AI tools triple commit-level output on a per-task basis, but the number of shipped releases grows only about 30%, and the actual usage impact remains unclear.

Bottlenecks shift over time in line with broader AI history, so monitoring downstream stages is crucial to understand AI’s ultimate impact on software output.

Increases in code volume from AI do not automatically translate into more projects or releases due to bottlenecks in reviewing, integrating, testing, and releasing.

Task-level productivity gains from AI range roughly 15% to 50% across domains like software development, writing, and customer support, with bigger gains as tools evolve.

Despite attenuation, these gains remain economically meaningful and may grow as downstream bottlenecks are addressed through better review, automation, discovery, and adoption.

Authors use public GitHub histories and Microsoft data with a matched event-study design to estimate causal effects of AI tool adoption on coding activity and output.

The study tests whether task-level productivity translates into final output, addressing the bottleneck or 'weak links' in software production.

Evidence from app marketplaces shows more app releases but uneven consumer engagement, indicating supply expansion has not yet boosted software consumption.

Task-level gains accumulate across tool generations: autocomplete ~40% output at commit level, sync agents ~140%, and async agents ~180%, with larger effects for less active developers.

The main implication is that aggregate output growth depends on downstream bottlenecks; task-level gains cannot be naively extrapolated to final software output.