Before You Build AI in Healthcare: Nail the Objective and the Data First
Here's the simple truth: AI won't fix a vague problem or messy data. That's the core message emphasized by Rob Havasy of HIMSS and Karla Eidem of PMI. Before you code a model or sign a vendor contract, set a clear outcome and stand up strong data governance.
Do that, and AI becomes useful. Skip it, and you'll collect "pilots" that never move the needle.
Start with the problem, then the metric
Define the clinical or operational gap first. Make it specific, measurable, and time-bound. If you can't quantify success, you can't manage it.
- Clinical: reduce 30-day readmissions by 10% for CHF patients in 9 months
- Quality: cut time-to-antibiotics in sepsis by 15 minutes in the ED
- Operations: decrease prior auth turnaround by 25% without hurting denial rates
- Workforce: save clinicians 2 hours per week on documentation while keeping note accuracy
Also define the user and the moment in workflow. Who acts on the output? In which screen? What is the default if the model is silent or wrong?
Governance that prevents regret
Strong governance isn't paperwork. It's how you avoid rework, safety events, and patient trust issues. Make these non-negotiable:
- Data stewardship: ownership, access, lineage, and approval paths clearly defined
- Privacy and security: enforce least-privilege, full audit trails, and breach response plans
- PHI controls: document use cases, retention, encryption, and de-identification standards
- Bias and equity review: track performance by subgroup; document mitigations before go-live
- Clinical safety: prospective validation, human-in-the-loop, and rollback plans
- Lifecycle: monitoring for drift, periodic re-validation, and sunset criteria
For structured frameworks, many teams align policies with the NIST AI Risk Management Framework and HIPAA Security Rule.
Make data warehousing work for AI, not against it
Your model is only as good as your data layer. Build a path from source systems to actionable features with quality checks at every hop.
- Single source of truth: standard vocabularies, version control, and business definitions
- Interoperability: consistent mapping of EHR, claims, device, and patient-generated data
- Data contracts: schemas, freshness SLAs, and fail-fast alerts when feeds break
- Feature store: reusable features (e.g., risk scores, recent labs) with lineage and owners
- Quality gates: completeness, timeliness, and anomaly detection before model scoring
If your data warehouse can't provide reliable, timely inputs, pause the AI build. Fix the plumbing first.
Procure or build with clear gates
Whether you build or buy, run candidates through the same filters:
- Evidence: peer-reviewed results or internal validation on your population and workflows
- Security posture: third-party assessments, breach history, and response commitments
- Integration: EHR-native workflow, APIs, data export, and alert fatigue safeguards
- Contracts: data use boundaries, model transparency, IP, retention, and termination terms
- Total cost: implementation, change management, monitoring, and ongoing tuning
Pilot to production without drama
Keep pilots tight and accountable. Treat them like experiments with graduation criteria.
- Run in a sandbox first; confirm data quality and latency
- Shadow mode to baseline performance and false alerts
- Train end users; adjust UX based on real actions, not opinions
- Define go/no-go thresholds; if passed, scale with monitoring and a rollback switch
People and roles you need
AI programs stall when roles are fuzzy. Assign clear owners.
- Executive sponsor: removes roadblocks and owns funding
- Clinical owner: defines use case, safety, and adoption
- Data steward: ensures data quality and lineage
- Privacy and security: approves data flows and access
- MLOps/IT: deployment, monitoring, and incident response
- Vendor manager: performance reviews and contract compliance
A quick checklist to pressure-test readiness
- We have a single, measurable objective and a baseline
- We know exactly who will use the output and where
- Data is mapped, governed, and monitored end-to-end
- Privacy, security, and bias controls are documented and tested
- We can validate safety before exposure to patients
- We have post-go-live monitoring and a rollback plan
- We know total cost and who owns each task
Bottom line
AI should come after clear outcomes and dependable data. That's the fastest way to real clinical impact and sustainable adoption-exactly what leaders like Havasy and Eidem are pushing for. Set the objective, fortify the data, then build.
Need structured upskilling for your team?
If your leaders, clinicians, or data teams need practical AI training aligned to job roles, explore these resources:
Your membership also unlocks:
