When Not to Automate: A Formal Protocol for Human Preservation in AI-Optimized Organizations
Citation. Rodríguez Arias, J., Chouliaras, S., & de la Chica, J. M. (2026). When not to automate: A formal protocol for human preservation in AI-optimized organizations. arXiv. https://arxiv.org/abs/2607.15944
Summary. This preprint proposes PHP-AIO, a deterministic, five-gate protocol for making pre-deployment automation decisions at the role level. The protocol scores tacit-knowledge erosion, resilience reduction, regulatory exposure, and socio-institutional capital degradation, then returns automate, augment, hybrid, or preserve with prescribed governance actions. Its examples and parameters are currently financial-services-oriented and expert-elicited rather than outcome-validated.
Key claims
- Standard automation ROI can omit long-horizon organizational risks from lost tacit knowledge, correlated automation failure, prospective regulatory constraints, and degraded trust or institutional capital.
- PHP-AIO proposes a pre-decision gate sequence and a process-level “automation debt” measure so that automation, augmentation, hybridization, and human preservation become auditable choices rather than cost-only defaults.
- In stylized role profiles, the framework produces different recommendations from a uniform cost-benefit automation rule; the authors report sensitivity to upward perturbations of at least 14% for three of four representative gate decisions.
- A formal human-in-the-loop anchor is presented as necessary but not sufficient: the relevant human capability, authority, and preserved practice must remain available for oversight and recovery to be meaningful.
Evidence & limitations
- Preprint (arXiv:2607.15944v1, July 2026) from Santander AI Lab; the Drive PDF exactly matches the public arXiv PDF by SHA-256.
- This is a conceptual protocol with expert-elicited weights, thresholds, country multipliers, and drift rates; its examples are stylized profiles rather than observed deployment outcomes.
- The authors report that retrospective validation, input-level inter-rater reliability studies, and cross-industry calibration remain underway; deterministic scoring does not resolve disagreement in the human judgments used as inputs.
- The arXiv API returned HTTP 429 during preflight, so title, author, date, category, and public provenance were verified from the public arXiv PDF and the exact binary hash match rather than API metadata.