Theory

Theory statement

AI assistance creates net engineering value only when generation, search, and cognitive-relief benefits exceed interaction, verification, coordination, rework, and lifecycle costs for the measured unit of work.

Scope and boundary conditions

Applies to contemporary generative coding assistants and agents. It does not predict a fixed effect size and must be re-estimated as tools and workflows change.

Underlying mechanism

net value = avoided production/search cost - interaction/verification cost - transferred downstream cost, conditioned on task, person, codebase, tool, workflow, and measurement horizon.

Supported predictions

Narrow checkable tasks can benefit; expert familiar-repository work can slow; individual satisfaction can improve while delivery outcomes degrade; automated checks can miss merge-readiness costs.

Failed predictions

None tested prospectively.

Supporting evidence

EV-AICODING-2026-10A1, EV-AICODING-2026-3B8F, EV-AICODING-2026-6E21, EV-AICODING-2026-C940, and EV-AICODING-2026-F5D2.

Contradicting evidence

No evidence yet shows invariant effects; absence may reflect the small evidence map rather than a true lack of counterexamples.

Alternative explanations

Differences may be entirely temporal (model capability), study sponsorship, participant selection, or measurement error rather than stable moderators.

Confidence rationale

Medium (0.60): the accounting model explains contradictions and is falsifiable, but moderator causality and generality are untested.

Open questions

Which terms dominate by task class, and which measures predict lifecycle value?

Required updates

Prospectively test EX-AICODING-2026-2D77 and update or reject the theory.