Automation and AI concierge procedures

Status: Working semantic operating manual; telephony, privacy, consent, and external-data routes require owner approval before field use
Date: 2026-07-14
Scope: Automation and AI assistance across the eight non-respite GUIDE service domains and GUIDE program operations. This manual does not give AI, Philippines staff, or Puerto Rico staff clinical or medical authority.

The business purpose#

Automation is not merely a way to reduce minutes. It should make the work safer and easier by:

• finding facts before the beneficiary or caregiver is asked to repeat them;
• showing the source, date, author, and uncertainty behind every important fact;
• preparing a lightly trained care navigator with the questions and prior commitments that matter now;
• preventing required work, refusals, contradictions, and promised follow-up from disappearing;
• helping the beneficiary or caregiver prepare for difficult calls and forms;
• reaching outside organizations through the least burdensome reliable route;
• moving failures to the right human without forcing the person to start over;
• converting conversations into specific actions with owners and return expectations; and
• confirming the real result rather than treating a send, appointment, referral, or generated note as closure.

The unit of automation is the whole care episode, not the isolated task. A transition episode may contain records, medicines, appointments, equipment, home services, caregiver instruction, clinical review, and Care Plan updates. Proxi should gather and reuse the shared facts once, complete as much connected work as the episode allows, and count each person's actual labor once.

Work allocation by act, authority, and cost#

Routes are selected by the act and authority required, not by passing through tiers. Software, AI, and permitted lower-cost staff prepare every route. L1 performs administrative exception and recovery work. L2 enters only for a required or requested human GUIDE service, unresolved ambiguity, disagreement or distress, a warm introduction, relationship-dependent work, or useful human confirmation. Repeated failure stays with the original worker and same-lane lead unless the remaining work is a concrete system, software, billing, Medicaid, workforce, nursing-leadership, medical-leadership, or legal act. A clinical packet routes directly to the authorized U.S. clinician. No role is a prerequisite for another unless the specific procedure requires it, and urgent care is not delayed for packet perfection.

Prepare every judgment before buying it#

The expensive or relationship-sensitive human should not open with a blank chart and the instruction “figure it out.” Before L2, a clinician, a same-lane lead, or a concrete enabling or escalation job enters, Proxi should complete all permitted retrieval and organization: source-linked facts, dates, contradictions, unknowns, prior promises/instructions, viable options and tradeoffs, the exact unresolved question, and a draft communication/action plan. The workflow names the real job and exact act.

For a preference-sensitive matter, software or bounded AI may collect stated preferences, remember earlier choices, explain source-backed options, and accept a clear choice from the beneficiary or decision-specific authorized representative. A participating caregiver may supply that caregiver's own preference or report but does not acquire decision authority by participating. L2 enters only when a human GUIDE service is required, the person asks for help, ambiguity/disagreement/distress remains, a warm introduction or relationship act matters, or human confirmation materially improves the result. “Preference-sensitive” does not itself create navigator labor.

For a medical matter, L0/L1 gathers and organizes the records, literal person report, discrepancies, missing items, and precise clinical question. The authorized U.S. clinician then performs the examination or clinically necessary history and executes the judgment: interpretation, reconciliation, urgency, diagnosis, treatment, prescribing, or another protected act. The clinician is not used as the default data gatherer, chart organizer, form filler, scheduler, or record chaser.

Preparation is not pre-decision. AI and nonclinical staff do not decide which clinical source is true or sufficient, and the clinician may request more information. A required live-human service or urgent safety/clinical route begins immediately while preparation continues around it.

Hard boundaries before AI touches a telephone call#

AI assistance may sit beside a human service, but it does not replace the human GUIDE service. In particular, CMS Appendix B 3.1 requires human support, not artificial intelligence, for the off-duty 24/7 route. The Philippines first-answer person therefore remains present and responsible for the human response even when software retrieves the record, creates a call brief, transcribes with permission, or drafts the recap.

Permission is not one checkbox. Before a person-specific call, the system must distinguish and establish each permission that applies:

The conservative national operating path is affirmative disclosure and permission from every participant before AI listens to, transmits, whispers into, transcribes, or records an external call. This is a Proxi risk-control design pending counsel approval, not a claim that every jurisdiction uses identical law. It avoids silently switching rules when participants are in states such as California, Florida, or Washington, whose statutes contain all-party-consent requirements for covered communications. If any participant declines, AI leaves the audio path immediately; the beneficiary/caregiver may use a static pre-call checklist or a disclosed human-only call instead.

Permission comes from each participant for that participant's own participation. A representative may grant permission for another person only after authority for that specific act is established. Preserve who made the disclosure, the disclosed AI functions and retained artifacts, the call purpose, each participant's response, and the time of permission or withdrawal. Before permission, AI remains outside the human audio path; any non-PHI menu-control exception requires a separately approved policy. Withdrawal stops future AI listening, capture, and call-content processing immediately. Previously created artifacts follow the disclosed approved retention/deletion rule; staff must not promise retroactive deletion unless it is required and operationally available.

For an outbound AI-generated voice, do not assume that the purpose is exempt merely because it concerns health care. The approved call class must establish the applicable consent, emergency purpose, or exemption before dialing. The call must identify the initiating entity and provide the required callback/disclosure information; any applicable opt-out or revocation must stop future calls in that class. Telemarketing is outside this manual.

A transmission-only telecommunications provider may qualify as a conduit when its access to PHI is transient. A vendor that, on behalf of a covered entity or business associate, creates, receives, maintains, records, transcribes, translates, summarizes, or retains PHI is not treated as a conduit merely because it provides telephone technology. Proxi must determine the vendor's HIPAA role and execute a BAA when required. Before activation, Proxi must also approve security/risk controls, exact retention and deletion treatment, permitted model use, subprocessors, offshore access, and access/correction handling. “We did not save the audio” is insufficient when a transcript, embedding, prompt, summary, debug log, or model trace was retained.

Universal automated episode procedure#

1. Name the reason and intended value#

Before retrieval, outreach, or a call, state:

• the person-specific request, event, due obligation, open promise, or accepted goal;
• the service Proxi is trying to furnish;
• the useful result expected for the beneficiary or caregiver;
• why this person's participation is or is not needed now; and
• the human or outside endpoint that would make the result true.

An empty billing month, generic “engagement,” or available automation is not a reason.

2. Gather what can be known without asking again#

Retrieve permitted current facts from the Proxi record, EHR/HIE, CMS/Medicare route, claims, e-prescribing or pharmacy source, provider response, referral partner, prior conversation, assessment, Care Plan, or person-supplied document. For every fact preserve source, author/organization, event date, receipt date, and what the source can establish.

Software may merge exact duplicates for presentation while retaining every source. It must display stale, missing, conflicting, corrected, and late facts rather than choosing the most convenient answer. A claim proves a billed event, not what happened clinically today. A directory proves a listing, not availability. A medication source proves its own list, not actual use or a reconciled regimen.

3. Confirm identity, participation, and permission#

Before protected disclosure or action, establish the speaker, beneficiary, caregiver participation, decision-specific representative authority, purpose, allowed information, current communication channel, and any call/recording/AI-assistance permissions required by approved policy. The active worker's screen must show the established limits before the protected act begins. For an inbound 24/7 request, L1 answers immediately and establishes any missing limits during the live response; a completed brief is never a prerequisite to human connection.

If identity or authority is insufficient, the concierge may give only approved general information and may help establish the missing permission. It does not fill the gap by inference, prior familiarity, relationship title, shared phone, or caller confidence.

4. Prepare the smallest useful interaction#

Create a source-linked brief containing:

• the exact result sought;
• known facts and their dates;
• facts that must be confirmed from the person or outside party;
• prior attempts and replies;
• accepted choices, refusals, and topics not to repeat;
• one plain opening and the minimum questions needed;
• approved information the concierge or staff may provide;
• conditions requiring L1, L2, an exact Proxi program/privacy/compliance/legal/billing/quality/data specialist act, or clinical help; and
• the expected closure evidence.

The brief is an aid, not a script that forces irrelevant questions.

5. Dispatch the approved route that can produce the result#

Versioned deterministic rules select the permitted route by the act required: deterministic transaction; approved electronic communication; person-chosen AI self-service; approved automated administrative call; AI-assisted beneficiary/caregiver call; L1 administrative work; L2 human GUIDE work under the D-021 entry conditions; a named Proxi specialist act from the closed program/privacy/compliance/legal/billing/quality/data list; or authorized clinical/emergency work. These are parallel authority routes, not stages. A clear prepared choice proceeds directly, and a clinical act routes directly to the authorized clinician.

Approved deterministic rules select the permitted mode from the established facts, permissions, required authority, and known endpoint capability. AI may surface candidate missing facts or possible stop signals; it does not choose the service, role, urgency, disclosure, billability, closure, or escalation outcome. Missing or conflicting facts invoke the approved hold or authorized-human route. The workflow moves directly to a known human-only or clinician-only act without forcing lower-cost failed attempts first. Except for urgent and required-live-human entry, the receiving human gets the prepared packet rather than starting the retrieval and organization again.

24/7 carve-out. This ordering does not apply to an inbound request for the GUIDE 24/7 service. That contact routes directly to a live Philippines L1 first-answer person. AI may assist the live human or remain available as a separate optional tool, but it may not answer first, collect information as a prerequisite to human help, delay the human connection, or receive credit as the required human response. If the person reaches an automated front door, the first available action is immediate live-human connection; any information collected beforehand is optional.

6. Conduct the interaction and confirm important facts#

During the interaction, distinguish:

• what Proxi already knew and its source;
• what the beneficiary/caregiver says now;
• what the outside party says and that speaker's role;
• what the AI suggested;
• what a Proxi human explained or promised; and
• what an authorized clinician decided.

For dates, medicines, appointments, costs, addresses, names, instructions, choices, deadlines, and promises, read back the important fact and obtain confirmation or correction. AI may draft the read-back; the person or responsible human confirms it.

Read-back confirmation validates only what the confirming actor is authorized and competent to confirm. A beneficiary or caregiver may confirm their own report, preference, or choice; an outside party may confirm its own statement; and an authorized clinician confirms a clinical decision. Acceptance of an AI recap does not promote every extracted statement to confirmed fact or convert an AI suggestion into an instruction. Preserve each consequential fact, decision, promise, and action with its actual source and authority.

7. Treat AI-detected signals as stop candidates, not decisions#

AI may identify language or events that match an approved stop signal. The AI interaction then stops at the safe boundary, and the approved deterministic workflow connects the human assigned by rule. AI does not determine clinical meaning, urgency, legal authority, or final disposition. Stop and raise the approved routing signal when:

• the person asks for a human or uses the approved spoken help request;
• AI cannot understand identity, language, accessibility, intent, or the answer after one clarification;
• facts conflict in a way that affects the action;
• the person is distressed, angry, frightened, confused, or repeatedly misunderstood;
• the outside endpoint negotiates, refuses, changes terms, or asks for information beyond the approved purpose;
• a privacy, representative-authority, complaint, coverage, or consequential exception appears;
• a new symptom, possible danger, medicine question, clinical interpretation, or request for medical advice appears; or
• the interaction reaches its approved time/attempt limit without a usable result.

The context packet goes with the transfer so the person does not repeat the entire history. The receiving human restates the known issue briefly and asks the person to correct it.

8. End with an action plan, not merely a summary#

At the end, produce and confirm:

• the reason for the interaction;
• material facts learned and their speakers/sources;
• the exact result achieved;
• every promise or open action;
• one named owner for each action;
• the expected response or due condition;
• what the beneficiary/caregiver needs to do, if anything;
• the usable help/return route;
• what remains uncertain or unresolved; and
• whether another human or clinical service is already engaged.

Provide an accessible recap to the beneficiary/caregiver when permitted and useful. A transcript, AI summary, or task list remains a draft presentation artifact. Human acceptance may correct the presentation, but each consequential fact, decision, instruction, promise, and action still requires its own authoritative source or authorized confirmation.

9. Pursue the actual endpoint#

Automation tracks due responses and completes permitted follow-up. It changes channels or sends work to L1 when a destination is dynamic, no electronic route exists, a response contradicts the record, or the planned transaction fails. A clear prepared choice proceeds directly. L2 enters only when a human GUIDE service is required, the person requests help, ambiguity/disagreement/distress remains, a warm introduction or relationship act matters, or human confirmation materially improves the result.

The requested result is successfully closed only when objective evidence shows it was achieved, or when the person with authority explicitly declines, withdraws, or confirms that the need no longer exists. An accepted handoff, exhausted attempt, waitlist, unresolved external dependency, or named next owner is a truthful disposition, not successful closure. Proxi may end the current pursuit after documenting the non-success outcome, current risk treatment, next owner, and follow-up expectation, but every dependent promise remains open until completed, withdrawn, no longer applicable, or separately dispositioned. Attempt exhaustion never proves that the underlying need was resolved.

10. Attribute service and labor once#

One natural interaction may substantively furnish several GUIDE services, but each service receives credit only for content meeting its own requirement. Record each role's actual minutes once. Work performed inside the main conversation is not added again as a separate service allowance. Clinical and external work remains separately attributable to the people who performed it.

Authorized outside-data acquisition ladder#

Proxi should use the permitted source that is authoritative enough for the specific fact and current enough for the intended action, then ask the beneficiary or caregiver only to confirm, correct, explain, choose, or supply what remains missing. No feed is universally strongest: the person is authoritative for preference and person-reported use; a payer for its own assignment or coverage record; a provider for its own order or record; and an authorized clinician for clinical interpretation or reconciliation. Sources that can only generate candidates remain labeled as candidates. Access is never assumed and requires the source-specific enrollment, contract, identity, permitted purpose, patient authorization, technical onboarding, or trading-partner predicate.

Every active ladder row must cite a registered primary source before field use. A row without a source-register ID is a research or procurement hypothesis only. Product capability does not establish Proxi's entity entitlement, permitted purpose, patient-match accuracy, completeness, or authority to act.

Every imported fact keeps the source organization, source-native identifier, event/effective date, retrieval date, authorization or permitted-purpose basis, and limitation. Software may reuse a confirmed fact across connected tasks; it may not silently promote a candidate to a confirmed fact.

Determining the current primary care provider#

There is no single Medicare lookup that safely answers “Who is this person's current PCP?” Blue Button claims and other claims can generate recent primary-care candidates. HETS does not supply a current-PCP designation. Medicaid or MA payer data may identify an administratively assigned PCP, depending on the payer, but assignment is not the same as the person's actual current treating relationship.

Use this order:

1. Software checks the current EHR care team, recent PCP-authored records, hospital responsible-practitioner information, payer assignment, and recent claims, preserving source and date and labeling every unconfirmed result as a candidate.
2. Present the prepared candidates and ask the beneficiary or authorized caregiver only to confirm, correct, or supply what the sources could not establish, including whether the office still provides ongoing primary care.
3. L1 verifies the office, current relationship, destination, and desired communication route before sending PHI or attributing a PCP duty.
4. Preserve conflicts; software does not select the most recent candidate or payer-assigned clinician as the actual PCP.
5. Permit “no current PCP” or “not yet confirmed” as a truthful result and route the resulting care-coordination need.

Daily Companion and Caregiver View — Proxi product layer#

The Proxi Daily Companion is a permission-scoped digital product between scheduled human services. On configured channels and cadence, it can offer approved information, reminders, help with current actions, source-attributed self-report, and a request for routed human follow-up. It is not continuous monitoring, emergency response, clinical review, synchronous human availability, or a guarantee of immediate human help. This is a Proxi product design under D-023, not a CMS-required daily contact, a ninth GUIDE pillar, or a substitute for the human service required by Ongoing Monitoring, 24/7 Access, Caregiver Support, or any other pillar.

The Caregiver View lets an authenticated, permission-scoped caregiver see the latest source-dated information without receiving a synthetic “the patient is fine” conclusion. It shows only what the caregiver is permitted to see: the last successfully completed Proxi interaction and time; the source and as-of time for each material fact; the beneficiary's exact self-report or confirmation; completed and open actions with owners; and what is stale, conflicting, or unknown. Silence, a missed interaction, and stale or conflicting information remain unknown, never reassurance.

At a natural point in the caregiver experience, Proxi may ask a short optional caregiver pulse, such as “How are you holding up?”, “Would practical help be useful?”, or “Would you like us to route a request for human follow-up?” The response remains the caregiver's own report and may trigger routed review. It never becomes a beneficiary fact, burden diagnosis or score, urgency decision, completed caregiver or clinical assessment, completed one-on-one support call, completed service, or payment evidence.

Daily companion and caregiver-check procedure#

1. Confirm identity and current permission. Verify the beneficiary or caregiver and the information/action scope available to that person. Do not infer caregiver authority from family relationship, contact status, shared address, or prior portal access.
2. Show the latest source-dated information. Display the source, as-of time, known limits, open actions, and any stale or conflicting facts. Silence, stale data, and conflict remain unknown and never become “stable,” “fine,” or another reassuring conclusion.
3. Offer the beneficiary interaction. The beneficiary may participate, defer, decline, change channel, or request a person. Approved content, reminders, and action help remain inside the bounded software lane.
4. Offer the caregiver pulse when applicable. Capture the caregiver's exact current report, requested help, and preferred response route. Preserve it as caregiver-owned; it may trigger routed review but does not establish beneficiary fact, burden, urgency, completed assessment/support, service completion, or payment evidence.
5. Prepare the smallest useful packet. Retrieve related plan items, promises, known facts, contradictions, and the exact unresolved question. AI may identify a configured stop candidate but cannot decide urgency, risk, authority, or destination.
6. Route to the actual rostered performer. Deterministic rules send administrative or live-first-answer work to Philippine L1; required/requested nonclinical GUIDE relationship or caregiver work to Puerto Rico L2; nursing, behavioral, or safeguarding judgment to the applicable U.S. L3 professional; prescribing, diagnosis, treatment, reconciliation, or higher medical judgment to U.S. L4; outside acts to the exact L5 endpoint; and a specialist act to the applicable Proxi program, privacy, compliance, legal, billing, quality, or data specialist with the exact specialty and act named.
7. Let the human perform only the human act. The human begins with the permitted source-linked summary, asks for correction, and performs the required relationship, clarification, support, or protected judgment. The person does not restart the story.
8. Return one useful action plan. State what was reported or confirmed, what happened, who owns each action, what remains unknown, and when the person should expect the next result.
9. Credit only actual service. The episode may open work in one or more existing pillars, but Daily Companion or caregiver-pulse activity alone creates no GUIDE human contact, completed assessment or support call, service completion, closure, or payment evidence.

Canonical Proxi software surfaces#

Activation mechanics remain open under O-061, coordinated with O-042, O-056, and O-059. They do not reopen the settled D-023 exclusions.

Concierge operating modes#

AI self-service conversation#

Use when the person chooses AI, the purpose and content are approved, identity/permission are sufficient, and no human or clinical act is required. This mode is never the first-answer route for an inbound GUIDE 24/7 request.

1. Explain that the person is interacting with AI and how to request a human.
2. State the bounded purpose in one sentence.
3. Present source-linked known facts and ask only for confirmation, correction, missing facts, or the person's choice.
4. Deliver approved information in the person's language and format.
5. Stop and transfer on any authority, comprehension, distress, safety, clinical, or material-conflict trigger.
6. Read back important facts, choice, and next action.
7. Send the confirmed result into the responsible service and give the person the recap/help route.

Self-service completion does not prove the person understood an individualized clinical issue, that an outside service occurred, or that a human GUIDE contact requirement was met.

AI asks only questions and provides only content selected by the approved deterministic procedure for this purpose. After an approved clinical stop signal, AI may draft the exact clinical question from the person's literal report, source-linked facts, and discrepancies. It may not invent a clinical issue, determine urgency, decide that clinical judgment is required, select the receiving professional, or decide that a human requirement has been satisfied.

Automated outbound administrative call#

Use only for an approved nonclinical call class whose consent/exemption, destination, content, identity, disclosure, and stop rules are established before dialing.

1. Confirm the number, permitted recipient/message treatment, local time, language/accessibility need, artificial-voice authority, and whether a live person or voicemail is allowed.
2. State that the voice is automated, identify Proxi/the responsible initiating entity, give a callback route, and explain the narrow purpose without disclosing more than the verified recipient may receive.
3. Offer a human at the start and whenever the person asks, appears confused or distressed, or gives an answer outside the approved path.
4. Perform only the approved administrative act: confirmation, reminder, status, scheduling within known choices, or collection of a bounded nonclinical response.
5. Read back consequential dates, names, selections, and promises.
6. Honor stop/revocation immediately and prevent redial by another automated campaign for the same call class.
7. Transfer any negotiation, complaint, representative-authority issue, clinical content, safety issue, or material contradiction with the context packet intact.
8. Record the call disposition and the exact bounded administrative act with its evidence: for example, appointment confirmed with attributable date, status obtained from the responsible party, or submission receipt received. Also retain answered/transferred, declined, revoked, wrong party, inaccessible, voicemail left/not left, failed route, or no answer. “Completed” never means the GUIDE service or underlying need is complete, and a placed call is not a furnished service.

Private caregiver whisper during an external call#

The whisper is a coaching channel, not secret AI participation. It may be used only after the final telephony/privacy policy confirms disclosure, permission, recording/transcription treatment, and the external call's allowed purpose.

1. Before dialing, show the caregiver what the AI will hear, what it can whisper, what will be recorded or retained, and how to pause it or summon a person.
2. Agree on the desired result and the facts/questions the caregiver wants help presenting.
3. During the call, whisper short prompts only to the caregiver: a missed question, a fact to verify, a request to slow down, a plain-language request for repetition, or a reminder not to agree before understanding.
4. Never whisper a clinical conclusion, false statement, coercive negotiation, or instruction to conceal the AI.
5. When the call exceeds the approved scope or the caregiver requests help, raise the approved stop signal so deterministic rules connect the directly authorized Proxi human with the prepared context.
6. At the end, ask the caregiver to confirm what was decided, promised, and still unclear; prepare the action plan and recap.

During the call, AI may offer only a pre-approved nonclinical prompt tied to the prepared call objective and confirmed facts. A newly generated question that could affect clinical care, urgency, consent, benefits, or legal rights is not whispered. AI stops whispering, preserves only the permitted literal statement, and raises the approved stop signal; the preparation layer builds the packet and versioned deterministic rules assign the authorized human.

Co-dial and three-way conversation#

Use when the beneficiary/caregiver wants support reaching an outside party and the purpose, recipient, information, and authority are established.

1. Prepare the call goal, permitted information, source facts, questions, and closure evidence.
2. Confirm who will speak for which facts and decisions. Proxi does not impersonate the beneficiary or professional.
3. Dial the verified destination and identify participants according to approved policy.
4. Let deterministic/AI assistance handle menus, waits, and retrieval of approved facts where permitted.
5. Philippines L1 handles an administrative call. Puerto Rico L2 joins only when a human GUIDE service is required, the person requests help, ambiguity, disagreement or distress remains, a warm introduction or relationship act matters, or human confirmation materially improves the result. A clear prepared choice proceeds without L2.
6. If the outside party raises a medical question, pause that portion and connect the separate authorized clinician rather than asking L2 to answer.
7. Read back the outcome, responsible party, date, reference number, and next step before ending.
8. Send the caregiver an accessible recap and carry every promise into follow-up.

Human takeover without repetition#

When the person requests a human or an approved deterministic rule assigns one after an AI stop signal:

1. The human receives only information and artifacts whose collection, transfer, and retention remain permitted. If transcription was not permitted, transfer source-linked facts and an authorized human-authored note rather than creating or implying that a transcript exists.
2. The human opens with a one-sentence summary and asks the person to correct it.
3. The human completes only the relationship, exception, negotiation, or coordination work that required a person.
4. AI may continue note/action-list support only while the applicable AI-processing permission remains in force, and it does not speak over the human's authority. If AI participation was declined or withdrawn, the remainder of the interaction and documentation is human-only.
5. The person receives one combined recap, not separate AI and human stories.

Same-lane and concrete-department recovery#

When the same destination or workflow fails repeatedly, return the work first to the original worker and the lead of that actual lane. The Philippine L1 lead repairs administrative ownership and approved backup coverage. The Puerto Rico navigator lead repairs navigator assignment or navigator-service ownership. The original clinician or clinical supervisor corrects clinical work inside scope.

If the remaining problem is not service work, route only the concrete act: configuration/access/telephony operation to the System Administrator; incorrect implemented behavior to Software Engineering; claim/payment work to Billing/Revenue Cycle; Medicaid authority questions to Medicaid operations; roster/leave/credential-file/capacity work to HR/Workforce; nursing-practice quality to the Head of Nursing; physician or medical quality to the Chief Medical Officer; and a genuine uncovered legal question to Legal Counsel. The recovery closes only when affected people and obligations are reconciled, not when a ticket is opened or the system is technically restored.

Clinical transfer#

When a clinical question appears:

1. For non-urgent work, L0/L1 completes all permitted retrieval and organization before clinician time begins: the person's exact words, current location/callback route, relevant source documents and dates, literal contradictions, missing items, already-authorized current instructions, and the exact question requiring clinical judgment. When an urgent or required-live-human route is triggered, the clinician or live human enters immediately while the preparation layer completes the remaining packet in parallel. Packet incompleteness and unresolved facts remain explicit. L2 adds only relationship or person-choice context that is genuinely needed.
2. Staff states plainly that a clinician must answer the question and does not offer reassurance or a preliminary medical opinion.
3. The correct separate U.S. professional receives the bounded question and source packet.
4. The clinician reviews the prepared packet, personally elicits or examines what is clinically necessary, requests additional facts when needed, and performs only the protected act within actual scope and beneficiary-location authority. Another authority is added only when the question genuinely requires it.
5. The authorized decision is communicated through the approved route and carried into the relevant plan/service without reinterpretation.
6. Administrative pursuit, navigator explanation or requested human assistance with the beneficiary/caregiver's choice, and clinical minutes remain separately attributable.

High-value external-call procedures#

Hospital or facility after discharge#

1. Detect the transition from the permitted ADT/HIE feed, received record, facility notice, beneficiary/caregiver report, or 24/7 contact; preserve the source and event time rather than treating a model inference as the transition.
2. Retrieve the discharge record, medication sources, pending-test/result list, follow-up orders, appointments, equipment, home-service arrangements, warning instructions, and named receiving clinicians or agencies. Produce one missing-item list without deciding the clinical meaning of any discrepancy.
3. Open separate linked obligations for each result that still has to happen: record receipt, clinical review, appointment, medicine question, equipment delivery, home-service start, transport, caregiver instruction, and promised callback.
4. Use the approved electronic route first. L1 clusters the remaining record and administrative pursuit into as few calls as possible, capturing the employee's name, department, direct route, verbatim answer, promise, and date.
5. Route the prepared symptom, instruction, pending-result, or medication question directly to the authorized U.S. clinician. The clinician personally elicits or examines what is clinically necessary, interprets, reconciles, and decides the clinical response.
6. Present already-prepared practical options to the beneficiary/caregiver. L2 joins only for requested human help, ambiguity, distress, disagreement, a needed warm introduction, or a relationship barrier; L2 does not recreate the transition packet.
7. Return one plain recap showing what is done, what is still open, who owns each item, and when Proxi will check again.

Receipt of a discharge summary and assignment of downstream owners complete only those substeps; they do not close transition recovery. Each linked obligation receives its own objective result or truthful non-success disposition.

Physician or specialist office#

1. Name the call purpose as exactly one or more separate objectives: obtain a dated record, schedule within confirmed constraints, or obtain an attributable answer to a precise clinical question.
2. Attach the specific artifact, date, beneficiary match, current callback route, prior attempts, and permitted disclosure scope. For a response request, attach the prepared source facts and exact question rather than “please review chart.”
3. Send the approved electronic request and track its matched acknowledgement before using human time. Cluster unresolved objectives into one manual office call when that does not mix permissions or recipients.
4. The concierge/L1 navigates the office, confirms the correct team, and captures every response verbatim with speaker, role, time, callback route, and promised date. Administrative staff do not interpret or relay a new medical instruction as their own.
5. L2 joins only for a requested warm introduction, a communication/relationship barrier, or human help with a prepared nonclinical choice. The prepared clinical question routes directly to the authorized clinician; an office employee's acknowledgement is not that clinician's answer.
6. Close each objective separately on the requested record, scheduled/confirmed appointment where that was the goal, attributable clinician response, or truthful non-success outcome with next owner. The original request may be dispositioned as superseded when the authorized person accepts an alternative, but every resulting appointment, record request, clinician response, referral, or callback remains open until independently achieved, withdrawn, no longer applicable, or truthfully dispositioned.

Pharmacy#

1. Assemble prescription/order, dispense/fill, claims, facility/discharge, bottle/photo, and person-reported-use sources in separate columns with dates; do not call the packet a reconciled regimen.
2. Split the work before contact: administrative facts (hours, stock, readiness, price, delivery); practical supports (packaging, synchronization, pickup); and clinical/prescriber questions (start/stop/hold, dose, interaction, side effect, substitution, appropriateness, refill authorization, or reconciliation).
3. Software checks connected status routes first. The concierge/L1 asks only the remaining bounded administrative questions and records the pharmacy's answer verbatim, including whether a prescriber response was received.
4. Present current practical options directly to the beneficiary/caregiver. L2 joins only if the person requests human help or ambiguity, distress, disagreement, or a practical relationship barrier remains. An approved teach-back prompt may ask the person to repeat source-faithful instructions already communicated by an authorized clinician or pharmacist; ambiguity routes to the clinical/pharmacy lane.
5. Send the prepared clinical/prescriber packet directly to the prescribing-authority clinical or pharmacy lane. L2 does not explain clinical rationale, correct a regimen, decide which source is current, or confirm clinical adequacy.
6. Track order, prescriber response, dispensing, possession, reported use, understanding, and clinical decision as separate events. None substitutes for another.

Payer or Medicare-related support route#

1. Prepare identity/authentication needs, authority, exact question, relevant dates of service, service/item code, prior determination or reference, supporting documents, desired result, and the person's permitted participation.
2. Use an approved API/portal route first. For a call, AI may navigate menus and absorb hold time only under the approved calling/permission policy; connect the person before any required authentication, attestation, or personal choice.
3. Capture the representative's name/identifier, reference number, exact stated result, effective dates, conditions, missing material, appeal/reconsideration route if stated, and promised decision or reprocessing date.
4. AI may explain cited administrative language and populate confirmed fields but does not decide eligibility, coverage, medical necessity, appeal merit, or legal rights. L1 handles routine status and missing-document pursuit. L2 joins only when requested human help, ambiguity, distress, disagreement, or sensitive nonclinical coordination remains after the packet is prepared. Billing/compliance/legal/clinical owners perform only their reserved acts.
5. Convert every future promise into dated verification work. Receipt of the attributable determination or requested status closes only that inquiry substep. The underlying coverage, access, appeal, documentation, or next-action obligation remains open until separately resolved or truthfully dispositioned; a call reference number alone proves neither.

Community service or referral partner#

1. Verify current operation, geography/population served, availability meaning, cost/funding facts, waitlist, documents, and intake route. Preserve unknowns and the source/date for every material fact.
2. Compare every serious option against confirmed hard constraints and stated priorities. Present the honest comparison or truthful no-match directly; software does not hide exclusions or select for the person.
3. Record the beneficiary/caregiver's own choice and the exact recipient/content/purpose permission needed for outreach. L2 joins only for requested human help, ambiguity, disagreement, distress, a requested warm introduction, or a relationship barrier.
4. Prepare the smallest permitted referral/intake packet, help populate it from confirmed facts, and obtain the person's own attestation/signature where required.
5. Transmit or introduce through the approved route; then separately confirm exact-referral receipt, intake disposition, and a dated person-specific first service.
6. Resolve denial, waitlist, no response, or failed intake by carrying forward the confirmed packet and returning the next prepared choice to the person.
7. When the purpose warrants it, ask the service seeker after first use whether the service actually helps the stated need. A referral, authorization, claim, acceptance, or appointment never proves service start or usefulness.

Form completion#

1. Retrieve authoritative fields and show their source/date.
2. Ask the person only for missing facts, choices, and attestations.
3. Distinguish person-supplied answers from AI-proposed wording.
4. Explain the form's plain administrative meaning from approved material; prepare the source-linked facts and exact reserved question first, then route only the clinical, legal, coverage, consent, or other protected judgment to its authorized owner.
5. Read back consequential entries and obtain confirmation.
6. Show the final form and destination before the authorized submission/release act.
7. Retain submission receipt and every later response; a completed draft is not a submitted or accepted form.

Coaching before or during a difficult external call#

1. Prepare the purpose, confirmed facts, desired result, likely branches, exact questions, documents, and prompts to request a written answer, representative name, reference number, and deadline.
2. Let the beneficiary/caregiver remain the speaker of record when that is the chosen route. AI may provide only disclosed, permitted, pre-approved factual or logistical whispers.
3. On the person's first request for a human—or when distress, conflict, consent/authority, clinical/safety content, or an outside-approved-content question arises—raise the approved stop signal so deterministic rules connect the directly authorized human with the prepared packet. Do not run a retention script or make the person repeat the story.
4. End with one source-linked recap, the person's correction, a short debrief when useful, and every promised result opened as trackable work.

No-repeat transfer packet#

Every handoff carries the person's exact words, source-attributed confirmed facts, contradictions and unknowns, what was already attempted, the permitted information/participants, and one precise unresolved question. The receiving actor opens with a one-sentence readback for correction and performs only the act requiring that actor's authority. Transfer attempt and receiving-owner acceptance are separate facts. Deterministic rules route directly to the authorized L1, L2, clinical, external, or Proxi program, privacy, compliance, legal, billing, quality, or data specialist lane, with the exact specialty and act named; they do not force a Philippines-to-Puerto-Rico-to-clinician sequence.

Source-linked recap and action plan#

Produce five literal outputs after a substantive episode:

1. a speaker-attributed recap, with unraised subjects marked not discussed rather than normal;
2. atomic actions, each describing one observable result;
3. one named owner for every action;
4. an actual due date or return condition; and
5. the next planned contact or an explicit statement that none is currently needed.

Separate Proxi commitments, beneficiary/caregiver choices, clinician decisions, and external-party promises. Link each open item to the owning service. A human corrects only the statements or judgments within that person's authority; source-attributed raw facts may be preserved immediately with confirmation status rather than disappearing until someone signs a note.

Outreach retry and endpoint repair#

An attempt never equals contact or completion. Automation stops the routine retry path when an endpoint changes or conflicts, wrong-party disclosure is possible, a channel redirects unexpectedly, the person withdraws permission, or delivery failure is reported. One L1 exception packet contains the full attempt history, destination provenance, allowed alternatives, and exact objective. Repair the endpoint without overwriting its history, suppress the dead or prohibited route, tell the person the truthful status when material, and return the episode to its original result—not merely to “another attempt sent.”

Natural episodes that should co-join work#

What every task expansion must add#

Each existing SOP task will be expanded with these plain sections:

• literal steps;
• required facts and obtainable external sources;
• definitely-software work;
• deterministic-plus-AI work;
• applicable concierge mode;
• Philippines, Puerto Rico, exact Proxi specialist, clinical, and external residual work;
• neighboring tasks that can be completed in the same natural episode;
• useful person-facing result;
• objective completion and non-completion evidence;
• missing/stale/conflicting/late/declined/unreachable/unsafe handling; and
• low / ordinary / high remaining human minutes by role, separating preparation, direct interaction, documentation, follow-through, and clinical work.

This expansion does not turn task prose into a software schema. Its purpose is to make the human job executable and to expose which work can disappear from the human queue only after the automated path proves the real result.