THE INSTRUCTION GAP

The bottleneck in AI-assisted development has shifted from prompting to execution. Most failed builds trace to incomplete human execution of correctly specified steps, not model error or poor prompt construction. The field has not updated its framing to reflect this shift.

Working memory capacity is the primary constraint on instruction-following performance, documented across five decades of psychological research. The failure pattern is not random: early steps in a sequence get completed, middle and later steps do not. The failure is cognitive, predictable, and patterned.

A 2016 study in Memory and Cognition found that executing each step immediately upon reading it, rather than reading the full sequence and then executing from memory, significantly improves completion rates. The mechanism is enactment: doing the step encodes what comes next.

A 2015 study in Scientific Reports found that demonstrated instructions produce significantly better recall and completion than written or spoken ones. This explains why video walkthroughs consistently outperform written documentation even when the written version is more complete.

Milgram's obedience research across 19 experimental variations shows that instruction-following compliance drops measurably when no authority structure enforces completion. Most AI implementation environments have no such structure, and the data predicts exactly what practitioners observe.

Fewer than 20 percent of enterprise AI projects achieve their expected ROI (Boston Consulting Group). Instruction-following failure as a root cause does not appear in standard post-mortems because it has not been named or measured. The paper identifies the specific mechanics of how enterprise sequence execution breaks down across distributed teams.

The paper closes with five principles of good instruction-following practice: enactment before advancement, explicit verification, sequence integrity, single-thread execution, and debrief on failure before diagnosing tool error.