A framework for how AI reshapes a job.
Most studies of AI and work measure the effect at one layer - a task, a worker, a firm - and miss where it lands. This framework follows the effect across the whole chain. It starts by explaining why even the most rigorous single study can't generalize.
Four assumptions that intellectually rigorous AI studies get wrong.
There are four things these studies often get wrong. With these fundamental misframings, the studies - no matter how rigorous - produce results that cannot easily be generalized.
- 01
They assume a fixed capability frontier.
A productivity study estimates an effect at time T. AI capability is non-stationary; it changes every few weeks. So a point-in-time productivity study is partly obsolete by the time it's published. Prior tools used to be static long enough to be measured. That's no longer true.
- 02
They assume value doesn't migrate - that productivity gains are captured at the same point as yesterday.
AI's productivity gains are often captured not by the firm or worker adopting it, but upstream by the tool provider, or by the players above the algorithm. So the surplus migrates out of the frame the study is pointed at. This is why worker-level productivity can rise with no wage capture, and why the "AI boom will show up in the statistics eventually, like the Solow paradox resolved for computers" expectation is misplaced. The gain isn't late to arrive. It's accruing at a different altitude, and firm-level metrics cannot count it.
- 03
They treat adoption as a one-time deployment.
In previous tech shifts, you could give the tool to workers and measure the lift. But AI adoption is continuous capability absorption, with tools, workflows, and practices co-evolving, plus an ongoing human-in-the-loop bill to be paid for verification, evaluation, and re-tuning. Deployment is continuous, so before-and-after studies that assume one-time deployment aren't even structured correctly. There is no static "after" to measure.
- 04
They assume the unit of work stays the same.
A study measures a task or a role. But foundation models cut across functional boundaries and unbundle the role into fragments that get rebundled into other roles, teams, and even organizations - a task you performed could now be performed by your customer. The study measures one role or one org, and misses that the work has moved to an entirely different location.
AI reshapes a job in five steps.
If no single study generalizes, what does? A frame that follows the effect across every layer it crosses. AI runs a job through one causal chain: capability, substitution, interface, bundle, capture. Read each step two ways - the task-centric view most studies take, and the system-centric view the framework takes.
AI gets better at specific tasks. Benchmarks measure the boundary moving.
Substitution looks like automation. The view misses that the substituted task was part of a bundle held together by a constraint. Removing the task removes the constraint.
Tasks reshuffle across roles. The view misses that interfaces are where constraints used to bind, and that reshuffling is possible because the constraint is gone.
Jobs change as tasks change. The view misses that a bundle is held together by a constraint; when the constraint shifts location, the bundle re-forms as a different bundle, not a modified old one.
Wages move; some workers gain, others lose. The view misses that the control point is whoever now holds the new constraint - often not the worker at all.
Capability gains widen which constraints - risk-bearing, coordination, scarcity, judgment - machines can hold.
Substitution dismantles the binding constraint. The bundle starts to come apart - unbundling begins.
With the constraint dismantled, the cost of cutting between teams and firms collapses. New interfaces form where the next constraint will sit.
The constraint shifts to a new location. Bundles rebundle around the new constraint.
Value migrates to whoever holds the new constraint - usually a tool or infrastructure provider, not the worker.