FollowThrough / permissioned proposal-owner pilot Status: study plan only. No interviews, customers or ROI results collected. Recruit 3-5 people responsible for agency proposals. Obtain their permission to observe the session and separately obtain permission before sending any real source material to Nebius. Prefer a sanitized brief. Never recruit or email participants automatically from this repository. Give each participant a brief, an evidence pack and a later amendment. Use two comparable fictional packs and alternate the order of manual and assisted review to reduce practice effects. Do not reuse the model development examples as an independent evaluation set. Have a second human label obligations, amendment effects and acceptable evidence before the session. Record, for both manual and assisted review: - Elapsed time from opening the brief to the participant's final reviewed pack. - Mandatory obligations missed, unsupported claims, missing attachments and incorrect deadlines, checked against the independent labels. - Amendment effects identified, false alarms and residual mistakes at export. - Number of interventions, clarification questions and accepted exceptions. - Model token usage, failures and provider latency, plus actual invoiced cost if available. Do not call token estimates actual currency expenditure. - Whether the participant could explain why approval was blocked/reopened. Ask: What would stop you using this on your next proposal? Which finding was useful, misleading or missing? Would this change a real review decision? Ask for the participant's normal workflow before introducing feature ideas. Publish every participant's paired measurements anonymously only with consent. Report median and range as exploratory results with this small sample. Do not extrapolate to contract wins, revenue, legal compliance or production accuracy. Record limitations and negative findings alongside benefits. Worksheet columns: participant_alias, consent_recorded, scenario, method, order, minutes, mandatory_missed, unsupported_claims, attachments_missing, deadline_errors, amendments_missed, false_alarms, model_failures, provider_ms, input_tokens, output_tokens, actual_cost_if_available, clarification_count, qualitative_notes