Who owns the conclusion? Ex ante verification agreements as an accountability mechanism for human and AI work in project governance
Abstract
Generative artificial intelligence (AI) has made reports, analyses and presentations nearly free to produce, while the effort required to verify them has not declined. In project and portfolio governance, this asymmetry shifts work from authors to reviewers and blurs accountability for conclusions: the sender relies on one model, the recipient on another, and errors pass between them in fluent prose. This paper asks how organizations should assign accountability for conclusions and design verification when work products are produced by people and AI agents. Using a conceptual, theory-building approach grounded in documentary cases from IT portfolio governance in a large energy company, public cases and illustrative vignettes, the paper develops the ex ante verification agreement (EVA): a specification, attached to a task before work begins, of the claims the work product must support, the errors the acceptor considers material, the checks that will detect them, and who applies the checks and decides acceptance. Four principles follow: the sender owns the facts, the inference and the clarity; checks are agreed with the task and matched to the feared error; checks are owned by the acceptor and cannot be changed by the executor, human or AI; and checks are themselves tested by seeding the feared error. A typology of five work-product errors (omission, substitution, inference error, fabrication and check circumvention) shows that each check detects only the error class it was designed for. Six propositions are formulated for empirical testing.
Keywords: generative AI, verification, accountability, project governance, separation of duties, automation bias, AI agents, principal-agent theory
Declarations
Data availability
No new datasets were generated or analyzed in this study. The documentary cases draw on internal company documents that cannot be shared for reasons of confidentiality; their relevant content is reported in the text. Public sources are listed in the references.
Ethics statement
This study did not involve human participants, animals or personal data and did not require ethics approval.
Author contributions
Vladimir Gaylun: Conceptualization, Methodology, Investigation, Formal analysis, Visualization, Writing – original draft, Writing – review & editing.
Funding
This research received no external funding.
Competing interests
The author is Head of Product at a U.S. company that develops AI systems for legal practice. The company did not fund this research or influence its conclusions.
