The work that remains

On January 28, 1986, the space shuttle Challenger broke apart 73 seconds after launch. Seven crew members died. The cause was a rubber O-ring—a seal designed to prevent hot gases from escaping the solid rocket boosters. The O-ring had worked flawlessly in previous launches, but on an unusually cold Florida morning, it lost flexibility, failed to seal, and triggered a catastrophic chain reaction.

The Challenger disaster became a case study in systems thinking. The O-ring was one of thousands of components, costing pennies compared to engines costing millions. But in a system where every component must work for the system to work, the weakest link determines the outcome. The shuttle wasn't 99% successful because 99% of its parts functioned. It was 0% successful because one part failed.

Seven years later, economist Michael Kremer formalized this insight in "The O-Ring Theory of Economic Development." His core observation: many production processes are multiplicative rather than additive. Output equals the product of component qualities, not their sum. When tasks are quality complements, improving nine of ten components while neglecting the tenth may accomplish nothing. The bottleneck constrains everything.

I've been thinking about O-rings constantly over the past month, as a remarkable convergence of research has reshaped our understanding of how AI affects work. A new National Bureau of Economic Research (NBER) working paper from Joshua Gans and Avi Goldfarb—titled, appropriately, "O-Ring Automation"—argues that standard AI workforce projections are built on a foundational error. They assume tasks add together when they actually multiply. And this isn't a minor methodological quibble. It means that virtually every AI productivity forecast you've seen—from McKinsey, from Goldman Sachs, from the consulting firms advising your board—is systematically wrong.

The standard approach to measuring AI's workforce impact follows a simple logic: identify what tasks comprise a job, determine which tasks AI can perform, aggregate the results. This methodology underlies the most cited research in the field—Frey and Osborne's automation probabilities, Webb's patent-text analysis, Felten's AI exposure indices, and Eloundou's GPT-4 assessments.

All of these studies use some version of weighted linear aggregation:

If an occupation consists of ten tasks and nine are highly automatable, this formula yields "90% exposed." The implicit assumption is that automating 90% of tasks eliminates roughly 90% of the job's value—or at least 90% of the human contribution.

Gans and Goldfarb's critique is devastating in its simplicity: this math is wrong when tasks are quality complements. Under O-ring production, output is multiplicative:

If the tenth task is a binding bottleneck—a task whose quality constrains overall output regardless of performance on the other nine—then automating the other tasks doesn't eliminate 90% of the job. It may not eliminate any of the job. The worker simply reallocates their time to the bottleneck task, potentially performing it at higher quality than before.

This is what Gans and Goldfarb call the "focus mechanism." When some tasks are automated, workers don't lose those tasks and keep everything else constant. They concentrate their fixed time endowment on the remaining tasks. A worker who previously allocated one hour each to ten tasks now allocates two hours each to five tasks. Quality on the remaining tasks increases. And if those remaining tasks are bottlenecks—if they're the binding constraints on output—then total value may actually rise.

The O-ring framework might have remained an elegant theoretical critique if not for three empirical studies published in the past four months. Together, they provide the data to test whether the bottleneck problem is real—and the answer is unambiguous.

OpenAI's GDPval (October 2025) asked the capability question: Can AI perform economically valuable tasks at expert quality? They constructed 1,320 tasks across 44 occupations representing significant GDP contribution, then had professionals with an average of 14 years experience grade AI outputs against human expert outputs in blind evaluations. The headline finding: frontier models like Claude Opus 4.1 achieved a 47.6% win rate against human experts. AI can now match or exceed human performance on roughly half of tested tasks, while operating 100x faster and cheaper.

This is impressive capability research. But notice what it doesn't tell you: whether those tasks are bottlenecks or complements within actual production processes. GDPval measures what AI can do in isolation. It doesn't measure how AI capability translates to workforce outcomes.

Anthropic's Economic Index (January 2026) asked the usage question: How is AI actually being used? They analyzed over a million Claude.ai conversations and a million API transcripts, classifying tasks by complexity, skill requirements, and success rates. The findings were striking: usage is highly concentrated (top 10 task categories account for 24% of all conversations), success rates decline with task complexity (70% for simple tasks, 66% for complex), and AI disproportionately handles higher-education tasks within occupational profiles.

But here's the critical finding: Anthropic tested what happens when you model tasks as complements rather than substitutes. Under standard separable-task assumptions, their data implies 1.8 percentage points of annual productivity growth from AI. When they apply a constant elasticity of substitution framework with σ = 0.5—meaning tasks are complements—that projection drops to 0.7–0.9 percentage points. Add success rate adjustments and you're at 0.6 percentage points.

The adjustment from separable tasks to complementary tasks reduces projected productivity gains by 50-67%. This is the first large-scale empirical confirmation of the O-ring framework's predictions.

Microsoft's AI Diffusion Index (January 2026) asked the adoption question: Who is using AI? Using aggregated telemetry adjusted for device market share and internet penetration, they found global adoption reached 16.3% by the end of 2025. But the geographic distribution is unexpected: the UAE leads at 64% adoption, Singapore at 60.9%, while the United States—despite leading in AI infrastructure and frontier model development—ranks 24th at 28.3%.

Microsoft's data reveals that capability development and adoption leadership are different things. Countries that invested early in policy coordination, trust-building, and institutional preparation (UAE appointed the world's first AI minister in 2017, five years before ChatGPT) are realizing benefits that capability leadership alone cannot deliver.

Anthropic's most provocative finding was what they called the "deskilling effect." When you remove AI-covered tasks from occupational profiles, the remaining work requires lower educational attainment on average. AI handles the sophisticated parts; humans keep the routine residual.

Their examples are vivid: Technical writers lose "analyze developments to determine revision needs" (18.7 years average education) while keeping "draw sketches to illustrate materials" (13.6 years). Travel agents lose "plan and sell itinerary packages" (13.5 years) while keeping "print tickets" (12.0 years).

This sounds alarming. But the O-ring framework suggests a more nuanced interpretation. The deskilling effect is real when AI removes high-value tasks and leaves commodity residual. But it reverses—becomes superskilling—when AI removes commodity tasks and leaves genuine judgment bottlenecks.

Consider a property manager whose role includes rent analysis, bookkeeping, maintenance coordination, and CRM management alongside lease negotiations, owner relationships, and strategic asset decisions. AI excels at the first cluster—it's pattern work, sophisticated in details but systematizable. What remains is bottleneck work: negotiating with difficult tenants, managing owner expectations, making strategic decisions that determine whether properties thrive or churn.

For this property manager, AI doesn't deskill—it liberates. They can focus entirely on the judgment work that was always their real contribution, performing it at higher quality and higher volume. The same phenomenon that deskills a technical writer superskills a property manager.

The difference isn't the technology. It's the composition of the role—specifically, whether the tasks AI handles are bottlenecks (deskilling) or complements (superskilling).

Here is the problem with the current research landscape: we now have excellent data on AI capability (GDPval), usage patterns (Anthropic), adoption breadth (Microsoft), and theoretical frameworks (Gans-Goldfarb). But none of these answers the question that organizations and individuals actually need answered:

GDPval tells you AI can write a legal brief at expert quality. It cannot tell you what percentage of a specific lawyer's role involves brief-writing versus client relationships, court appearances, and judgment calls.

Anthropic tells you AI usage concentrates in certain task categories with declining success rates for complexity. It cannot tell you whether the tasks AI handles in your role are bottlenecks or complements.

Microsoft tells you 28.3% of Americans use AI tools. It cannot tell you whether your specific industry, function, or role will experience deskilling or superskilling.

The O-ring theory tells you bottlenecks matter more than average exposure. It cannot tell you which tasks in a specific role are the bottlenecks.

This is the diagnostic gap: the space between macro research findings and micro transformation decisions. The research establishes that composition determines outcomes—but provides no methodology for assessing composition at the level where decisions are actually made.

Over the past two years, my team has been developing a diagnostic methodology designed to fill this gap. We call it the PJRC framework—Pattern, Judgment, Relationship, Creativity—and it's the core of what we're building at The Scaffold.

The framework classifies work across four dimensions based on a five-source weighted synthesis—drawing on Frey & Osborne, Felten, ILO sector data, Anthropic usage patterns, and McKinsey skill bands:

Pattern Work: Systematizable tasks suitable for AI automation. In O-ring terms, these are complements—their automation frees human attention for bottleneck tasks. When AI handles pattern work, the focus mechanism kicks in: workers concentrate time on remaining tasks, potentially increasing their quality.

Judgment Work: Tasks requiring contextual decision-making, ambiguity navigation, and the application of wisdom that can't be reduced to rules. These are often binding bottlenecks that become MORE valuable through the focus effect when pattern work is automated.

Relationship Work: Tasks centered on trust-building, stakeholder alignment, negotiation, and the human connection that clients and colleagues value. This is frequently the most durable bottleneck—the work that strengthens rather than weakens as AI handles analytical tasks.

Creativity Work: Tasks requiring novel synthesis, aesthetic judgment, and the generation of ideas that didn't exist before. This includes both artistic creativity and the strategic creativity that identifies opportunities others miss.

The diagnostic answers the question the research cannot: what is the composition of this specific role across these four dimensions, and does AI engagement leave bottleneck remainder (superskilling trajectory) or commodity remainder (deskilling trajectory)?

The PJRC framework is operationalized through what we call the AI Mirror diagnostic. The assessment creates what we've come to call a "Mirror Moment"—the uncomfortable confrontation between how individuals perceive their work composition and what empirical analysis suggests.

Here's what we've found in early deployments: most knowledge workers believe they do roughly 40% pattern work. The diagnostic typically reveals 60-80%. That 20-40 percentage point gap is the perception gap—the space between self-image and structural reality.

This gap exists for understandable reasons. We experience our work phenomenologically: the decision feels like judgment, the analysis feels like expertise, the communication feels like relationship-building. We don't experience our work structurally: noticing that the decision follows a pattern we've repeated dozens of times, the analysis applies a framework we could articulate as rules, the communication follows templates we've internalized.

The Mirror Moment doesn't judge—it reveals. And what it reveals is the answer to whether AI engagement leads to liberation (focusing on genuine bottlenecks) or vulnerability (watching AI capture your actual contribution).

The property manager who discovers that 65% of their role is pattern work—rent analysis, maintenance scheduling, communication drafting—is seeing liberation potential. Remove that pattern work and they can focus entirely on the 35% that's genuine judgment: tenant negotiations, owner relationships, strategic decisions.

The technical writer who discovers that 70% of their role is pattern work—formatting, consistency checking, template application—and that the remaining 30% is also trending toward pattern (AI is rapidly improving at analytical documentation) is seeing vulnerability. Their bottleneck is weakening.

Individual diagnostics are valuable, but the strategic question is organizational: across departments, functions, and roles, where does transformation pressure concentrate? Which teams face superskilling potential, and which face deskilling risk? How does pattern work distribution vary across the organization?

This is what we're building with Organizational Intelligence—the aggregation layer that transforms individual AI Mirror assessments into strategic workforce insights.

The capability surfaces patterns that would be invisible at the individual level. Imagine seeing that Operations shows 66.7% of employees highly exposed to automation, while Human Resources shows only 28.6%. That Sales has the highest concentration of highly-exposed employees, while Product has the most balanced distribution across exposure categories. This is the organizational view of the O-ring problem—not abstract exposure indices, but the actual bottleneck structure of your workforce.

A heat map of automation exposure by department and skill band tells a different story than aggregate exposure indices. It shows where transformation pressure will manifest and where the focus effect creates opportunity. When you can see that your organization's average pattern work is 49.6% with a standard deviation of 20.8%—ranging from employees at 21.7% (highly resilient) to 79.6% (highly exposed)—you're no longer working with abstractions. You're planning transformation with structural clarity.

This is the diagnostic infrastructure that translates the O-ring framework from academic insight to organizational strategy.

The convergence of the Gans-Goldfarb framework with Anthropic's empirical confirmation creates a specific window. The "50% discount" heuristic is entering mainstream discourse—Evan Armstrong at The Leverage articulated it this week, and I expect it will spread rapidly through the CHRO community over the coming months.

This creates both opportunity and urgency. Opportunity because organizations are beginning to recognize that their exposure-index-based AI strategies may be built on flawed foundations. Urgency because the window for establishing diagnostic methodologies is finite—the major consulting firms will eventually adapt their frameworks, and the HR tech vendors will eventually build assessment tools.

But there's a deeper reason this matters now. The research synthesis reveals something that hasn't been widely understood: the successful organizations facing the greatest AI transformation risk are successful precisely because they systematized knowledge work into repeatable patterns. That systematization—the thing that made them scalable—is exactly what makes their work automatable.

I call this the Pattern Paradox: organizations that scaled through standardization created the pattern work that AI now handles best. Their competitive advantage became their vulnerability. And the workers who excelled at executing those patterns—the reliable performers who followed best practices—may be more exposed than the mavericks who never fit the system.

This is uncomfortable. It means that organizational success and individual competence, as traditionally measured, may inversely correlate with AI transformation resilience. The diagnostic question isn't "are you good at your job?" It's "is your job composed of patterns or judgments?"

Let me be clear about what the Pattern-Judgment framework is and isn't.

It is a diagnostic methodology grounded in converging research from labor economics (O-ring theory), AI capability studies (GDPval), usage analytics (Anthropic), and adoption patterns (Microsoft). It provides the individual-level assessment that translates macro findings into micro decisions.

It is not a prediction machine. No methodology can tell you exactly how AI capabilities will evolve or which bottlenecks will remain irreducibly human. What looks like judgment work today may become pattern work at GPT-7. What looks like commodity residual may turn out to be surprisingly durable because of physical presence requirements, regulatory constraints, or pure human preference for human interaction.

The honest position is this: we now have the framework to ask the right question—bottleneck or complement, superskilling or deskilling—and a methodology to answer it based on current evidence. We don't have certainty about trajectories. We have informed assessment that requires continuous updating as capabilities evolve.

Armstrong's personal audit approach is exactly right: with each major model release, reassess whether a bottleneck was just removed. The diagnostic isn't a one-time exercise; it's a continuous practice of understanding work composition as the landscape shifts.

The O-ring that failed on the Challenger wasn't different from the thousands of other components in the shuttle. It wasn't obviously critical. It was simply the bottleneck that day—the single point whose failure propagated through the entire system.

The tasks that comprise your job work similarly. Some are bottlenecks—binding constraints whose quality determines overall output. Others are complements—important contributions that nonetheless don't constrain the whole. AI is rapidly becoming capable of handling the complements. The question is what remains.

If what remains is genuine judgment work—ambiguity navigation, relationship depth, contextual wisdom—then AI represents liberation. You can focus entirely on the bottleneck, performing it at higher quality than when you were splitting attention across pattern work. Your value increases.

If what remains is commodity residual—administrative tasks, physical presence requirements, legacy processes—then AI represents vulnerability. The sophisticated work gets automated; you're left with work that doesn't justify your compensation. Your value decreases.

Standard exposure indices can't distinguish these trajectories because they assume tasks add together. They don't. They multiply. And that means the diagnostic question isn't "what percentage of my tasks can AI perform?" It's "are those tasks bottlenecks or complements?"

That's the question the research now frames. That's the question the Pattern-Judgment framework answers. And that's the question everyone—every organization, every worker, every leader planning transformation—needs to be asking.

The work that remains determines the outcome. Know what remains.

The Scaffold Platform provides the diagnostic methodology described in this article. The AI Mirror assessment reveals individual work composition, while Organizational Intelligence aggregates assessments into strategic workforce insights—showing where transformation pressure concentrates and which teams face superskilling versus deskilling trajectories. For organizations and practitioners interested in pilot deployments, contact us at thescaffold.io.

Gans, J.S. & Goldfarb, A. (2026). "O-Ring Automation." NBER Working Paper No. 34639.

Patwardhan, T., et al. (2025). "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks." arXiv:2510.04374.

Appel, R., Massenkoff, M., McCrory, P., et al. (2026). "The Anthropic Economic Index Report." Anthropic Research.

Microsoft AI Economy Institute. (2026). "Global AI Adoption in 2025." Microsoft Research.

Kremer, M. (1993). "The O-Ring Theory of Economic Development." The Quarterly Journal of Economics.