The Maze: AI coding tools are doing exactly what they promised: making software easier to produce. The catch is that software still has to be reviewed, packaged, discovered, trusted, downloaded, and used. A new NBER paper finds that coding agents lift upstream developer activity sharply, but the gains shrink as they move toward releases and user adoption. The app economy did not run out of code. It ran out of attention.
The supply curve moved first. The visible iOS evidence shows monthly app releases moving from a 2024 baseline of 100 to roughly 180 by early 2026. The broader NBER paper finds the same direction across major software marketplaces: iOS accelerates sharply, Chrome extensions roughly double from 2023 levels by mid-2026, and Google Play breaks from a prior decline. AI lowered the cost of creating more software. That is real productivity. It is just not the same as real demand.
Usage refused to follow the release curve. In the iOS visual, apps with significant usage sit near the baseline for almost the whole period and end around 92. App reviews are noisier, but they also do not rise with releases and end around 75. The paper's marketplace section reaches the same conclusion across platforms: cohort-level usage within the first three months is flat or declining, even as entry rises. More apps entered the store. Users did not create more hours in the day.
The bottleneck starts before the app store. The paper's developer-side estimates are stark: autocomplete, sync agents, and async agents raise commits by about 40%, 140%, and 180%. But that 180% cumulative effect falls to about 50% for projects and about 30% for actual releases. The authors estimate an elasticity of substitution of 0.25 between AI-generated upstream output and downstream human effort. Translation: code and human judgment are complements, not plug-in substitutes.
Discovery may be the next weak link. The paper tests the nicer version of the story: maybe new apps improve matching even if total usage does not rise. The evidence is not kind. The share of low-audience releases rises after January 2025, from about 79% to 86% on iOS and from about 18% to 31% on Chrome. That does not prove every marginal app is bad. It does suggest the marginal app is increasingly invisible.
This is a warning for ecommerce operators, not just developers. Every marketplace has the same physics. Lower creation costs flood supply. Demand then sorts brutally through trust, distribution, reputation, ranking, and habit. AI can make the catalog thicker. It cannot guarantee the shelf gets shopped. The scarce asset shifts from production capacity to market access.
Why it matters: AI's first-order effect is abundance. Its second-order effect is congestion. For retailers, SaaS tools, app stores, seller platforms, and creator marketplaces, the question is no longer "can this be built?" It is "who will notice, trust, and repeatedly use it?" The winners will not be the teams that ship the most artifacts. They will be the teams that connect AI-made supply to a real demand path.
Sources: NBER working paper | LinkedIn source post
