Organizations treat AI deployment as a single job. Look at what's actually happening inside any deployment and you find three distinct activities, each with its own failure mode.
Specification work happens before the agent runs: making a task explicit enough to delegate. Boundaries, acceptable outcomes, recovery paths. In autonomous chemistry labs, researchers define materials, objectives, and experimental budgets before the automated cycle starts. The system operates inside that envelope.
Execution work is the agent performing the task. Merged code, resolved tickets, processed cases. This phase absorbs most organizational attention because it produces countable outputs.
Adoption work makes those outputs usable: reviewing results with enough context to accept or refuse them, detecting failures, maintaining the organizational capacity to recover when something breaks.
Execution scales fast. Specification and adoption require human judgment that doesn't compress the same way, and that asymmetry is where most deployment-to-value gaps originate.
Specification — Workflow redesign, constraint-setting, defining "done" before anything runs. McKinsey's 2026 survey found high performers were 3x more likely to have redesigned workflows before scaling.
Execution — Agent runtime. One large-scale study found 24% more merged pull requests per engineer, but the authors called that a proxy for output, not delivered value.
Adoption — Review, exception handling, recovery. 68% of production agents run ten steps or fewer before a human intervenes. That intervention is adoption work.
Why they collapse — Execution produces metrics that fit existing reporting structures. Specification and adoption produce preconditions and recovery capacity — real organizational value that doesn't register as productivity.