An Agentic AI Framework for Master Data Synchronization, Validation and Self-Healing Across Integrated Healthcare Information Systems

Background. Large healthcare organizations run many specialty applications, such as prescribing and dispensing, warehouse and inventory, laboratory, blood bank and billing, that depend on shared master data, chiefly the patient identifier (MRN) and the product identifier with its description. Standards such as HL7 v2 and FHIR can propagate a new or changed record from an origin system, but propagation failures (rejections, unmapped codes, missing mandatory fields, duplicates and timing dependencies) still rely on manual diagnosis and repair. The resulting gaps surface as held or failed transactions at the point of care and supply. Failures also arise without any interface error, when a record exists but its local build or configuration (for example location, charge or dispensing parameters) is incomplete. Objective. To specify a human-governed, closed-loop framework in which AI agents act as integration architect, builder, verifier and support engineer: designing field mappings, configuring receiving systems against an approved build standard, verifying cross-system data validity and local build completeness, and diagnosing and remediating propagation errors. Methods. The study followed a design-science approach. The synchronization gap is formalized, a seven-class failure taxonomy is derived, and a four-layer architecture (sense, reason, act, govern) is specified with tiered autonomy grounded in the Parasuraman-Sheridan-Wickens model of types and levels of automation. An explicit tier-decision rule and tool contracts are defined, verification checks are mapped to established electronic health record data quality frameworks, the closed loop is illustrated with a synthetic worked example, and a staged evaluation protocol with quantified safety gates is specified. Results. The framework specifies agent roles and tool contracts; a proactive build-completeness check for local configuration; a default and maximum autonomy tier for each of seven failure classes; a decision rule in which the permitted tier is the minimum of a policy tier, a confidence tier and a criticality cap; safeguards (patient identity actions reserved for humans, grounding in versioned build standards, deterministic pre-execution validation, and treatment of message content as untrusted data) mapped to a threat model; and evaluation metrics (time to consistency, held-transaction rate, false-fix rate, harmful-action rate) with the sample sizes needed to demonstrate safety bounds, for example 149 consecutive correct fixes to show with 95% confidence that the false-fix rate is below 2%. This is a conceptual and design paper; no empirical results are reported. The tier-decision rule was checked exhaustively against ten stated properties over 11,440 input combinations, with no violations, and six deliberately seeded faults were each detected (Section IV-I). Conclusions. The framework proposes that an AI agent layer can be added over existing integrated systems without replacing them, turning error queues from passive backlogs into a supervised remediation loop. Whether it achieves this is untested. The paper supplies testable hypotheses, pre-specified safety gates and a staged path from shadow mode to limited autonomy for future empirical validation.