Delegation Value Measurement
Define comparable accepted quality and representative tasks before measuring time. Record human baseline, active supervision, review, rework, failures, one-off setup, cash cost and completed volume. Keep elapsed model latency separate from human attention when they overlap. Missing review time is unknown, not zero.
Use a calculation tool. Recurring human effort is active supervision plus review plus rework. First-task effort also includes setup. Amortise only over an explicit supported horizon; label forecasts and observed completed volume separately. Do not hide setup or double-count overlap. Compare quality and guardrails before interpreting savings.
A slower recurring AI path is not a saving just because text appeared quickly. A quality-matched positive trial can support a bounded adoption recommendation; avoid an always-reject policy. Report uncertainty and do not extrapolate one trial to a whole occupation or service.
Set review cadence and stop rules for unacceptable errors, review overhead or changed inputs. Success can justify only the next authorised boundary. If costs are incomplete, propose instrumentation rather than inventing ROI. If sensitive-data or approval gates fail, favourable economics do not override them.
Return the accounting table, assumptions, observed-versus-forecast label, quality evidence, bounded decision and missing costs. Keep measured examples synthetic when they are fixtures; a model evaluation does not prove production labour savings.
Provenance
Original distillation; source restrictions retained. No action authority is granted by these instructions.
