Reconstruction Is Not the Same as Explainability
Model explainability describes how an AI system generally reaches outputs, using techniques such as feature weights or decision boundaries. It answers why a model tends to behave a certain way. Reconstruction is a different requirement: it asks what specifically happened during one historical decision. For a disputed prior authorization or claims adjudication, an auditor or appeals reviewer does not need a general theory of model behavior. They need the exact model version, the exact prompt or context, the exact tool calls, and the exact data snapshot used for that single decision.
Coverage determinations increasingly involve multi-step agentic workflows, including policy lookups, retrieval steps, and calls to external tools, rather than a single model inference. A general explainability report on the underlying model cannot account for what a specific agent pipeline actually did at a specific point in time. Compliance and audit functions require reconstruction because their obligation is to defend an individual decision, not to characterize a model's overall tendencies.
Explainability characterizes general model behavior. Reconstruction produces evidence for one historical execution: model version, context, tool calls, policy version, and data snapshot.
Regulatory Drivers Behind Reconstruction Requirements
Several regulatory and oversight frameworks now expect payers to account for how specific AI-driven decisions were reached. CMS's Interoperability and Prior Authorization Final Rule requires impacted payers to provide specific reasons for prior authorization denials and to support standardized, API-based workflows, with compliance obligations phasing through 2026 and 2027. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, directs state regulators to expect insurers to maintain governance documentation covering AI data inputs, decision logic, and ongoing monitoring sufficient for regulatory review.
The EU AI Act classifies AI systems used to evaluate eligibility for health and life insurance as high-risk, subjecting them to record-keeping obligations that require automatic logging capable of tracing system functioning across its lifecycle. ONC's HTI-1 Final Rule requires certified health IT incorporating predictive decision support to make available source attributes describing data sources, development, and intended use. NIST's AI Risk Management Framework identifies documentation and traceability of inputs, outputs, and decision context as core risk-management functions applicable to healthcare. State regulators, including Colorado's Division of Insurance, have advanced rules requiring insurers to test, document, and explain algorithms used in underwriting and claims-related decisions.
None of these frameworks are satisfied by a model card or a general explainability statement. Each presumes the payer can produce evidence tied to a specific decision.
Implementation Tradeoffs
Reconstruction capability introduces real operational tradeoffs that governance leaders should weigh explicitly. Log retention periods must align with applicable appeal and regulatory examination windows, which often extend well beyond default system log retention settings. Reconstruction obligations extend to third-party or subcontracted AI tools embedded in a payer's workflow, not only internally built systems, which means vendor logging practices need scrutiny during procurement.
When multiple vendors, models, or agent frameworks participate in a single determination pipeline, standardized log schemas become necessary to avoid fragmented or inconsistent records that cannot be correlated. Comprehensive tool-call and data lineage logging also carries performance and storage overhead that should be evaluated against existing claims and utilization management system throughput requirements before broad deployment. None of these tradeoffs eliminate the need for reconstruction; they determine how much engineering and governance work is required to make it reliable at scale.