The technical base is not the stranded asset. The stranded asset is a career that never left the depreciating half of the unbundled job. If you run L&D for that population, cataloguing new tools will train people to operate inside a smaller premium. Recutting ownership uses the same tools and produces a different wage story.
ADP Research recently published Unbundling jobs: Measuring the value of tasks in an AI economy, with the Stanford Digital Economy Lab. Nela Richardson and Andrew Wang do not ask whether AI is deleting occupations. They price the work inside them.
Treat that as an operating hypothesis, not as a finding in the Richardson–Wang paper. The research prices tasks. It does not measure cycle time, defect rates, or skill transfer from end-to-end ownership. Those claims fail or hold in delivery organizations, not in a wage equation.
That loop is correct as far as it goes, and it is incomplete. Tool adoption without a change in what gets reviewed, shipped, and promoted produces power users of a discounted task. Sharing prompt libraries is not a career recut. Putting a senior engineer on specification and evaluation of AI-generated changes, with production accountability, is.
Phillips does not hedge. “I believe the premium on design, judgment, and critical thinking represents a fundamental and lasting shift, not a temporary market adjustment.” For talent leaders, “this is not a trend to hedge against. It is one to build strategy around.”
Phillips’s redesign thesis is that organizations should recut roles around higher-value outcomes rather than task execution. In software development, she argues, AI can reduce handoffs and let one person own more of the lifecycle. Broader ownership, in that telling, improves speed and quality and builds adjacent skill through the work itself.
The shift is inside the role
Contrast two planning postures. One holds the job constant and asks how many of them AI replaces. The other holds the outcome constant and asks which activities still require a person who can specify, evaluate, and decide. Those produce different requisitions, different spans, and different promotion criteria. Only the second one is consistent with the wage signal.
The talent-leader job she describes is to create room for responsible experimentation in daily work, then install mechanisms that move what one team learns onto other teams. “Experimentation drives adoption; adoption combined with knowledge sharing drives reskilling at scale.”
That is not a permission slip to cut the maintenance headcount and call it strategy. It is a requirement to stop treating “software engineer” or “systems analyst” as a stable bundle of equally priced work.
The practical split is binary. Hire for the capacity to frame a problem, write a specification, and live with a tradeoff. Then put that person on work where the specification is theirs and the outcome is scored. Do not hire for monitoring fluency and hope a later rotation manufactures evaluation skill.
Recut the role around outcomes
AI is changing which tasks inside a role employers pay for, rather than eliminating whole roles. Headcount plans built on job titles are therefore buying the wrong object.
The paper does not contain a test of permanence. It contains task prices in IT work under current AI diffusion. Building long-term architecture on a single cross-section is a choice, not a proof. The opposite choice is worse on the evidence we do have. Hedging means keeping job architectures, learning programs, and promotion criteria pointed at a task mix the wage signal has already marked down, on the bet that monitoring and maintenance regain relative value. Nothing in the current price signal supports that bet.
A companion Main Street Macro note names the frame. “What will matter more is the changing value of the granular tasks and activities that make up a job.” Job creation and destruction statistics will keep circulating. They will keep answering the wrong question.
Career growth in Phillips’s account is increasingly horizontal: skills, adaptability, contribution across functions. Cross-functional mobility is not a culture slogan. It is how you keep paying people after the task mix inside their original title has been repriced. Look at the last promotion cycle in your engineering org and ask what actually moved compensation. If the answer is still ticket volume and stack fluency, the career system is pointed at the depreciating half of the unbundled job.
The paid work is design and evaluation
Walk through a current-year workforce plan and the artifact is familiar: FTE counts by function, backfill rules, span-of-control targets, a freeze list. None of that tells you which activities inside those FTEs are still in the paid cluster. Design, specification, advising on technology use, directing technical work: those hold or gain value. Routine monitoring and maintenance: those lose it.
That is the correct unit of analysis. It is also the unit most talent systems still refuse to use.
On judgment, Phillips does not pick a side. It is both hired and developed. Organizations can screen for critical-thinking ability, business acumen, and decision-making potential. “Judgment is ultimately refined through experience, exposure, and accountability.”
The question, then, is not whether AI will eliminate the job title on the org chart. It is whether the work inside that title is being repriced faster than the organization is willing to redesign it.
The “reskill now” question is not “which course.” It is “toward which priced activity.” Toward design of systems, writing of specifications, advising on where the tools belong, and evaluation of outputs against business constraints. Away from a professional identity built only on routine monitoring and maintenance.
Mid-career IT is an expansion problem
The job may remain. The task mix will not.
Accountability is the load-bearing word. Experience without decision rights produces observers. Exposure without consequence produces tourists. A learning program that teaches “critical thinking” in a classroom and then returns people to a role where they monitor systems AI already watches is not developing judgment. It is decorating a task that has lost price.
Some technical skills that paid well before AI have seen their premium shrink. Phillips refuses the replacement story. “This is less about abandoning existing expertise and more about expanding it.” The valuable profile, in her account, is deep technical knowledge combined with the ability to apply AI to business problems and accelerate outcomes.
The paper is an early proof of concept, announced as part of a broader unbundling project at Davos in January 2026. It uses a selected sample of IT jobs, controls for age, gender, and company differences, and finds that certain IT tasks are less valued by employers today than they were prior to 2022 and the first wide release of AI tools. That validation is operationally meaningful but not independent in the audit sense. What it measures is relative prices inside IT work, not a labor-market census.
The architecture consequence is still real. Job families that enumerate tools and tickets will keep promoting people for work the market is discounting. As roles become less defined by a task list and more by capabilities — business acumen, critical thinking, collaboration, judgment — career frameworks that only move people up a ladder of the same tickets will starve the path that still clears the market.
Build around the premium. Do not wait for it to revert.
You already have people whose compensation was set against a task mix this research now marks down.
The premium sits around design and evaluation work. Organizations have always paid more for critical thinking, judgment, and business acumen than for routine execution. What Phillips says has changed is the share of the workforce now expected to do the expensive work, including early-career talent, because the cheap work is being automated.
You cannot run last year’s job families and this year’s task prices at the same time. First movers stop paying senior rates for activities the market has already discounted, and they assign evaluation work to a wider population before the external market does it for them. Slow movers keep the titles and lose the people whose skills still clear the market, or they keep the people and overpay for work whose relative value is falling.
We discussed a number of operating areas with Emma Phillips, ADP’s Division Vice President of Human Resources for Global Products and Technology: how you plan headcount, how you hire for design and evaluation, what you tell mid-career IT, whether you treat the premium as permanent, and what separates organizations that recut work from those still staffing titles. She is answering from inside the company that produced the research. Keep that dual interest in view.
That claim is directional, and it is larger than what an IT-task analysis can prove. The paper shows relative prices inside IT activities. It does not show that every junior role in every function has been rewritten around tradeoffs. The hiring implication does not wait on that proof. If the paid work is advising, specifying, and evaluating, then pipelines optimized for stack fluency and ticket throughput are selecting for the depreciating half of the bundle. Most interview loops still overweight demonstrable execution of a known stack. That is the opposite of the premium.
