activities in all — and worker wages were regressed on which of those activities the job actually contained, controlling for age, gender, year-specific effects, and company-specific characteristics. [Fact] Those controls matter. Without them you would simply rediscover that senior people at rich firms earn more.

Why the design matters more than the result

Nearly every widely cited AI-and-jobs number traces back to one method: experts read a task description and rate how automatable it looks. Frey and Osborne did it in 2013, and most successors refined that approach rather than replaced it. The weakness is structural. It measures what informed people believe a technology can do, and beliefs about AI have swung violently since 2022.

Linked postings-and-payroll data does something else entirely. It observes what employers actually paid when a job contained a given task, before and after the technology arrived. Nobody is asked to forecast anything. If the price of monitoring systems fell, it fell in the payroll file whether or not anyone predicted it.

That is why a study of 7,000 IT workers deserves attention out of proportion to its sample size. It is a different kind of evidence, not more of the same kind. [Claim]

Which tasks carry the wage

Of the 25 activities examined, 8 were associated with higher wages and 6 with lower wages. [Fact]

At the top: advising others on the design or use of technologies, and designing databases. At the bottom: diagnosing system or equipment problems, and monitoring the operation of computer or information technologies. [Fact]

Put those two lists side by side and a pattern appears that has nothing to do with difficulty. Diagnosing a failure is hard. It is also reactive, bounded, and verifiable — someone tells you the thing is broken and you find out why. Advising on architecture is open-ended and carries consequences that land months later. The market is paying for judgment under uncertainty, not for skill.

Which is precisely the boundary that current AI systems sit on.

The finding that breaks the standard advice

Here is the part worth stopping on.

Five tasks showed reduced compensation in 2023-2025 compared with 2019-2022 — the before-and-after of widely available generative AI. [Fact] One of them is "develop models of systems, processes, or products," a task that still ranks among the higher-value activities overall. [Fact]

The standard advice of the past three years has been to move up from execution toward design, modeling, and architecture, because that is where AI cannot follow. This data says part of that advice is already dated. Modeling work still pays a premium. The premium is shrinking.

Run the arithmetic on the counts and it gets sharper. Of the 25 activities, 14 produced a clear wage signal in one direction or the other, which means 11 — nearly half — showed no distinguishable wage difference at all. [Estimate — our own derivation from the counts reported.] Against that backdrop, five tasks moving downward together after 2022 is not obviously noise.

What this release does not give you

I have to be blunt about a limitation, because it changes how much weight everything above can carry.

The published write-up reports which tasks are associated with higher or lower wages. It does not report by how much. [Fact] There are no percentage premiums, no dollar figures, no standard errors, and no discussion of statistical significance in the public version. A task ranked "higher value" might carry a 15% wage premium or a 1.5% one, and from this release you cannot tell which. [Claim]

The scope is also narrower than the headline linkage suggests. Roughly 7,000 IT workers is a respectable sample for IT and tells you nothing about nursing, logistics, or teaching. And it runs through one payroll provider's client base, which is not a random draw of American employers.

Treat this as a direction, not a measurement. The direction is well identified. The magnitudes are unpublished.

How to use this if you work in IT

The practical implication is narrower than "learn architecture."

Take your own week and sort it into the two lists. How many hours go to monitoring systems and diagnosing failures — the tasks this data ties to lower wages? How many go to advising other people on decisions they will be held to, or designing structures other people will build on?

If your ratio leans heavily toward the first list, the risk is not that your job disappears. It is that your job stays and its price falls. That is the quieter form of AI displacement, and it does not appear in unemployment statistics at all.

The same logic reframes the argument about AI replacing developers. If tasks are priced separately and AI compresses the price of some of them, an occupation can hold its headcount while losing its wage. ADP's stated conclusion is that "AI's biggest and most immediate effect will be on tasks, not entire occupations." [Fact] That is a claim about pay structure at least as much as about employment.

Our task-level exposure and automation figures for the roles this study covers are here: software developers, computer systems analysts, database administrators, and computer network support specialists.

One last thing worth noticing. The task sitting highest in this data — advising others on the design or use of technologies — is the one that requires another human being to accept your advice. That requirement is doing a great deal of work.

The eleven tasks nobody prices

The overlooked line in this data is the middle of the distribution.

Eleven of the 25 activities produced no distinguishable wage signal in either direction, which means the market currently does not pay differently for a job that contains them versus one that does not. [Estimate — derived from the reported counts.] That is where I would look next, and it is the part of the study nobody will write about.

An unpriced task is one whose value is invisible to the wage-setting process. Invisible value is the easiest kind to automate away without anyone recording a loss. The tasks people argue about — coding, modeling, advising — carry a price and therefore a constituency willing to defend them. The other eleven have neither. If AI takes those first, the payroll data will show nothing at all until the work that depended on them starts failing.

Sources

• ADP Research and Stanford Digital Economy Lab (Nela Richardson, Andrew Wang), "Unbundling jobs: Measuring the value of tasks in an AI economy," published July 29, 2026 — https://www.adpresearch.com/research/unbundling-jobs-measuring-the-value-of-tasks-in-an-ai-economy

About this analysis

This article was written with AI assistance. Task lists, sample sizes, and the count of tasks that declined in value are taken directly from the published ADP write-up. The 14-of-25 and 11-of-25 breakdown is our own arithmetic on those reported counts and is labeled [Estimate]. The critique of the missing effect sizes, and the career interpretation, are ours and not ADP's.