Delegated agency across the venture cycle: the bright and dark sides of agentic AI in entrepreneurial ventures

Replication package guide: dataset and code for "Delegated agency across the venture cycle" 1. Overview The replication package accompanies the manuscript "Delegated agency across the venture cycle: the bright and dark sides of agentic AI in entrepreneurial ventures" and consists of two files: the workbook Agentic_AI_Panel_Dataset.xlsx (all data, coding, evidence logs and headline results) and the script Analysis_Replication.txt (plain-text Python that reproduces every headline statistic from the workbook). A versioned copy of both is deposited at https://doi.org/10.5281/zenodo.24112346. No data are fabricated: every factual record carries a source URL, an evidence tier (A–D) and an access date of 10 November 2026. 2. The dataset workbook (Agentic_AI_Panel_Dataset.xlsx) The workbook contains 18 sheets. Core data sheets hold the firm master (60 ventures: 8 verified adopters, 32 claim-only adopters, 20 matched controls), the three treatment dates per venture (announced, available, verified deployment), 115 funding events totalling US$21,327.3 million, venture status histories, five verified governance incidents, and the two-pass, seven-dimension oversight rubric scores (140 dimension scores; Krippendorff ordinal α = 0.792). The Firm_Period_Panel sheet is the analysis dataset: 512 venture-half-years (2022H1–2026H2) in which funding per period, the binary funding indicator, log funding, event time, and the two post-treatment flags are computed by live formulas (SUMIFS against the Funding_Events sheet), so every panel value can be traced to its underlying dated events. Audit sheets preserve the evidence trail: an Evidence_Log of 40 sourced items, a Screening_Log of 27 excluded candidates with reasons, a Search_Log, the coding matrix, the seven process-tracing chains, headline analysis results, the power analysis, and a missingness/coverage report. Table 1. Workbook sheets at a glance Sheet Contents Cover, Data_Dictionary Version, citation, and variable definitions for every sheet Firm_Master 60 ventures: group, region, sector, founding, status, evidence tier Agent_Adoption Three dated adoption events per venture with evidence tiers Funding_Events 115 dated rounds (US$M) with source URLs Firm_Period_Panel 512 venture-half-years; formula-computed outcomes and treatment flags Status_Events, Governance_Incidents Exit/restructuring histories; five verified incidents with severity Oversight_Rubric, Oversight_Analysis Two-pass scores, α = 0.792; exposure-adjusted event rates Evidence_Log, Screening_Log, Search_Log 40 sourced items (A–D tiers); 27 exclusions; query audit trail Coding_Matrix, Process_Tracing Dimension-level scores; seven dated causal chains Analysis_Results, Power_Analysis All headline statistics (R0–R11); MDES and cluster requirements Missingness_Coverage Coverage and missing-data accounting Note: panel outcomes are live formulas; editing a funding event updates the panel and results on recalculation. 3. The replication code (Analysis_Replication.txt) The replication file is a self-contained Python script in plain-text format. It reads the workbook directly, rebuilds the firm-half-year panel from Funding_Events and Firm_Master (independently of the workbook's formulas, using the same half-year conventions), and prints results blocks R0–R11. Running it end-to-end reproduces every headline number reported in the manuscript and in the workbook's Analysis_Results sheet. ● R0–R2: sample construction, coverage, and the primary verified-adoption difference-in-differences (β = 1.630, cluster-robust p = 0.046; randomization-inference p = 0.010, 500 permutations, fixed seed). ● R3–R6: any-adoption specification (+83%), announcement-date timing comparison (+550%), placebo test (p = 0.130), and the event study with leads, lags and the joint pre-trend test. ● R7–R9: oversight reliability (α = 0.792), oversight–severity correlations with exposure-quarter denominators, and adverse-exit comparisons (Fisher exact tests). ● R10–R11: covariate balance (standardized mean differences) and the power/MDES analysis that frames the estimates as exploratory. Requirements are minimal: Python 3 with pandas, numpy, scipy and statsmodels, plus openpyxl to read the workbook. To replicate, save the file as replication.py in a folder containing Agentic_AI_Panel_Dataset.xlsx and run: python3 replication.py. All randomness is seeded, so output is deterministic. 4. How the pieces fit together The workbook is the source of record; the script is the independent check. The workbook stores each observation once with its source and tier; the script reconstructs the analysis sample from the raw event sheets rather than trusting the pre-computed panel, so any discrepancy between formula-computed and script-computed values would surface immediately. Manuscript tables I–IV and Appendix C are drawn exclusively from the Analysis_Results sheet, which the replication script regenerates in full.