Weak Links Cap Aggregate Gains Even Under Widespread Automation

# Weak Links Cap Aggregate Gains Even Under Widespread Automation A system's throughput is set by what it cannot improve, not by what it improves fastest. Automating the automatable part of a process does not deliver a proportional gain, because the un-automated remainder becomes the binding constraint — and it binds harder the more the rest speeds up. Aghion, Jones and Jones put it as: growth "may be constrained not by what we do well but rather by what is essential and yet hard to improve." The cleanest evidence is from software, where automation went furthest first. After coding agents arrived, lines of code generated rose **10–20x**. Software releases rose about **30%**. App usage did not rise at all. Code generation was never the constraint; review, integration, coordination, judgment about what to ship, and users' appetite for new software all were. Multiplying the one input that was already cheap moved the output almost not at all. This is why "AI can do X% of tasks in this job" is a bad predictor of anything. It measures the wrong end. The number that matters is what share of the *essential* tasks remain un-automated and how hard they are — because those tasks set the ceiling on the gain and, in a labor market, hold the returns to the people who do them. It also explains a result that otherwise looks contradictory: strong labor-productivity growth and a stable unemployment rate at the same time. Output per hour rises because the automatable work got cheap; hours don't fall because the weak links still need humans.  ## Why This Matters For **reasoning about AI's economics**, this is the structural reason displacement forecasts keep overshooting. Forecasts extrapolate from capability on measurable tasks; systems are governed by the un-measured essential ones. Any projection that doesn't name its weak link is projecting an unconstrained system that doesn't exist. For **work Courtney actually does**, it's a planning rule. Before adopting a tool that makes some step 10x faster, ask what the next constraint is — because that's what you'll be buying. A content pipeline that generates drafts instantly still bottlenecks on review, accuracy checking, and the audience's attention. A plugin-development workflow that writes code instantly still bottlenecks on testing, compatibility, and release review. For the **long-run growth question**, it's the load-bearing objection to singularity models. Automating innovation itself only produces runaway growth if the automation is *complete*; any essential task that resists — for technical reasons or societal ones — keeps growth finite and keeps the labor share of income elevated. ## Counterpoints to Hold - **Weak links move.** They're a snapshot, not a permanent feature. The 30% releases figure is from a moment when review tooling lagged generation tooling; if review automates next, the ceiling lifts. The claim constrains the *rate* of gains, not their eventual size. - **"Essential" is doing a lot of work undefended.** Some tasks look essential because they're currently done by humans, not because they must be. Distinguishing genuinely irreducible tasks from merely un-automated ones is the whole difficulty, and this framing doesn't supply the test. - **The source has an interest.** Anthropic's economics team concluding that Anthropic's product augments rather than displaces labor is not disqualifying, but the framework is more comfortable for the vendor than the alternative. - **Aggregate stability can hide distributed harm.** A capped aggregate gain is fully compatible with specific workers in specific roles losing everything. "No macro effect" is not "no one was hurt." ## Links - [[AI Cost & Tokenomics MOC]] — why spending on the cheap step returns little once the constraint sits somewhere else - [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]] — the same mechanism at the individual and team scale; this note is its economy-wide form, with the 10–20x/30% gap as the measurement - [[3. Resources/Permanent Notes/A System Failure That Looks Like a Performance Failure Never Gets Fixed]] — applies-in: when the weak link is misdiagnosed as slowness, teams buy more of the cheap input instead of addressing the constraint - [[3. Resources/Permanent Notes/Bubbles End From Cost Not From Lost Belief]] — the financial consequence: if weak links cap the payoff, the cost of carry has longer to run before returns arrive - [[3. Resources/Permanent Notes/AI Tools Don't Eliminate the Apprenticeship Layer They Relocate Its Cost]] — the weak links are staffed by experts, which raises the unanswered question of how the next generation of them gets made - [[3. Resources/Permanent Notes/Maintainer Attention Is the Scarce Resource in Open Source]] — the open-source instance: contribution volume was never the constraint, review capacity was, and AI raises the first while leaving the second - [[3. Resources/Permanent Notes/Perm · The middle is disappearing as AI commodifies previously-expert work]] — contrasts: that note reads the same automation as hollowing expertise, where this reads it as concentrating value into whatever expertise remains essential - [[3. Resources/Permanent Notes/Automation Removes the Social Checks a Process Silently Depended On]] — the mechanism by which the weak link got weaker: the informal check that used to sit before review disappeared with the automation - [[3. Resources/Readwise/Articles/How to Redesign a Broken Delivery Flow|How to redesign a broken delivery flow]] — evidence-for, from the delivery-process side: "Your AI tools made the coding faster. They did nothing to help anyone see where the work actually waits" ## Open Questions - What is the weak link in Courtney's own publishing pipeline — is it drafting, verification, or distribution? The rule only pays off if the constraint is named honestly. - Does the 10–20x / 30% gap hold up outside software, or is it an artifact of release cadence being deliberately conservative? - If review automates, does the constraint move to trust and accountability — tasks that may resist automation for *societal* rather than technical reasons? ## Source - [[3. Resources/Literature Notes/Lit · Why hasn't AI increased unemployment]] — Peter McCrory, Anthropic economic research, 2026-07-24 - See also: [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]]