Care Transfer Agent: AI-Assisted Discharge Placement

Discharging a medically complex patient to the right post-acute facility is still largely done by hand. For a single high-complexity case, a case manager typically spends 75 to 90 minutes reading the chart, cross-referencing level-of-care criteria and payer requirements, assembling a shortlist from a personal list of facilities, and placing five to ten phone calls to confirm beds, network status and clinical capabilities. The questions are the same every time, the facility list differs by case manager, and almost none of the reasoning survives in a form anyone can audit later.

This lab puts you in the case manager's seat and walks three patients through that workflow with an AI agent alongside you. The agent screens the chart and flags blockers. You confirm the biopsychosocial assessment, confirm the level of care, sign the physician attestation, choose which facilities to call, read the verification transcripts and record the patient's choice. The agent narrates its reasoning at every step so you can disagree with it. At one gate you will, overruling its highest-scoring facility because the family declined it.

The architecture is deliberately explicit about which component does what. Facilities are real, drawn from public California registries: Department of Public Health facility and services files, HCAI language-capability data and the CMS Care Compare long-term care hospital dataset. Bed availability, network participation and some clinical capabilities are simulated for the demonstration. Matching is deterministic weighted scoring against each patient's own clinical requirements, not a model, so the same patient and roster always produce the same ranking and the number can be explained to a physician who disagrees with it. A large language model handles clinical reasoning, narration and the conversation during the simulated verification calls. The facts a facility is judged on, and whether it can accept the patient, always come from its record. The lab shows you where that line sits so you can check it rather than take it on trust.

No clinical background is required. The lab explains each measure as it appears and opens with a short orientation to the discharge-planning process itself.

By the end of this lab you will be able to:

• Describe the hospital discharge-planning workflow end to end, and identify which steps require human judgment rather than automation
• Distinguish the three post-acute levels of care (LTACH, IRF and SNF), and explain why placing a patient at the wrong level leads to either a readmission or a denied authorization
• Evaluate an AI recommendation by reading the reasoning behind it, rather than accepting or ignoring the conclusion
• Identify which parts of an agentic system are deterministic and which are model-generated, and articulate why facility capability decisions should never be delegated to a language model
• Explain why a system that stops short of completing a task can be more valuable than one that finishes on incomplete information
• Exercise a human override of an AI recommendation, and describe why the recorded rationale is what makes the override auditable

The lab is built for case managers, clinical operations and utilization-review leaders evaluating AI for discharge workflows, and for technical audiences assessing how human-in-the-loop controls are implemented in an agentic application.

Application — Python 3 web application built with Streamlit, orchestrated as an eight-phase state machine using LangGraph, containerized and running on a single Linux virtual machine. No GPU required.

Language model — accessed through the WWT AI Proving Ground LLMaaS gateway, an OpenAI-compatible endpoint. The lab runs meta/llama-3.3-70b-instruct. The model performs clinical reasoning, step narration and verification-call dialogue; it does not determine facility capabilities, match scores or accept/decline outcomes.

Facility matching — deterministic weighted scoring in Python across capability coverage, language match, distance and bed availability. No model, no embeddings, no vector store.

Facility data — a statewide California long-term acute care hospital roster compiled from the California Department of Public Health facility and services files, HCAI language-capability data, and the CMS Care Compare long-term care hospital dataset. Facility names and locations are real; bed availability, network participation and some clinical capabilities are simulated for demonstration purposes.

Patient data — entirely synthetic. No real patient information is present in this environment.

Access — a Windows jumpbox and the application virtual machine, both inside the lab environment. You connect by RDP to the jumpbox and open the application in a browser there. Nothing is installed on your own machine.