Agentic Corporate Actions Processing

The platform proposes; it never decides. An analyst confirms, overrides, or escalates at every decision point, and no number is ever computed by a model.

**Multi-Agent Decision-Support Framework:**

A supervised, checkpointed agent graph moves each event through confidence-gated stages: interpret, calculate, elect, reconcile, diagnose, and explain. Foundational agents handle ingestion and extraction, source reconciliation, validation, anomaly detection, and explanation, while corporate-actions specialists cover event classification, entitlement orchestration, election management, and break diagnosis.

Critically, the reasoning agents call out to deterministic tools—a calculation engine, withholding-tax engine, FX conversion, and fractional-share logic—rather than computing anything themselves. Every figure carries a complete calculation trace showing formula, inputs, and rule applied. Confidence gates govern each stage: high-confidence results auto-pass, medium is flagged, and low confidence or conflicting sources block and require a human decision. Where sources disagree on a date, rate, or ratio, both values are shown side by side until an analyst arbitrates.

**Business Capabilities:**

• Interpret issuer and custodian notices from PDFs, scans, emails, and message feeds into structured, provenance-linked records
• Grade every extracted field by confidence, blocking low-confidence or conflicting values for human decision
• Calculate entitlements deterministically across dividends, splits, mergers, spin-offs, and rights issues
• Apply per-account withholding tax, multi-currency FX, and fractional-share handling with full calculation traces
• Manage the election lifecycle with eligible-account notification, live response tracking, and deadline escalations
• Resolve non-responders to documented defaults and close events with audit-ready aggregates
• Detect post-pay-date breaks and diagnose ranked, evidence-backed root causes
• Capture analyst corrections as a feedback loop that informs future break diagnosis

**AWS-Native Architecture:**

The platform is designed to run entirely within a firm's AWS environment, deployed alongside the existing corporate actions system rather than inside it. Notice ingestion begins with Amazon Textract, which converts issuer and custodian documents into structured fields; Amazon Bedrock, with guardrails enforced, then reasons over those fields to classify events and interpret terms. Amazon Bedrock AgentCore and AWS Step Functions maintain the supervised agent graph, holding each event at its confidence gate until an analyst decision releases it.

Deterministic calculation is kept firmly outside the model boundary: entitlement, withholding-tax, FX, and fractional-share engines run as AWS Lambda functions invoked as tools, so every figure originates from code rather than inference. Amazon Aurora PostgreSQL holds the checkpointed run state and append-only audit log, making any event resumable and any decision reconstructable, while Amazon OpenSearch Service retrieves comparable historical events and prior analyst corrections as evidence for break diagnosis. Notices and generated explanation reports are retained in Amazon Simple Storage Service (Amazon S3), and custodian, reference-data, and market feeds arrive through Amazon API Gateway, Amazon EventBridge, and AWS Glue.

Access is scoped by role through AWS Identity and Access Management (IAM), with encryption under AWS Key Management Service (AWS KMS). Amazon CloudWatch and AWS CloudTrail trace every agent action against a shared correlation identifier, and AWS Config, AWS Security Hub, and AWS Control Tower extend governance across accounts.