Superintelligent Agents for Hospice Agencies: Intake and Visit Notes

An AI agent for hospice agency admin is most useful when it handles the repetitive paperwork around referrals, intake packets, scheduling and visit documentation, then routes each item to the licensed staff member who must decide. Clinical judgment, eligibility decisions and signatures stay with nurses, physicians, social workers and the rest of the interdisciplinary team. The practical starting point is one narrow role with defined inputs, outputs and approval steps, not a general assistant for the whole agency.

Key takeaways

• Begin with intake paperwork or visit-note drafting, where the rules are written down and errors are easy to catch.
• Let the agent draft, route and check. Licensed clinicians decide, edit and sign.
• Give each agent only the memory and tools its single role needs, with scoped access to each system.
• Require an approval step and a log entry for every action that leaves the agency.
• Measure success by fewer re-keyed fields, fewer incomplete packets and fewer returned drafts, compared with a baseline you record before the change.

Why intake and visit notes come first

Hospice intake depends on assembling pieces from many sources: the referral, the physician's order, benefit and insurance information, family contact details, advance directive documents, and the agency's own assessment forms. Most delays at this stage are administrative. Items are missing, the same information is typed into two forms, and a coordinator waits for a nurse to call back about a field someone already recorded. The work is slow because it is scattered, not because it is hard.

Visit notes carry a similar load. A clinician sees the patient, then spends time after the visit putting the same observations into the agency's format, often at the end of a long day in several homes. The narrative must match the structured entries, the visit schedule and the care plan, and any mismatch creates follow-up work for someone else.

These tasks are repetitive, rule-bound and checkable against written standards, which makes them a good fit for an agent. They also sit next to clinical content, so the boundary has to be drawn before any tool is connected. An agent that drafts a note from entries the clinician provided is helpful. An agent that decides whether a patient qualifies, or quietly changes what a nurse documented, is a different kind of system with a different level of risk. The same handoff pattern appears in senior living communities, where family inquiries and staff notes pass between people who each hold part of the record.

What the agent may do, and what it must not

Use this split as a first draft of the approval boundary, then review it with your clinical leadership before configuring anything.

• Collect referral documents from named sources and file them to the correct intake record.
• Compare the intake packet against the agency's required-items list and report what is missing.
• Draft an intake summary for the coordinator, with each statement linked to the document it came from.
• Propose visit slots within the scheduling rules, but hold confirmation until a coordinator approves.
• Draft a visit note from the clinician's structured entries or dictation, using the agency template.
• Flag contradictions between a new note, the prior note and the care plan, and leave the resolution to the clinician.
• Never determine eligibility, certify prognosis, interpret symptoms or recommend medication changes.
• Never alter a clinical entry after signature, and never sign, submit or transmit a clinical record without a licensed reviewer.
• Never give clinical advice to families. Route those questions to the named nurse or on-call clinician.

Each line should name the systems the agent can read or write and the person who approves its output. If a line cannot be completed that way, it is not ready to automate.

A step-by-step intake workflow

Illustrative scenario: a referral arrives by email and through a secure portal upload. The agent reads both, checks whether the patient already has an open record, and compares the packet with the agency's required items. It finds that the advance directive is unsigned and the caregiver's phone number is missing. It drafts a request to the referral source and places it in the coordinator's queue. The coordinator approves the outreach, and the agent sends it from the agency's address using the approved template. When the documents arrive, the agent files them, updates the checklist and tells the coordinator the packet is ready for the nurse's admission visit. The nurse reviews the summary, corrects one statement, and signs the admission record. Nothing reaches the clinical record until that signature exists.

The sequence the agent follows in that scenario is short:

• Receive: capture the referral and attach every source document to one intake record.
• Match: check for an existing patient or prior episode before creating anything new.
• Check: compare the packet against the required-items list and record each gap.
• Request: draft missing-item messages and hold them for coordinator approval.
• Summarize: prepare an intake summary in which every statement points to a source.
• Hand off: route the packet to the named clinician with a due time and the open items listed.
• Close: after the clinician signs, log the final state and every change the agent made.

How visit-note drafting keeps the clinician in charge

The clinician completes a short structured entry on a mobile form or dictates observations after the visit. The agent drafts the narrative in the agency template, using only what was entered for that visit, the care plan items due, and the medication list version the clinician confirmed. It does not fill gaps with general clinical language. Where the draft would need a fact that is not in the entries, it leaves a visible marker for the clinician to complete.

The clinician edits the draft, signs it, and the system keeps the version history. A draft that has not been signed never files automatically. If the clinician is working without connectivity, the agent queues the draft and does not guess at the content while waiting. Timeliness rules for documentation remain the agency's responsibility, and the agent's job is to make the on-time path the easy path by reminding the clinician of open notes, not by filing notes on their behalf.

The review step is where quality is decided. Ask clinicians to report, in a simple log, each time a draft needed a substantive correction. Those corrections are the best signal for whether the template, the prompts or the input fields need work.

Memory, tools and permissions

An agent needs memory to be useful across weeks: the agency's templates, the required-items list, scheduling rules, and the contact details of each referral source. It does not need every clinical detail about every patient held in a general memory. Scope each memory store to a single role, set retention rules with the agency's privacy officer, and remove stored material when the role changes. Health information must be handled under the privacy law that applies to the agency, such as HIPAA in the United States or the LGPD in Brazil, together with the agency's own policies.

Tools should be granted narrowly. The intake agent might read the referral inbox, write drafts to the intake queue and read the schedule, but it should not hold broad write access to the clinical record. Use scoped, revocable grants for each system, and check them on a schedule. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) describes the principle behind this: every request should be checked against what the requester is allowed to do, not against what the system can technically do.

Referral documents and uploaded files can contain text that reads like instructions, such as a line asking the system to skip a check. The agent must treat document content as data to be summarized, never as a command. The [OWASP Top 10 for LLM Applications](https://genai.owasp.org/llm-top-10/) lists this class of risk, and it should be part of the agency's test plan before go-live.

How Be Human handles this: in a first engagement, Be Human defines one role, the memory and tools it needs, the approval points and an observable success criterion. In managed operations, Be Human monitors the agent, reviews its permissions on a recurring schedule, maintains the integrations with the schedule and document sources, and handles incidents such as a misrouted packet or a draft that cites the wrong document. The agency keeps authority over clinical content throughout.

How to tell whether the agent is working

Choose measures that a coordinator or clinical manager can observe without special tooling. Record a baseline before the agent starts: how often staff re-key the same field, how many packets reach the admission visit complete, how many visit drafts are returned for substantive edits, and how long missing items take to arrive. Compare the same measures after the change. Avoid targets that reward speed alone, because a faster draft that needs heavy correction is not progress.

Build an evaluation set from past intake packets and visit notes that have been de-identified and reviewed by a clinician. For each case, write down the correct checklist result and the facts a correct summary must contain. Run this set before every change to prompts, tools or permissions. Anthropic's guide to [demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) explains why this kind of repeatable testing matters more for agents than for single responses. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers a structure for mapping these risks to owners and controls.

Finally, keep an operational log that records which source document each statement came from, what the agent changed, which person approved it and when it was signed. That log is what lets a clinical manager answer a question about a single packet weeks later.

• Baseline the four measures before the agent starts.
• Build a de-identified evaluation set reviewed by a clinician.
• Confirm that every outbound message and clinical draft has a named approver.
• Review tool permissions and memory scope on a fixed schedule.
• Record every incident, including near misses, and review them monthly.

If you want a starting point, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent write hospice visit notes on its own? No, and it should not be set up that way. It can draft a note from the clinician's structured entries or dictation, mark anything it cannot support from those entries, and route the draft for review. The licensed clinician edits, signs and owns the final record.

Does an agent make admission or eligibility decisions? It should not. Eligibility, prognosis and admission decisions belong to the physician and clinical team under the agency's policies. The agent can check that the required documents are present, summarize them with sources, and flag anything that needs a clinician's attention.

How is patient data protected when an agent is involved? Protection depends on scoped access, limited memory, logging and the privacy law that applies to the agency, such as HIPAA or the LGPD. Agents should hold only the data their role needs, and every access should be traceable. Your privacy officer should approve the design before any patient data is connected.

How does a first engagement start? It starts with one role, usually intake or visit-note drafting, and a written approval boundary agreed with clinical leadership. The engagement defines the memory and tools needed, the approval points and an observable success criterion. Managed operations can then cover monitoring, recurring evaluations and permission reviews after the agent goes live.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.