Agentic Medication Reconciliation Across Transitions of Care

1. Problem / Context

Medication reconciliation across admissions, transfers, and discharge is one of the most error-prone moments in care. Lists arrive incomplete, duplicative, or outdated. Pharmacists spend valuable time calling community pharmacies, paging clinicians, and comparing EHR lists to patient-reported histories. In mid-market health systems, lean pharmacy teams and variable data quality create a persistent safety and efficiency gap. The result: delays in first dose, adverse drug events from omissions or duplications, and hard-to-audit processes that strain compliance teams.

Agentic AI can change that standard. Instead of brittle scripts or manual hunts through multiple systems, agentic workflows ingest the right data, reason about conflicts, and coordinate the next best action—always with a human pharmacist in the loop and full auditability for regulators and internal quality teams.

2. Key Definitions & Concepts

• Medication reconciliation: A structured process to create the most accurate list of a patient’s medications and compare it against orders at each transition of care, resolving discrepancies.
• Agentic AI: A governed “sense–think–act” pattern where software agents read signals, use tools (APIs, knowledgebases), reason about options, and coordinate tasks—under explicit policies and human oversight.
• FHIR and ADT: The workflow ingests FHIR MedicationStatement and MedicationRequest resources, plus HL7 ADT events (admit, transfer, discharge) to trigger reconciliation at the right moments.
• eRx/PBM fill history: Recent dispensings from e-prescribing networks and pharmacy benefit managers add real-world adherence and active therapy signals.
• RxNorm/NDC normalization: Standard vocabularies to unify drug names, strengths, and formulations so reasoning and deduplication are reliable.
• Drug knowledgebase APIs: Clinical interaction, allergy, and contraindication checks that inform safe alternatives and discontinuations.
• Databricks Workflows: The orchestration layer that sequences data ingestion, reasoning, human-in-the-loop (HITL) review, and EHR write-back with retries and monitoring.

3. Why This Matters for Mid-Market Regulated Firms

Mid-market hospitals and integrated physician groups face the same safety and compliance expectations as large systems but without sprawling teams. Safety committees and pharmacy leadership need traceable, explainable reconciliation; compliance teams need audit trails for Joint Commission surveys and payer reviews; and operations leaders need measurable cycle-time and error-rate improvements. Budgets are tight, EHR integration will vary by site, and vendor lock-in must be avoided. An agentic approach balances safety and throughput: clinical reasoning to handle messy, incomplete data; human approvals where it matters; and governance primitives that make auditors comfortable.

Kriv AI, a governed AI and agentic automation partner for mid-market organizations, focuses on exactly these constraints—building workflows that are safe, auditable, and practical to operate with lean teams.

4. Practical Implementation Steps / Roadmap

1. Trigger points: Listen to ADT events to initiate med rec at admit, transfer, and discharge. Scope to target units (e.g., medicine floors) for an initial rollout.
2. Data ingestion: Pull FHIR MedicationStatement and MedicationRequest from the EHR. Enrich with recent eRx/PBM fill history to capture active community meds.
3. Normalization: Map all sources to RxNorm concepts, retain NDCs for specificity, and standardize dose/route/sig to a normalized schema.
4. Allergy and interaction context: Fetch allergy lists from the EHR and run prospective interaction checks via drug knowledgebase APIs.
5. Reasoning over conflicts: Use agentic logic to detect duplicates (generic/brand, strength variations), therapeutic duplications, and dosing conflicts; flag gaps (e.g., insulin type but no dose) and propose safe alternatives or discontinuations.
6. Human-in-the-loop review: Present a consolidated list in a pharmacist console that highlights evidence (last fill date, source provenance, interaction flags). The pharmacist edits, resolves conflicts, and approves.
7. EHR write-back and tasks: After approval, write updated orders or create tasks in the EHR, document reconciliation notes, and notify the care team.
8. Orchestration and reliability: Use Databricks Workflows to chain jobs, manage credentials, and implement retries/backoff on connector timeouts. Ensure idempotency to prevent duplicate orders.
9. Observability and audit: Persist inputs, decisions, and outputs in Delta tables with lineage and versioning; log who reviewed, what changed, and why.
10. Rollout and feedback: Start with one unit, compare baseline to post-implementation metrics, then expand coverage.

[IMAGE SLOT: agentic medication reconciliation workflow diagram showing EHR FHIR/ADT, eRx/PBM fill history, RxNorm normalization, drug knowledgebase APIs, pharmacist HITL console, and EHR write-back orchestrated by Databricks Workflows]

5. Governance, Compliance & Risk Controls Needed

• PHI policies and access: Apply Unity Catalog policies to restrict PHI access by role, mask fields in non-production, and separate secrets from code.
• Lineage and time travel: Use Delta lineage to trace every source (FHIR, PBM) and every decision (dedup, alternative suggested), with versioned snapshots for reproducibility.
• Approval gates: Route model, rule, and terminology updates through change control with clinical sign-off. Implement safe rollback for rule changes.
• Human oversight: Enforce required pharmacist approval for any medication list change; sample post-implementation audits for quality.
• Vendor lock-in avoidance: Favor standards (FHIR, RxNorm, Delta) and portable orchestration patterns to prevent dependency on a single EHR UI workflow.
• Reliability controls: Retries/backoff on integrations, circuit breakers for downstream outages, and queueing for after-hours admissions.

[IMAGE SLOT: governance and compliance control map highlighting Unity Catalog PHI policies, Delta lineage, approval gates, and human-in-the-loop checkpoints]

Kriv AI often helps clients set these foundations—data readiness, terminology services, MLOps pathways, and governance artifacts—so pharmacy leaders can trust what’s in production without slowing operations.

6. ROI & Metrics

Executives should insist on a tight measurement plan from day one:

• Cycle time: Minutes from ADT trigger to pharmacist approval; target 30–50% reduction.
• Pharmacist time saved: Minutes saved per reconciliation multiplied by daily volume, converted to FTE capacity.
• Discrepancy detection rate: Omissions/duplications identified per patient; intervention acceptance rate by clinicians.
• First-dose timeliness: Time to first inpatient dose for high-risk meds.
• Safety outcomes: Reduction in reconciliation-related errors and near-misses; fewer ADEs tied to omissions/duplications.
• Audit completeness: Percent of reconciliations with full provenance and approval trails captured.

A realistic example: a 250-bed regional hospital with ~80 admissions/day, 60 requiring med rec. Baseline pharmacist time averages 25 minutes per patient. An agentic workflow that cuts this to 13 minutes saves ~12 minutes × 60 = 720 minutes/day (12 hours). At a fully loaded pharmacist cost of $70/hour, that’s ~$840/day or ~$300K/year in capacity. If platform and enablement costs are ~$150K in year one, payback occurs within 6 months—while safety and compliance posture improves.

[IMAGE SLOT: ROI dashboard with cycle-time reduction, pharmacist time saved, discrepancy detection, and audit completeness visualized]

7. Common Pitfalls & How to Avoid Them

• Treating this as RPA: UI macros break on minor EHR changes and can’t reason about incomplete data. Use APIs, standards, and knowledgebase checks to make decisions resilient.
• Weak normalization: Missing RxNorm mapping leads to false duplicates or missed interactions. Maintain a terminology service and regularly sync local formularies.
• Over-automation: Never bypass pharmacist approval. Keep a HITL step with clear evidence and rationale for each recommendation.
• Missing audit trails: Without lineage and decision logs, surveys and payer reviews become painful. Persist inputs, model versions, and actions in Delta with time travel.
• Fragile integrations: Lack of retries/backoff and idempotency causes duplicate orders or dropped reconciliations. Bake reliability patterns into orchestration.
• Change control gaps: Rule/model updates pushed directly to production can introduce risk. Require clinical sign-off and enable safe rollback.

30/60/90-Day Start Plan

First 30 Days

• Define scope: target units, patient cohorts, and transition events (admit, transfer, discharge).
• Connectivity: Set up FHIR and ADT connections; establish eRx/PBM data access.
• Governance baseline: Configure Unity Catalog roles, PHI policies, and secrets management; create Delta tables for inputs/outputs.
• Terminology readiness: Stand up RxNorm/NDC mapping and a lightweight terminology service.
• Success metrics: Baseline cycle time, discrepancy rates, and audit completeness.

Days 31–60

• Build the agentic workflow: ingestion, normalization, interaction/allergy checks, reasoning, and HITL console.
• Orchestrate with Databricks Workflows: add retries/backoff, idempotency, and alerting.
• Pilot in one unit: pharmacists review proposed lists; collect feedback on UX and recommendations.
• Governance in motion: implement approval gates for rules/models; start audit dashboards.
• Evaluate: compare pilot metrics to baseline; refine recommendations and mappings.

Days 61–90

• Scale to additional units; tune workloads and autoscaling.
• Strengthen monitoring: lineage views, quality checks, exception queues.
• Train staff: pharmacists, hospitalists, and nursing on the revised workflow.
• Formalize change management and rollback; schedule periodic terminology updates.
• Present ROI and safety outcomes to leadership; plan broader rollout.

9. Industry-Specific Considerations

• EHR nuances: Epic, Oracle Health, and MEDITECH expose FHIR differently; confirm available resources and write-back patterns early.
• Transition complexity: Observation stays and inter-facility transfers often have the noisiest data—pilot here to maximize value.
• Allergy data gaps: Reconciling historical free-text allergies may require targeted cleanup before automated checks are fully reliable.
• After-hours volumes: Queue HITL reviews with clear SLAs so night/weekend admissions aren’t blocked.

10. Conclusion / Next Steps

Agentic medication reconciliation brings clinical reasoning, standards-based data, and human oversight together to reduce risk and accelerate care. With Databricks Workflows orchestrating ingestion, normalization, decisioning, and EHR write-back—and with Unity Catalog and Delta providing strong governance—you can deliver safer, faster, fully auditable med rec.

If you’re exploring governed Agentic AI for your mid-market organization, Kriv AI can serve as your operational and governance backbone. As a mid-market-focused partner, Kriv AI helps with data readiness, terminology services, MLOps, and the HITL operating model so your teams can move from pilots to reliable production quickly and safely.