The automation-augmentation binary.
The dominant frame bounds AI's labor effect to two possibilities - AI does the task, or AI augments a less-skilled worker. The frame misses how bundles actually come apart, and gets the labor relationship inside human-in-the-loop work exactly backwards.
What the binary claims.
Most AI-jobs analysis runs on a two-option mental model. For any task in a job:
- Automation - AI does the task. The human is displaced from this task. If enough tasks get automated, the role disappears.
- Augmentation - AI helps the human do the task better, faster, cheaper. Or AI lets a less-skilled human do what previously required an expert. The human stays in the workflow, often with compressed pay.
The framing is intuitive, easy to communicate, and matches the reassurance most reports want to deliver: "some tasks get automated, but other tasks need the human, so the role survives." Both McKinsey-style task-unbundling reports and most institutional commentary run on this binary.
It misses the third mechanism.
A task can also leave a role without AI doing it and without AI augmenting a less-skilled human. The task moves to a different role entirely. The bundle that held it broke; the work relocated to brand teams, project managers, platform contracts, institutional accountability seats - different humans, different contracts, in places the original bundle no longer reaches.
The binary has no language for this. In its mental model, "AI didn't do it" and "AI didn't help a less-skilled human do it" add up to "the work survives in this role unchanged." But in the actual data, the work can persist while its bundle home dissolves.
In the translator case study, half the translator's ten-component bundle breaks via this third mechanism: cultural mediation moved to brand teams, self-QA to separate Human-in-the-loop evaluation layers, client scoping to PMs, confidentiality to platform terms of service, certification to institutional accountability seats. None of those was automated. None was augmented. The binary reads these as "untouched" and misses the actual restructure.
It inverts the Human-in-the-loop relationship.
The "augmentation" half of the binary carries a strong implication: AI assists the human, and the human stays valuable because they bring judgment, nuance, and domain expertise to the AI's output. This is the standard "human-in-the-loop" reassurance.
In the public job market for translators, 89% of postings invert this relationship. The worker exists to plug AI gaps - to label data, evaluate output, transcribe audio that feeds back into someone else's AI training pipeline. The consumer of the worker's output is an AI system, not a paying client. The human augments the AI, not the other way around.
The binary's vocabulary cannot name this configuration accurately. It will keep calling it "augmentation" - because the human is technically still in the loop - even though the labor relationship is structurally the opposite of what the word implies.
A bundle frame, not a binary.
The corrective is to drop the binary and ask bundle questions:
- What constraint binds this role? (Scarcity, risk, coordination.)
- Where is the constraint going under AI pressure? (Collapsing, persisting, shifting to another constraint.)
- Where are the bundle's components moving to? (Inside the model, to a less-skilled worker, or to a different role entirely.)
- Who is the consumer of the worker's output? (A paying client, or an AI system the firm is selling?)
These questions surface the three mechanisms of bundle breakage, not two - and they distinguish AI-as-leverage configurations from Human-in-the-loop gap-filler configurations directly. The binary cannot make either distinction.
The translator data.
In the translator case study, the three-mechanism scoreboard in the Mechanism section shows the full breakdown: 4 of 10 components tool-absorbed, 1 of 10 skill-premium-eaten, 5 of 10 unbundled-and-relocated. The binary sees only the first two columns. The third column - 50% of the bundle - is what it cannot frame.