POLARIS: Typed Planning and Governed Execution for Agentic AI in Back-Office Automation

TL;DR

POLARIS tackles the gap in enterprise back-office automation by introducing a typed planning and governed execution framework for agentic AI. It uses a plan–select–act loop where CoAPlanner proposes diverse, type-checked DAGs, ReasoningAgent selects a compliant plan via a rubric, and Guarded Execution with a validator-gated repair loop ensures safe, auditable side effects before any action. Policy guardrails, anomaly handling, and complete audit traces yield decision-grade artifacts suitable for regulated environments, demonstrated on synthetic invoice tasks and the SROIE dataset with strong extraction performance and precise anomaly routing. The work provides a benchmark-oriented reference for policy-aligned Agentic AI, offering a methodology and evaluation suite that can extend to other regulated domains beyond finance, such as supply chain and compliance.

Abstract

Enterprise back office workflows require agentic systems that are auditable, policy-aligned, and operationally predictable, capabilities that generic multi-agent setups often fail to deliver. We present POLARIS (Policy-Aware LLM Agentic Reasoning for Integrated Systems), a governed orchestration framework that treats automation as typed plan synthesis and validated execution over LLM agents. A planner proposes structurally diverse, type checked directed acyclic graphs (DAGs), a rubric guided reasoning module selects a single compliant plan, and execution is guarded by validator gated checks, a bounded repair loop, and compiled policy guardrails that block or route side effects before they occur. Applied to document centric finance tasks, POLARIS produces decision grade artifacts and full execution traces while reducing human intervention. Empirically, POLARIS achieves a micro F1 of 0.81 on the SROIE dataset and, on a controlled synthetic suite, achieves 0.95 to 1.00 precision for anomaly routing with preserved audit trails. These evaluations constitute an initial benchmark for governed Agentic AI. POLARIS provides a methodological and benchmark reference for policy-aligned Agentic AI. Keywords Agentic AI, Enterprise Automation, Back-Office Tasks, Benchmarks, Governance, Typed Planning, Evaluation

Figures (4)

• Figure 1: Overview of POLARIS orchestration with governed plan selection. Raw inputs (prompts/triggers/PDFs) are normalized by the InputNormalizerAgent. The CoAPlannerAgent proposes multiple structured execution plans; a Reasoning Agent scores them and selects a single plan. The chosen plan executes as a DAG: a DocumentParserAgent and DataValidatorAgent run with a bounded repair loop (orange) to improve extractions, after which only the required Data Processor Agents are invoked (e.g., RecordMatching, PolicyRetrieval, AnomalyDetection, RiskControl, APIAccess). Reconciliation Agents (Approval, ReportGenerator) produce decision-grade outputs (Structured Report, Approval/Decision Action, Inference/Alert Result). Dashed lines indicate data/control flow across stages. DSPy few-shot learning supplies planning priors.
• Figure 2: Dependency-aware parallel scheduler. The scheduler converts the planner’s ordered list of agents into a DAG that enables safe parallelism. DocumentParser runs first; middle-stage checks run concurrently; sinks wait for all upstream checks.
• Figure 3: Extraction performance by scenario for POLARIS. Grouped bars show Precision, Recall, and F1 across VU (Unknown Vendor), VL (Layout Drift/Noise), CC (Clean), and CM (Month-End). The plot highlights high recall under noise (VL) with a precision dip, and strong balance on clean/compliant settings.
• Figure 4: POLARIS: Inputs/Outputs, DSPy plan trace, and validator-gated repair loop (synthetic negative scenario VL). Left: a low-quality invoice is normalized by the InputNormalizer into a canonical JSON record. Middle–left: the CoAPlanner proposes multiple typed plans; the ReasoningAgent selects one (shown) that executes DocumentParser$\rightarrow$DataValidator$\rightarrow$PolicyRetrieval$\rightarrow$Approval. Middle–right: a bounded parser$\leftrightarrow$validator loop issues targeted feedback (e.g., fix missing/invalid invoice number, date, vendor; normalize currency/dates) until validation passes or the attempt budget is exhausted. Right: reconciliation shows a policy retrieval error due to missing/invalid parsed fields; the ApprovalAgent returns approved: false with a structured rationale and final output object.