Healthcare Clinical Workflow Design: AI-Assisted Discharge

A discharge process that was safe but slow

A five-hospital nonprofit system in the mid-Atlantic ran roughly 71,000 inpatient discharges a year. The clinical outcomes were sound, but the discharge process itself had become the throughput bottleneck. A hospitalist finishing a shift often faced eight to twelve discharge summaries, each requiring a manual synthesis of the admission note, the problem list, medication changes, pending labs, and follow-up instructions. Median time to complete a summary was 47 minutes, and summaries frequently trailed the physical discharge by a full day. The finance office measured the effect as 2.1 excess days of patient throughput held in beds that were clinically ready to turn over.

The executive team had circled the same question for several quarters. Every proposal to speed the work raised the same objection from the chief medical officer, and the objection was correct: discharge is safety-critical, and a faster process that introduces medication-reconciliation errors or drops a pending result is worse than a slow one. The Stratenity engagement was scoped to produce a decision the team could defend to its own clinicians and its regulators, not a pilot that would stall the first time a summary went out with a wrong dose.

The evidence base was uneven at the start. The electronic record held clean, current data on orders and results, but the tacit knowledge of how an experienced hospitalist actually assembled a safe summary lived in a handful of senior clinicians and had never been written down. The first two weeks were spent shadowing discharges across the medicine, surgery, and cardiology services and turning that tacit process into an explicit field-by-field map the whole team could read together. That map, not the model, became the foundation of the redesign.

AI drafts, the clinician holds the override

The design principle was simple and non-negotiable. AI would draft; the clinician would decide. We mapped the discharge summary into fields and classified each one by risk. Low-risk narrative fields, such as the hospital-course summary, could be AI-drafted and accepted with a glance. Safety-critical fields, medication reconciliation, pending results, and follow-up escalation, required an explicit clinician action to accept, edit, or reject, and the interface would not let the summary sign without one. Every AI suggestion carried a citation back to the source in the chart, so the clinician saw why the model proposed a discharge dose before confirming it.

We piloted on two of the five hospitals for ten weeks before any expansion. A worked case set the pattern. A 68-year-old patient admitted for a heart-failure exacerbation had four medication changes across the stay. The model drafted the reconciliation in seconds, flagged the diuretic dose increase against the admission list, and cited the order that changed it. The hospitalist accepted three lines, corrected one where the model had missed a hold placed verbally on rounds, and signed. Total time was under four minutes, and the corrected line became a labeled training signal rather than a silent error. The same case under the old process would have taken the better part of an hour and would have relied entirely on the hospitalist's recall of a hold placed verbally on rounds, precisely the kind of change that slips through when a tired clinician reconstructs a stay from memory at the end of a shift.

We also designed the operating cadence before expansion, not after. A weekly review pulled every clinician correction from the prior week, sorted the recurring ones, and fed them back into the prompts and the field rules. A single physician executive owned the workflow and its safety metrics, so when a pattern of missed holds appeared, one person had the authority to change the rule rather than a committee debating it. The cadence was the mechanism that let the system keep improving after we left the room.

Faster discharge, fewer errors, a full audit trail

• Median discharge-summary time fell from 47 minutes to 16 minutes across the two pilot hospitals within ten weeks.
• Medication-reconciliation errors caught in downstream pharmacy review dropped 38 percent, because the model surfaced every change for explicit sign-off rather than relying on recall.
• Excess throughput days attributable to documentation lag fell from 2.1 to 0.7, freeing an estimated 41 beds of daily capacity across the pilot sites.
• Every discharge carried a complete decision trail: each AI suggestion, its source citation, and the clinician action taken, queryable for any patient in the audit system.
• Clinician acceptance held above 90 percent by week eight, because the override was real and the interface never signed a critical field on the clinician's behalf.

What the engagement taught

• Field-level risk classification is the whole design. Treating a discharge summary as one document forces a single trust level; splitting it lets low-risk text move fast while critical fields stay under explicit human control.
• The override has to be structurally enforced, not advisory. A prompt the clinician can click past provides no safety; a field the system will not sign without an action does.
• Citations changed adoption more than speed did. Clinicians accepted suggestions they could trace and rejected the tool when it asked for blind trust.
• Corrections are the most valuable data in the system. Each clinician edit is a labeled example, and capturing them turned the workflow into a compounding asset rather than a static tool.
• Piloting on two sites before five kept the failure surface small enough that the first wrong suggestion was a learning event, not a system-wide credibility loss.

How to run this in your system

• Decompose the target clinical document into fields and assign each a risk class before you write a single prompt.
• Define, in the interface, which fields the system may accept passively and which require an explicit clinician action to sign.
• Require a source citation on every AI suggestion so the clinician can trace it to the chart before confirming.
• Instrument the audit trail to capture each suggestion, its source, and the clinician action, queryable per patient and per clinician.
• Pilot on a bounded set of units, measure summary time and downstream error rate against a protected baseline, and expand only once acceptance and safety both hold.