Verified Agentic Tasks for Business Process Engines: An Add-On with Deterministic Guarantees

Business Process Management (BPM) engines provide reliable process orchestration, durable execution, and human oversight, but do not inherently guarantee the semantic correctness of individual service-task outputs. This limitation becomes particularly significant when integrating Large Language Model (LLM) agents into business workflows, where plausible but incorrect results can propagate into downstream decisions. This paper presents an implemented agentic BPM architecture that integrates AI-driven cognitive task execution with deterministic, neurosymbolic verification. Its central component, VERA, is an LLM-independent verification gate that evaluates structured outputs against business invariants using exact arithmetic, fail-closed execution semantics, and bounded minimal-conflict diagnosis. The verification mechanism provides a formally defined, one-sided guarantee: detected inconsistencies violate explicitly established invariants beyond their stated tolerances, while consistency alone does not establish factual correctness. The architecture separates cognitive reasoning from deterministic verification and process execution, enabling AI agents to participate in governed business workflows. It supports integration with established BPM platforms and includes a standalone execution engine with durable process instances, human approval workflows, scheduling, parallel execution, and BPMN 2.0 interoperability. The documented evaluation includes 361 workflow and BPMN tests, 371 verification tests, five end-to-end reference processes, and an ablation study involving 500 synthetic business documents. Within the evaluated invariant-violating perturbations, VERA achieved complete detection, with a median verification overhead of approximately 6 milliseconds per document. Invariant-preserving errors remain an explicitly identified limitation. The contribution establishes an architectural and computational foundation for incorporating verifiable AI-generated outputs into business process automation, combining the flexibility of agentic systems with the operational governance and deterministic controls of traditional BPM engines.