Does fixing AI adoption take different playbooks for industries of many small firms versus a few giant ones?

This explores whether AI spreads differently through industries made up of many small, scattered firms than through industries run by a few giants. To be direct: none of these notes compares fragmented and consolidated industries head-on. What the corpus does offer is a strong argument about *where* AI scaling actually gets stuck. Once you see where, you can reason about how industry structure changes the picture.

The main claim is that the barrier is coordination, not capability. Does AI adoption require rethinking how organizations coordinate? builds a formal model showing that AI improves individual tasks but makes decisions more variable. When decisions depend on each other, you only get the big payoff if the organization redesigns how those decisions fit together. Does easier tool-building actually solve enterprise adoption problems? adds a second barrier: most workers don't recognize their own tasks as automatable, and enterprise adoption means decisions that cross departments and drag out over long timelines. Put these together and you get a hypothesis the corpus doesn't test but strongly implies. A consolidated firm can, at least in principle, run that redesign internally under one leadership team. In a fragmented industry, the interdependent decisions are spread across many separate companies, so the 'organizational redesign' becomes an industry-wide coordination problem with nobody in charge of it.

The same pattern shows up at the level of agents. Do single agents always hit organizational limits? argues that real work needing varied expertise, parallel steps, and independent checking outgrows any single agent loop, so you need systems of specialized agents. That is a technical mirror of the organizational point: the limit is how the work is structured, not how smart the model is. Why do agent benchmarks not predict real economic value? supports this from the evaluation side. Agents win contests but struggle with long, multi-step professional workflows. Those workflows are often exactly the messy, handoff-heavy processes that fragmented industries run between firms.

There is a counterweight worth knowing about. Fragmented industries may struggle with coordination, but they could gain more per worker. Does AI help lower-skilled workers catch up to top performers? found that lower-skilled consultants gained the most from GPT-4, though only when use was mandatory. Discretionary adoption risked reversing the effect, and in fragmented sectors with no central mandate, adoption is almost always discretionary. Is the AI capability gap really an interface problem? suggests that interface design, not model power, decides whether less experienced users actually capture those gains. That matters most for small firms with no internal AI team to smooth things over. Cost points the same way: Does model efficiency matter more than peak capability for real work? argues that cheap, compact models trained for follow-through can rival larger ones on multi-step work. That could lower the entry price for small players.

The takeaway you might not have expected: the corpus suggests the real gap between fragmented and consolidated industries is less about access to AI and more about who can do the reorganizing AI demands. Consolidated firms can mandate adoption and redesign workflows. Fragmented industries may need someone outside any single firm, such as a platform, a trade body, or a shared agent standard, to do that coordinating for them. If you want empirical comparisons by industry structure, this collection doesn't have them yet.