Delegated Agency Across the Venture Cycle: The Bright and Dark Sides of Agentic AI in Entrepreneurial Ventures

Description of the Research Data Workbook File: Agentic_AI_Case_Study_Dataset.xlsx Companion to: “Delegated Agency Across the Venture Cycle: The Bright and Dark Sides of Agentic AI in Entrepreneurial Ventures” (JEEE Special Issue submission) Compiled: 10 November 2026 Purpose The workbook is the complete replication dataset for the study’s comparative multiple-case analysis of seven agentic-AI ventures. Every descriptive figure reported in Section 4 of the manuscript is computed directly from this file, either through built-in formulas or via the accompanying replication script. Sheet-by-sheet contents Sheet Contents Cover Study metadata: design, observation window (Jan 2022 – Nov 2026), search period (28 Oct – 10 Nov 2026), corpus size (35 documents), screening totals, coding reliability, and the evidence-tier definitions. Case_Inventory Profiles of the seven included cases (Cognition/Devin, Harvey, Sierra, 11x, Rabbit, Adept, Vambe): HQ, founding year, venture-cycle stage, agentic offering, reported funding and valuation, key documented outcome facts, and source name/URL for every figure. Screening_Log All 27 screened candidates with the inclusion/exclusion/borderline decision, the primary reason, and an evidence-sufficiency rating — the audit trail behind Appendix B of the manuscript. Evidence_Log The full 35-document corpus: issuer, date, source type, evidence tier (A–D), the specific claim or event captured, and URL for each item. Coding_Matrix The case-by-code matrix: six agentic features (F1–F6), human-oversight level, five bright-side and six dark-side mechanism codes, outcome type, and venture status. Feature-count and dark-code-count cells are live formulas. CrossCase_Summary Formula-driven descriptive summaries reproducing the manuscript’s Section 4 tables (corpus composition by tier, cases by venture-cycle stage, feature prevalence, oversight vs dark-side codes, mechanism frequencies), with three embedded charts. All counts recalculate automatically if the coding matrix is edited. Search_Log The seven documented searches (S1–S7): platform, date, query string, filters, and records identified/screened/full-text assessed, with formula totals (214 / 96 / 58). How it connects to the replication code The plain-text replication file contains a Python script that reads this workbook and prints every Section 4 table (screening outcome, tier composition, stage distribution, feature prevalence, mechanism frequencies, and the oversight gradient). No figure in the manuscript is hard-coded; all derive from the Evidence_Log and Coding_Matrix sheets. Data-integrity note All funding, valuation, revenue and outcome figures are transcribed from the public sources listed in each row, with access date 10 November 2026. Company-reported figures (Tier C) and analyst aggregations (Tier D) are labelled as such and are not treated as independently verified facts.