AI Isn't Devaluing Your Job. It's Devaluing Five Specific Tasks.
AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)
On 29 July 2026, ADP Research and the Stanford Digital Economy Lab publisheda study measuring the value of tasks rather than the value of jobs. It is a small piece of work making a large methodological point, and the point is more useful than the usual occupation-exposure ranking.
The finding, in one line: within a sample of IT jobs, five specific tasks were paid less well after 2022 than before it, while advisory and directional tasks held their value. Not the job. Five tasks inside the job.
A job title is a bundle. AI is not arriving at the bundle evenly, which is why an occupation-level risk score averages across people whose actual exposure has almost nothing in common.
whatsmyedge, August 2026What was measured
ADP Research states the construction of the sample plainly, so we will reproduce it exactly. Starting from more than 5 million job postings by ADP client employers, linked to more than 9 million workers in ADP payroll data, the researchers identified a sample of approximately 7,000 workers in selected IT jobs during the years 2019 to 2025. That sample represents more than 20,000 worker-year observations drawn from more than 600 employers.
They then read the task descriptions out of those job postings, mapped them onto O*NET Intermediate Work Activity definitions to produce a set of 25 activities associated with IT jobs, and looked at how worker pay moved with the presence of each activity. The analysis controls for age, gender and company differences. The authors are Nela Richardson and Andrew Wang.
Note the shape of that funnel, because it is the single most important thing about this study and it is the thing most coverage has dropped. Nine million workers is the pool the sample was drawn from. Roughly 7,000 workers in selected IT jobs is what was actually analysed.
The five tasks that lost value
Comparing pay in 2023 to 2025 against pay in 2019 to 2022, these activities were associated with lower pay in the later window:
| Task that declined in value | What it looks like in a working week |
|---|---|
| Diagnose system or equipment problems | Triage, fault-finding, working out why something broke |
| Develop models of systems, processes or products | Producing the representation rather than deciding what it should be |
| Document technical designs, procedures or activities | Writing it up after the decision is made |
| Set up computer systems, networks or other information systems | Configuration and provisioning against a known pattern |
| Explain technical details of products or services | Translating something already settled for another audience |
Against those, the study reports that certain tasks retained or gained value: advising on technology design and use, designing databases, directing technical activities, and developing technical specifications.
There is an obvious pattern and it is worth naming carefully, because it is easy to overstate. The declining set is largely work that turns a settled decision into an artefact — a diagnosis, a model, a document, a configuration, an explanation. The holding set is largely work that makes or directs the decision. ADP Research's own summary describes high-value work as clustering around advisory and directive activity, with execution and monitoring associated with lower pay.
That is a real distinction, and it is also the distinction that every consultancy report has been asserting without evidence for two years. What is new here is that somebody attached payroll numbers to it instead of a survey.
What this does not show
Four limits, all of which the study is open about and none of which survive contact with a headline.
It is IT jobs. Roughly 7,000 workers in selected information technology roles. Nothing here licenses a claim about nursing, logistics, law or teaching. The method transfers; the findings do not.
It is not causal. The comparison is before and after 2022, which is when AI tools became widely available — and also a period containing a substantial technology hiring slowdown, rate rises and a correction in technology pay generally. A before-and-after split around a date does not isolate the cause. ADP Research frames the work as an early proof of concept, which is the right register.
It measures pay attached to tasks, not jobs disappearing. A task being worth less is not the same as a task being automated, and neither is the same as a person being made redundant. Those three get collapsed constantly.
ADP is not a neutral party. ADP Research is the in-house research arm of a payroll and HR software company with a commercial interest in being read as the authoritative source on labour data. The Stanford Digital Economy Lab co-authorship is a meaningful check on that, and the methodology disclosure is better than most. It is still worth knowing who published it.
A sourcing note, because this one is checkable
This study was picked up quickly, and the version that travelled is not quite the version that was published. A trade write-up on 31 July described the research as analysingpayroll data from 26 million workers. ADP Research's own page states more than 9 million payroll-linked workers, with approximately 7,000 analysed.
We are not interested in scoring a point off one outlet, and we have no view on how the discrepancy arose. We raise it because both pages are public, the check took under a minute, and the direction of the error is the one that always travels: the sample got bigger on the way to the reader, never smaller. If you are about to quote a number from this study in a strategy document, open the ADP page and read the sentence yourself. That habit is worth more than any single finding here.
What to do with this
If you are an employee: audit your own week against the two lists rather than against your job title. The useful question is not whether "IT" is exposed, it is what proportion of your hours goes to producing artefacts from decisions somebody else made, versus making or directing those decisions. If that ratio is heavily weighted to the first, the evidence here suggests the market is repricing your week even if your title is untouched. Shifting it is gradual work, and it does not require leaving your field.
If you manage people: the finding cuts against how most teams are structured. Documentation, configuration and first-line diagnosis are exactly the tasks routinely assigned to junior staff as a training ground. If those tasks are being revalued downward, the traditional apprenticeship path is being narrowed from the bottom, and the people affected first are the ones with the least ability to see it happening. That is a management problem before it is a technology one.
If you are building a workforce plan on this: do not. One study, one sector, roughly 7,000 workers, no causal identification. What it earns is a change in the unit you think in — tasks rather than titles — which costs nothing and is almost certainly right. What it does not earn is a headcount decision. Wait for the method to be run on a second sector by somebody who does not sell payroll software.
Frequently asked questions
Which job tasks are losing value to AI?
In the ADP Research and Stanford Digital Economy Lab study of IT jobs, five tasks were paid less well in 2023 to 2025 than in 2019 to 2022: diagnosing system or equipment problems; developing models of systems, processes or products; documenting technical designs, procedures or activities; setting up computer systems, networks or other information systems; and explaining technical details of products or services. These are findings about a sample of roughly 7,000 IT workers, not about every occupation.
How big was the ADP and Stanford study?
Smaller than most coverage suggests. ADP Research states it began with more than 5 million job postings linked to more than 9 million workers in ADP payroll data, then identified a sample of approximately 7,000 workers in selected IT jobs between 2019 and 2025. That sample represents more than 20,000 worker-year observations drawn from more than 600 employers. The roughly 7,000 figure is the one that bounds the findings.
Why measure tasks instead of jobs?
Because a job title is a bundle of activities, and AI does not arrive at the bundle evenly. Two people with the same title can spend their weeks on completely different work, so occupation-level exposure scores average across people whose real risk differs. Measuring the pay attached to individual tasks shows which parts of a role are being revalued, which is a more actionable unit than a job-title risk score.
Does this prove AI caused the change?
No, and the study does not claim it. It compares pay attached to tasks before and after 2022, controlling for age, gender and company differences. A before-and-after comparison around a date when AI tools became widely available is suggestive, not causal. Other things changed in the technology labour market over the same period, including a broad hiring slowdown. ADP Research describes the work as an early proof of concept.
What should I do if my work is mostly on that list?
Treat it as a signal about task mix rather than about your job. The tasks that held or gained value in the same sample were advisory and directional: advising on technology design and use, designing databases, directing technical activities, and developing technical specifications. The practical move is to shift the share of your week spent on judgement and direction relative to execution and documentation, which is a gradual reallocation rather than a career change.