Verifiable Transaction-State Reconciliation for AI-Orchestrated Financial Workflows Under Partial Failure and Duplicate Events

Abstract

AI agents are increasingly tasked with orchestrating multi-step financial workflows that span internal services and external processors. These workflows are exposed to partial failures and duplicate events caused by retries and at-least-once messaging. This paper describes a reconciliation protocol that pairs a deterministic transaction state machine with idempotency controls, compensating actions, and a hash-chain ledger supporting Merkle proofs. On a synthetic workload of 50,000 workflow instances under controlled failure injection, the design reduced unresolved states and improved consistency relative to a simple retry baseline and a saga-only approach, at the cost of modest added latency. The resulting evidence chain is intended to support operational audit in regulated environments.