Sunil Venkataram
Back to garden
•
🌰 seedling
O-Ring Production and AI Automation - Why Partial Automation Can Raise Wages
type/noteailaboreconomics

O-Ring Production and AI Automation — Why Partial Automation Can Raise Wages

The O-ring idea

Michael Kremer’s 1993 paper “The O-Ring Theory of Economic Development” proposed that many production processes are multiplicative: the final output is the product of quality across all steps, not the sum. The name comes from the Challenger disaster, where one faulty O-ring destroyed the entire system. A worker who makes slightly fewer errors per task becomes dramatically more productive overall because the quality gains compound.

Gans and Goldfarb applied this framework directly to AI automation. In their model, a worker’s job has n tasks. Output follows an O-ring production function:

Y = product of quality across all n tasks

Each task’s quality scales with the time the worker spends on it. The worker has a fixed time budget spread across all tasks.

The focus effect

When a firm automates k tasks, the worker no longer spends time on those. All available hours go to the remaining n - k tasks. Each surviving task gets more time, which means higher quality. Because quality is multiplicative, those gains compound.

The math shows that output can increase even if the automated tasks run at slightly lower quality than the human originally achieved. The worker’s concentration on fewer tasks more than compensates.

This is the core counter-argument to the idea that any AI capability overlap with a job equals displacement. In an O-ring job, partial automation raises the worker’s marginal product. Higher marginal product typically means higher wages.

Where the focus effect holds

The effect is strongest in high-dimensional jobs with genuinely complementary tasks. A management consultant who offloads slide creation and data analysis to AI can spend more time on client relationships, strategic reasoning, and implementation support. Each of those tasks improves the quality of the others. The consultant becomes more valuable, not less.

Medicine works the same way. A radiologist who offloads routine pattern recognition to AI spends more time on complex cases and clinical judgment. Over 870 FDA-approved radiology AI tools exist as of 2026, and 66% of doctors use at least one AI tool. These tools augment rather than replace because the job has many complementary tasks beyond the automated one.

Where it breaks down

The focus effect fails when all or nearly all tasks in a job can be automated. If a job has only one or two core tasks and AI handles them both, there is nothing left for the worker to concentrate on. The job disappears. See Job Dimensionality - Why Low-Task Jobs Face the Highest Automation Risk.

It also fails when the automated task quality is far below what the human achieved. Low-quality automation on even one task can drag down the entire multiplicative output. The threshold: automation quality must be at least close to the worker’s original manual quality on those tasks for the focus effect to produce net gains.

Why this matters for policy

Standard AI exposure indices (like the widely cited Eloundou et al. “GPTs are GPTs” finding that 80% of workers have at least 10% of tasks exposed) measure task-level overlap. They do not distinguish between jobs where partial automation makes workers more productive and jobs where it eliminates positions. The O-ring framework says the structure of task complementarity within a job matters more than the percentage of tasks exposed.

Two jobs with identical exposure scores can have opposite displacement outcomes depending on whether their tasks are complements or substitutes.

Backlinks (4)
Connected Notes