Multi-Agent Based Automation of Medication Reconciliation Post-Discharge

Medication errors are a leading cause of preventable hospital readmissions. By automating discrepancy detection and verification, this workflow closes the most dangerous gap in the discharge handoff, preventing adverse drug events that drive costly rebospitalizations. The financial impact is direct: for a 500-bed hospital, a 20% reduction in medication-related readmissions can protect over $2M annually in CMS penalties and unreimbursed care.

Manual reconciliation is a tedious, error-prone chart review process that can take 30-45 minutes per patient. Our agentic workflow automates data extraction from discharge summaries (PDF/HL7), compares it to pre-admission lists in the EHR (e.g., Epic, Cerner) and patient-reported home inventories, surfacing only high-confidence discrepancies for human review. This transforms pharmacist and nurse roles from data clerks to clinical verifiers, reclaiming hundreds of hours per month for higher-value care.

The workflow ingests structured and unstructured data directly from source systems—EHR APIs, scanned documents, patient portal messages—using LLM-powered extraction agents. This bypasses error-prone manual transcription of medication names, dosages, and frequencies into the reconciliation tool. By ensuring data fidelity at the point of ingestion, the system drastically reduces the 'garbage in, garbage out' risk that plagues manual processes, improving the accuracy of the entire medication list.

The workflow enforces a consistent, documented reconciliation process for every discharge, creating a defensible audit trail. Each agent action—data pull, comparison, discrepancy flag, escalation to pharmacist—is logged with timestamps and confidence scores. This automated documentation satisfies Joint Commission requirements for medication reconciliation (NPSG.03.06.01), reduces survey preparation burden, and provides clear evidence of a closed-loop, safety-focused process.

Instead of waiting for a confused patient to call, the system proactively schedules a pharmacist verification call within 24-48 hours of discharge. An orchestration agent manages the outreach, using patient preference data to choose channel (SMS, IVR, app). This timely, personalized contact improves medication understanding, adherence, and trust, directly impacting HCAHPS scores related to discharge information and care transition.

This workflow transforms medication reconciliation from a fixed-cost, labor-intensive task into a variable-cost, automated service. As patient volume grows or value-based contracts expand, the multi-agent system scales without linearly adding clinical FTEs. The architecture integrates with population health platforms to prioritize high-risk patients, aligning operational execution directly with the financial incentives of risk-bearing entities and ACOs.

This agent ingests the unstructured discharge summary (PDF, HL7/FHIR document) from the EHR (Epic, Cerner). Using a fine-tuned LLM, it extracts the post-discharge medication list, including drug name, dosage, frequency, and route, and normalizes it into a structured JSON payload. It handles OCR errors, physician shorthand, and conflicting instructions, flagging ambiguities for human review before passing the verified list to the next agent.

This agent receives the extracted discharge list and queries two systems: the pre-admission medication history from the EHR and the current home inventory from a smart dispenser API or patient-reported log. It performs a three-way comparison, identifying additions, deletions, dosage changes, and therapeutic duplications. It generates a discrepancy report with clinical priority scoring (e.g., high-risk anticoagulant change vs. low-risk vitamin adjustment).

This agent manages the approval and exception routing logic. For low-confidence discrepancies or high-risk changes, it creates a task in the clinician's workflow (e.g., Epic InBasket, Slack channel) for a pharmacist or nurse to review. It attaches the full comparison report and patient context. The agent monitors for review completion, escalates unactioned items based on configurable SLAs, and logs all decisions for audit trails and compliance (JCAHO, CMS).

Once discrepancies are verified, this agent executes the final handoff. It automatically schedules a follow-up call in the health system's nurse scheduling platform (e.g., TigerConnect, Zoom Scheduler). It then initiates a personalized outbound call or SMS to the patient to confirm the appointment time, provide key medication instructions, and answer basic questions using a voice AI. It confirms the patient's understanding before closing the loop.

This is the central nervous system, typically built with LangGraph or Temporal, that defines the stateful workflow, error handling, and retry logic. It ensures data flows securely between agents and external systems (EHR, pharmacy, telephony). A dedicated dashboard provides real-time metrics on reconciliation cycle time, discrepancy rates, and agent performance, enabling continuous tuning of the business logic and model prompts.

The primary value is reducing 30-day readmissions caused by medication errors, which carry heavy financial penalties and reputational risk. Operational savings come from freeing 8-12 hours per week of clinical pharmacist and nurse time spent on manual chart review and phone calls. The workflow also improves patient satisfaction scores (HCAHPS) through timely, clear communication and reduces the risk of adverse drug events post-discharge.