Ask whether AI has devalued developers and you get two camps. One says the job is ending. The other says nothing has changed. Both are answering the wrong question. Payroll data now shows AI devaluing specific tasks inside IT jobs, not whole jobs. Routine execution tasks, like setting up systems and documenting designs, lost wage value. Judgment tasks, like advising on technology design, designing databases, and directing technical work, are the ones tied to higher pay.
On July 29, 2026, ADP Research and the Stanford Digital Economy Lab published a task-level wage study of roughly 7,000 workers in IT jobs at more than 600 employers. It compares pay in 2019 through 2022 with pay in 2023 through 2025. Several tasks lost value between the two periods: diagnosing system problems, setting up computer systems and networks, documenting technical designs. The tasks tied to higher pay include advising others on the design and use of technology, designing databases, and directing technical work.
We have built and maintained ColdFusion systems since 1998, and we now run AI delivery on those same systems. This data describes staffing calls we make every week, so it is worth reading past the headline.
Which IT tasks AI is devaluing, according to payroll data
The IT tasks that lost wage value in the ADP study share a shape: routine execution. By execution we mean work with a known procedure, a clear finish line, and an output you can check quickly. That is the work current AI tools compress best.
The higher-paid tasks share a different shape. By judgment we mean calls with consequences: deciding what to build, judging whether a design will hold, directing work when the path is unclear. No tool output settles those questions. Someone has to be right.
Two cautions keep the reading honest. The study measures how tasks relate to wages; it does not prove AI caused every shift. And a few tasks appear on more than one list, so no single task is safe by its label. The shape is the signal, not any one line item.
Execution is getting cheaper. Judgment is not.
Stanford employment research points the same direction as the wage data: the employment declines show up for entry-level workers in AI-exposed occupations, not for experienced ones. Working from ADP payroll records through June 2026, Brynjolfsson, Chandar, and Chen found that employment of workers aged 22 to 25 in AI-exposed occupations sits 19% below where it would be had it kept pace with less-exposed peers. They found no comparable gap for experienced workers. The declines concentrate in occupations where AI substitutes for human tasks; where it complements workers, employment is flat or rising.
Read together, the two studies describe one market. Entry-level execution is where AI competes. Experience, which is mostly accumulated judgment, is where it assists.
That judgment is not spread evenly across software work. Greenfield feature work gives the model a clean field and leaves the person a smaller role. The judgment premium concentrates where the ground is hostile.
Why legacy systems are where judgment earns the most
On a twenty-year-old system, the expensive question is never how to write the change. It is what is safe to change. Which of four near-identical functions is the live one. Whether the undocumented scheduled task that reads this table survives the migration. What the original author knew that the code no longer says.
An agent drafts the change in seconds. Deciding whether the change is safe means evaluating a system that predates every test, doc, and convention the agent expects. That evaluation is design work, the kind the data ties to higher pay. The routine work AI absorbed was never the hard part of ColdFusion migration and modernization. The hard part was knowing where the risk sits. We covered how to get that knowledge from evidence rather than from a plausible guess in our post on AI archaeology in ColdFusion systems.
What to do with your own team
Staff for judgment and automate execution. In practice that means three moves.
Price tasks, not titles. The unit of change is the task bundle inside each role. Find the hours your team spends on work the declining list describes. Move those hours to tooling.
Grow evaluators on purpose. Review, verification, and design judgment are now the scarce inputs. People build them through exposure to consequences, not through courses. If AI absorbs the junior execution work, give juniors another route to real consequences, or the next generation of evaluators never forms. We made the staffing case in our post on why AI moved the bottleneck to code review.
Point AI at execution. Keep people on judgment. Teams that invert this, with AI drafting designs while people do routine integration, get the worst of both curves.
If you are deciding where AI fits your organization, start by mapping which of your workload is execution and which is judgment. That map is most of what an honest AI readiness assessment produces, and it is a better guide than any headline about developers being finished.