AI in Skilled Nursing Facilities: Screening Hospital Referral Packets

Yes. A skilled nursing facility can use AI for a first-pass summary and missing-information checklist, provided the tool is suitable for protected health information. Before uploading a real hospital referral packet, confirm a signed business associate agreement (BAA) covers the service and its data handling. An admissions lead must still decide whether the person can be accepted.

The useful output is a traceable worklist, not an automated accept-or-decline label. A packet may say that someone can transfer tomorrow while omitting the current medication list, isolation status or required equipment. The system should mark those gaps and point to the page it read. A person then checks the source and contacts the referrer.

Decide what the first pass is allowed to do

Give AI three jobs: identify documents in the packet, extract a limited set of operational facts, and flag missing or conflicting information against your own admissions checklist. It should produce a short queue entry that a reviewer can verify against the original pages. Keep capacity checks, clinical judgement, payer confirmation and the final response under named staff ownership. A probability score or an apparently confident summary is not evidence that the facility can safely meet a person's needs.

Write the acceptance criteria before selecting a tool. The criteria will vary with the services actually offered, staffing and the current shift, so they must come from the facility's approved policy rather than a general model. A useful checklist asks whether a current transfer date, responsible contact, care needs, mobility support, treatment requirements, equipment, infection precautions, medicines, allergies and funding information are present. The checklist is a completeness aid; the clinical team decides what each item means.

For example, a referral can contain a discharge summary dated today and a medication sheet dated three days ago. AI may extract both dates and mark “medication list needs confirmation”. It must not silently treat the older sheet as current or infer that an unchanged dose is safe. That distinction turns a quick summary into an accountable request for clarification. The broader hospital referral intake workflow for home health teams addresses the reply stage; the screening job here is the packet itself.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Put the agreement and data route in place first

Under HIPAA, a cloud service that creates, receives, maintains or transmits electronic protected health information on behalf of a covered entity generally needs a BAA. The official guidance on cloud computing says this remains true even when the provider stores only encrypted data and holds no decryption key. A vendor's claim that its product is “HIPAA-ready” does not replace a signed agreement covering the precise service, permitted uses, safeguards, subcontractors, incident reporting and return or destruction of data. Ask a qualified privacy adviser to confirm the arrangement for your organisation.

Map the full path: hospital portal or secure inbox, packet download, AI processing, human review, record storage and deletion of working copies. Each step has an owner. Ask the supplier where uploaded documents and generated outputs are kept, who can access them for support, how long they remain, whether model training uses them and how export and deletion work. Document the answers rather than assuming that an existing agreement for another product covers this feature. Restrict staff access to their admissions roles and test the process with synthetic packets before any live information enters it.

A small but revealing test is a user opening the admissions queue on a shared reception computer. If packet summaries remain visible when that user signs out, the issue is access and session handling, not text quality. Fix it before rollout. If the AI vendor cannot provide an appropriate agreement or a clear data route, use a manually maintained checklist and do the analysis inside approved systems. The BAA questions for therapy AI vendors can help structure supplier questions, although your admissions workflow needs its own review.

Build a packet manifest before extracting facts

Referral packets arrive as scans, portal exports and attachments that may contain repeated pages. Start with a manifest: document name, version or date, page range, and whether a page is readable. This prevents a later summary from treating two versions of the same form as independent evidence. Set a rule that unreadable scans and missing pages go straight to manual review. Optical character recognition can misread a decimal point, a medicine name or a tick box; never let an extracted value stand alone without a page reference.

For an illustrative packet, the manifest might read: discharge summary, pages 1–4, dated day 1; medication reconciliation, pages 5–7, dated day 2; therapy assessment, pages 8–11, dated day 1; insurance page, page 12, partially cut off. The screening output should make page 12's defect visible instead of guessing the missing policy number. A reviewer can then request a clearer copy. This is more useful than asking the model to write a polished paragraph about “overall suitability”.

Keep the AI prompt narrow. Ask it to return “found”, “not found”, “conflicting” or “unreadable” for each field, and to quote only the shortest necessary source fragment with a document and page reference. Tell it to leave a value blank when the evidence is ambiguous. A sample instruction for a synthetic test packet follows; adapt its fields to the facility's approved checklist.

Read this test referral packet as an extraction assistant.
For each requested field return: status, value if explicit,
document name, page number and the exact short source phrase.
Fields: requested transfer date; referring contact; mobility aid;
medicine list date; allergy statement; isolation precautions;
equipment request; payer information.
Use NOT FOUND, CONFLICTING or UNREADABLE where appropriate.
Do not decide placement, clinical suitability or coverage.
Do not infer a negative answer from silence.

An illustrative AI output might say “isolation precautions: not found”. If the discharge summary actually says “screen pending” on page 3, the correct output is “pending confirmation, page 3”, and the reviewer should change it. Recording that correction matters: repeated omissions on the same field may indicate a prompt, scan or source-document problem. It should not be buried as an occasional reviewer preference.

Route every exception to the right human

Make the queue states explicit. “Ready for review” means the packet can be considered by the admissions lead; it does not mean ready for acceptance. “Clarification needed” lists specific questions for the hospital. “Technical failure” means an unreadable or unprocessed packet requires manual handling. “Escalate now” covers information that your own policy identifies as time-sensitive. The workflow should not auto-send an offer of a bed or an adverse decision from any of these states.

Assign each flagged field to a role. The admissions coordinator can chase a missing contact number; the clinical reviewer checks care needs, treatments and precautions; the payer specialist confirms the funding route; the admissions lead weighs capacity and makes the final decision. If two roles disagree, retain the disagreement in the worklist until resolved. A red flag should say who owns the next action and when it is due, rather than merely changing a dashboard colour.

A second illustrative case shows why this matters. The AI extracts “walks independently” from a nursing note but misses a later therapy page saying “two-person transfer”. The mismatch should create a conflict with both references. The clinical reviewer checks which note reflects current function and asks the hospital if needed. A single merged summary could hide the change and lead to an unsafe handover.

A third case concerns equipment. A referral mentions oxygen in a current care plan, while the proposed transfer message omits delivery details. The tool should flag the missing arrangement, not suggest that the facility can source equipment by the stated transfer time. Staff need to confirm the facility's actual capability and supplier timing. If the packet contains a treatment outside the facility's scope, escalate it to the appropriate clinician rather than asking AI for a substitute plan.

A numbered trial for a twelve-packet day

Consider an illustrative skilled nursing facility that receives 12 hospital packets on a busy day. Four are complete enough for initial review, five need one or more clarifications, two contain conflicting versions of a key document, and one scan is unreadable. These are sample figures for planning, not a claimed industry distribution. The facility records how long the existing manual first pass takes for 20 packets before testing AI. It also records how often staff later discover a missing field and how often a referrer must be called twice.

1. Prepare: choose 20 synthetic or properly approved historical test packets, include poor scans and conflicting dates, and remove any that the chosen tool is not authorised to process. Have the admissions lead approve a ten-field checklist and name the reviewers.
2. Extract: run the manifest and first-pass checklist, preserving document and page citations. Put each packet into a queue state and make every “not found” field visible. No automated decision is sent outside the facility.
3. Compare: two reviewers independently inspect the original packet and the AI output. They mark correct fields, incorrect fields, missed conflicts and invented statements. A field counts as correct only if its cited source supports it.
4. Repair: change the prompt or scanning process for recurring errors, then rerun the same cases and a fresh set. If a high-risk error remains, keep that field entirely manual. Keep the old result so the improvement can be audited.
5. Operate: on the live twelve-packet day, staff validate the four initially complete packets, chase the five gaps, reconcile the two conflicts and manually process the unreadable scan. The admissions lead signs the decision record for every packet.

Suppose the manual first pass was 18 minutes per packet and the assisted first pass plus verification is 13 minutes. Across 12 packets that is 60 minutes of staff time released in this illustration. It is not a return-on-investment claim: setup, vendor review, exception handling and quality checks also consume time. A trial succeeds only if the team catches at least as many material gaps as before while keeping decisions traceable. If verification takes longer or errors rise, narrow the fields or stop.

A fourth concrete example is a duplicate packet sent after a correction. The earlier copy lists a transfer date of the 14th; the amended message says the 15th. If the tool merges both without a version label, it may place the wrong date in the worklist. Require “latest verified document date” and “conflicting earlier date” as separate entries. The person who spoke to the hospital records which date was confirmed.

Use a one-page review record at the decision gate

The admissions lead needs a short record that links the AI's first pass to the original packet and the final human action. The record should identify the packet version, reviewer, outstanding questions, source pages, contacts made, decision, reason and time. Store it in the existing authorised record system. Avoid keeping a second, ungoverned archive of packet copies in the AI tool simply because it is convenient to search.

For a fifth example, imagine a packet where an allergy statement says “none known” on the admission form but the medication reconciliation contains an allergy entry. The system should not choose the newer page automatically. The clinical reviewer resolves the discrepancy with an appropriate source and records the answer. “No allergy” and “allergy not documented” are different statements, so a missing field must never become a negative finding.

Use the same discipline for family contacts. A hospital packet may identify one contact for discharge coordination and a different person for care decisions. AI should label the roles exactly as the source does. It must not elevate a contact into an authorised decision-maker because their name appears first. A responsible staff member checks the appropriate paperwork and contact arrangements.

Measure error rates alongside queue speed

Review a sample of completed packets each week during the trial. Record the number of fields checked, unsupported AI statements, missed conflicts, clarifications raised, clarifications resolved, and decisions returned for more information. Separate the source of each problem: an absent hospital document, a poor scan, an extraction error, a stale checklist or a human oversight. This tells you which part of the workflow to repair. The referral tracking and chase-letter tutorial covers the follow-up queue once a specific gap is known.

Set stop rules before launch. Pause processing if the tool assigns information to the wrong patient, loses a page, exposes a packet to an unauthorised user, or repeatedly converts “unknown” into “no”. Route those packets manually while the cause is investigated. Escalate a potential privacy incident under the facility's incident procedure. Do not rely on a generic vendor accuracy percentage: your own packet formats and acceptance checklist determine whether the output is useful.

For a sixth illustration, the team discovers that scanned handwritten medication pages have twice the correction rate of typed pages in its own pilot. It can route handwritten pages to manual review while using AI on typed discharge summaries. That is a narrower but defensible use of AI. It also creates a clear instruction for admissions staff, who otherwise might assume that every uploaded packet has received the same level of scrutiny.

Revisit the checklist whenever the facility changes services, staffing patterns or its referral process. Retest altered prompts on the saved synthetic cases and a fresh sample. The audit record should show who changed the criteria and why. A monthly summary of recurrent gaps can inform conversations with referrers, but avoid turning a documentation problem into an unverified judgement about a hospital's care quality.

Send a precise clarification request

A request back to the hospital should name the missing item, explain why it is needed for review and give a secure reply route. It should not quote the entire packet in an ordinary email. An illustrative message is: “Thank you for the referral. To complete our review, please confirm the current transfer support requirement and send the latest medication reconciliation. The therapy note and nursing summary appear to differ. Please reply through the agreed secure channel, referring to the existing referral identifier.” Staff check the wording and recipient before sending.

Do not promise an acceptance time merely because AI produced a summary quickly. A reply can say that the packet is under review and name the outstanding information. If the facility cannot make a decision by the requested transfer time, the admissions lead should communicate that through the established route. The final record should distinguish “declined after review”, “pending information” and “no capacity”, because these are operationally different outcomes.

A seventh illustrative case is a referral arriving just before the evening handover. The AI worklist marks an unanswered equipment question, but the hospital expects a reply before the next shift. The coordinator records the outstanding question, the person contacted and the agreed callback time. The incoming admissions lead sees the original packet and the pending clarification in one place. No one treats an old “ready for review” label as a fresh clinical approval. This handover test reveals whether the workflow remains safe when its first reviewer is unavailable.

The practical result is a controlled first pass: fewer pages to hunt through, explicit gaps to chase and a visible trail from the original packet to a human decision. Keep the BAA and data route current, make citations mandatory, and treat every AI summary as a draft worklist. That is the limit at which the tool helps admissions without quietly becoming the admissions lead.

Further reads

• AI Usage Policy for a Medical Practice: Template and HIPAA Rules — Set rules for staff using tools with patient information.
• Nursing Home Survey Readiness Checklist: Evidence Binders With AI — Prepare evidence records for a separate operational review.
• AI for Care Homes: A Registered Manager's 90-Day Implementation Plan — Plan a staged introduction for a care team.
• Patient Intake Automation in a Small Practice: Forms, AI Summaries — Compare packet review with front-door form summaries.
• AI for Nursing Homes: What a Small Facility Can Run Without IT Staff — Choose a manageable first set of tools.
• AI Tools for Nurses Who Run a Team: Rotas, Handover and Audits — A nurse manager's guide to AI for rotas, handover templates and audit write-ups, with a clear list of what must never go into a prompt.
• ChatGPT for Clinicians: Does the Individual BAA Cover Your Practice? — What the individual BAA in ChatGPT for Clinicians covers for each person in a practice, who is left out, and what an organisation-wide agreement needs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: HHS, Guidance on HIPAA & Cloud Computing; HHS, May a HIPAA covered entity or business associate use a cloud service to store or process ePHI?; HHS, Business Associate Contracts. Consulted November 2026.