Data Foundations

0/4

Clear ownership for core datasets (clinical, operational, financial)Without ownership, data quality degrades and no one is accountable for fixes.

Consistent definitions for key business metricsInconsistent metrics lead to conflicting AI outputs and eroded trust.

Data quality thresholds defined and monitoredAI models amplify data quality issues — garbage in, garbage out at scale.

Known sources of truth (no duplicate "gold" datasets)Multiple "sources of truth" create conflicting model training data.

Data Integration & Interoperability

0/4

Reliable pipelines across core systems (EHR, claims, ops)Broken pipelines mean AI models train on stale or incomplete data.

Standardized identifiers and entity resolutionWithout consistent IDs, you can't link patient data across systems.

Interoperability standards in place (FHIR / HL7 where applicable)Standards reduce integration cost and improve data portability.

Latency and freshness aligned with AI use casesReal-time AI needs real-time data; batch models need consistent snapshots.

0/4

Scalable, cloud-native or hybrid data platformAI workloads require elastic compute that on-prem often can't provide.

Separation of raw, curated, and feature-ready dataMedallion architecture enables reproducibility and audit trails.

Ability to support batch + near-real-time workloadsDifferent AI use cases require different processing patterns.

Clear lineage from source → model → outputLineage is essential for debugging, compliance, and model governance.

0/4

Data access controls aligned to least-privilege principlesAI systems shouldn't have broader access than necessary.

Auditability for data usage and AI outputsRegulators and stakeholders will ask how decisions were made.

Compliance requirements mapped (HIPAA, regulatory, internal policies)AI doesn't get a compliance exemption — if anything, scrutiny increases.

Clear rules for PHI / PII handling in AI workflowsMishandling sensitive data in AI can lead to breaches and fines.

0/4

Defined deployment model (cloud, VPC, on-prem)Deployment constraints affect model architecture and latency.

Monitoring for data drift and model behaviorModels degrade over time — you need to detect it before users do.

Human-in-the-loop processes where requiredHigh-stakes decisions need human oversight, especially in healthcare.

Clear ownership for AI systems in productionSomeone needs to be on call when the model misbehaves at 2am.

0/4

AI use cases tied to measurable outcomesAI for AI's sake doesn't survive budget reviews.

Clear success metrics beyond "model accuracy"Business impact matters more than F1 scores.

Stakeholders aligned across product, engineering, and operationsSiloed AI projects fail when they hit production reality.

Roadmap from pilot → production → scaleWithout a path forward, pilots become permanent experiments.