Answers are grounded in this paper and cite their sources.
Intelligent agents are now used in smart-grid and distributed-energy enterprises to automate predictive maintenance scheduling, control-loop parameter inspection, dispatch data querying, and cross-department information sharing. When an agent takes any of these actions, it has to access supervisory control and data acquisition (SCADA) records, fault logs, and control-loop parameters across organizational boundaries—which is exactly where the compliance issue appears. The agent’s access decision has to obey mandatory-deny rules coming from sector-specific regulation and internal governance, without exception, and a purely learned classifier is not enough. Three concrete gaps remain in existing permission models. The first is that no deterministic guarantee is enforced on the mandatory-deny rule set. The second is that the business context is represented in a flat form, so the relational dependencies among tasks, tools, personnel roles, and data types are lost. The third is that the false positive rate on high-variance domains stays high— $$16.7%$$ for the strongest prior hybrid. In order to close the three gaps together, this paper formulates the Compliance-Aware Permission Prediction (CAPP) task and proposes an implementation named NCF-CG. Its learned branch is trained end-to-end and jointly uses a neural collaborative filtering preference estimator, a graph attention encoder over a Business Knowledge Graph, and a compliance-margin loss on forbidden pairs. On top of this branch, a hard compliance gate overrides the learned score at inference time whenever a mandatory rule fires. On a public benchmark of 7563 permission decisions, NCF-CG reaches $$89.20%$$ full-coverage accuracy and $$11.73%$$ false positive rate, surpassing all seven baselines (including four modern deep-learning models) under a paired participant-level t-test ( p < 10⁻¹² against every baseline). On the subset of decisions the gate never touches it still leads the prior hybrid by 5.05 percentage points, and a cross-domain transfer probe on 500 expert-reviewed energy-production vignettes returns $$74.5%$$ accuracy on the gate-untouched subset (the subset where domain shift can affect the model output). The framework gives an auditable modelling pattern for AI-agent access control in smart-grid enterprises: deterministic where the rules require it, and learned everywhere else.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: