Human work and potential automation

Status: Current public-source business-semantic classification complete for all 603 tasks
Date: 2026-07-12
Scope: The residual work Proxi must perform across the 603-task inventory. Beneficiary, caregiver, participant-customer, and external-party actions are separated from Proxi labor. Respite delivery remains excluded.

Completion boundary: every current task has a documented band and actor split. The 103 gray classifications are intentional findings, not unclassified work. Leon's commercial acceptance and reconciliation against the missing controlling participant documents are the next decisions; they do not make this classification pass incomplete.

For this pass, “done” means every task is placed in the strongest band the current evidence supports, with the software work and every remaining actor named. It does not mean forcing all gray rows into a definitive band. Doing that would erase the uncertainty Leon asked to see, invent terms from unavailable participant documents, or make an unaccepted commercial decision on his behalf.

Correction to the prior labor model#

The prior model overused “human required.” A beneficiary choosing, a caregiver reporting, a participant clinician signing, and an outside provider delivering a service do not automatically create paid Proxi labor.

For business planning, each task must answer four separate questions:

What can software complete?
What, if anything, remains for a Proxi-paid person?
What must the GUIDE Participant or its authorized employee do?
What must the beneficiary, caregiver, or an external party do?

Only question 2 belongs in Proxi's direct labor model. The other actions still matter for service completion, customer friction, and external cost, but they must not be counted as Proxi staffing.

The four Proxi automation bands#

Band	Ordinary operating model	Paid Proxi human on the clean path
Definitely software	Deterministic software retrieves, calculates, matches exact facts, schedules, reminds, assembles, tracks, and detects known exceptions. It may transmit only where the current participant policy already permits automatic execution; current V1 outbound PHI packets are excluded.	None.
Deterministic + AI	AI extracts, explains, summarizes, converses, ranks, or drafts; deterministic rules constrain the permitted result and external action. A participant employee makes a narrow attestation or one-click release only when the act-specific requirement or current V1 control calls for it.	None or negligible quality sampling on an approved clean path; the named subject-matter owner performs any mandatory act or handles a defined exception.
Gray area	AI could plausibly provide the substantive service, but current CMS wording, the missing Participation Agreement, evidence of comprehension, relationship value, or legal/clinical authority leaves AI-only completion unsettled.	Human available on demand, human fallback, or limited review until resolved.
Unquestionably human	The service itself must be human: required navigator contact, required 24/7 human response, care-team service or facilitation explicitly required by CMS, clinical judgment, physical service, negotiation, or a genuinely ambiguous exception.	The named human performs the irreducible act.

A task can be Definitely software for Proxi while still requiring the beneficiary to choose or an outside provider to act. The band describes Proxi's residual work, not the entire world around the task.

These are capability and service-design classifications, not claims that every integration or automated action is already deployed. “Definitely software” means the task does not inherently justify paid human labor once the required data, permission, and connection exist. Actual field readiness is a later implementation question.

Prepare judgment before purchasing it#

Automation should not merely route an unprepared chart to a more expensive person. Before any navigator, clinician, or subject-matter owner enters, Proxi should retrieve the permitted information, preserve source and date, organize it around the actual problem, show contradictions and unknowns, prepare the viable options or exact clinical question, and draft the communication/action plan. The human performs the narrow act that still requires human presence or authority.

For a preference-sensitive task, the person may state preferences and make a clear choice through approved software or AI. L2 receives a prepared choice packet and enters only for a required human service, requested help, ambiguity, disagreement, distress, sensitive relationship work, or confirmation that genuinely benefits from a person. For a medical task, L0/L1 prepare the source packet; the authorized U.S. clinician performs examination, clinically necessary elicitation, interpretation, reconciliation, diagnosis, treatment, prescribing, urgency, or other protected judgment. Routine record hunting, form filling, and administrative chasing do not consume clinician time.

Preparation never becomes pre-decision. AI does not declare which source is clinically true or sufficient, and an urgent clinical or required live-human service is not delayed while the packet is perfected.

Portfolio result across all 603 tasks#

Task body	Tasks	Definitely software	Deterministic + AI	Gray area	Unquestionably human
Comprehensive assessment	49	5	24	2	18
Person-centered care plan	52	7	29	13	3
Care coordination and transitions	63	7	36	12	8
Referral and services	57	2	31	19	5
Medication management and reconciliation	38	4	19	7	8
Ongoing monitoring and support	37	8	13	0	16
24/7 access	29	5	8	0	16
Caregiver education and support	34	2	14	10	8
GUIDE program operations	138	52	63	9	14
Enhanced dementia-care candidates	106	2	42	31	31
Total	603	94	279	103	127

The direct business reading is:

94 tasks (16%) are definitely software. No paid Proxi human belongs on the ordinary path.
279 tasks (46%) are deterministic plus AI. Software does the substantive Proxi work; a participant-customer may make a narrow attestation or V1 release click, and Proxi handles only exceptions.
103 tasks (17%) are gray. They should be AI-first, with a named human available for failed comprehension, requested human contact, unresolved authority, sensitive coordination, or another defined fallback.
127 tasks (21%) are unquestionably human at the service boundary. They include explicit GUIDE human contact, clinical judgment, physical or hands-on service, facilitation, negotiation, and genuine exception resolution. They become Proxi labor only when the commercial package includes the corresponding role.

Thus 373 of 603 task types (62%) have no routine Proxi-paid human on the clean path. Another 103 are candidates for low-incidence fallback rather than permanent human handling.

The public operating inventory alone is 497 tasks: 92 definitely software, 237 deterministic plus AI, 72 gray, and 96 unquestionably human. The 106 enhanced tasks are more service-intensive: 2 definitely software, 42 deterministic plus AI, 31 gray, and 31 unquestionably human.

Specific reads from the user's examples#

GUIDE program operations: 115 of 138 rows are definitely software or deterministic plus AI. Twenty-nine deterministic-plus-AI rows explicitly reduce the participant's clean-path action to a submit, release, or attestation click. Nine are gray and 14 contain real human program, clinical, training, contracting, transition, audit, or claims-dispute work.
Caregiver education and support: scheduling and standard content delivery sit in the 16 software/AI rows; the 10 gray rows are principally AI-deliverable skills teaching with human availability until the counting and comprehension boundary is cleared; the 8 human rows contain the direct care-team conversation, facilitation, support call, or safety response.
Referral and services: 33 of 57 rows are software/AI on the clean path, including need intake, constraint collection, search, ranking, practical-fit explanation, choice capture, packet preparation, and ordinary follow-through. Nineteen are gray because live capacity, failed intake, unconnected organizations, Medicaid/AAA coordination, or sensitive barriers may require VEA or navigator fallback. Only five are unquestionably human, all involving the inter-organizational AAA or Tribal Aging Program relationship and operating agreement.
24/7 access: 13 of 29 rows are software/AI—routing, context, testing, failover, documentation, handoff tracking, and coverage analytics. The 16 human rows are the actual human response, support, and safety handoff CMS requires.

Deep dive A: scheduling and reminders#

Routine scheduling is software work.

Definitely software#

Read the beneficiary's or caregiver's stated time, modality, language, accessibility, travel, and communication constraints.
Read staff, clinician, vendor, or connected-provider availability.
Offer valid times, book the selected time, send confirmation, and provide self-service rescheduling.
Calculate due windows and prevent a proposed schedule from missing a known GUIDE cadence.
Send consented reminders through the approved channel.
Detect delivery failure, calendar conflict, missed contact, or an unconfirmed appointment and open the appropriate exception.
Record the appointment, communication, and delivery evidence.

No navigator, educator, trainer, or Proxi VEA should touch a clean connected scheduling case.

When the Proxi VEA enters#

The office or community organization has no connected scheduling path.
Availability, eligibility, price, coverage, waitlist, or required documents must be confirmed by phone.
A fax, voicemail, or manual portal is the only route.
A routine booking fails, the person cannot be reached, or the stated constraints cannot all be met.

The VEA verifies administrative facts, offers only beneficiary-approved options, books or pursues the appointment, and records the external answer verbatim. The VEA does not choose the provider, determine clinical suitability, or become the GUIDE care navigator.

When the navigator enters#

After the system has gathered the known facts, options, and tradeoffs, the beneficiary or caregiver requests or materially needs human help making or confirming a preference-sensitive choice.
The scheduling conflict reveals a care-plan, relationship, adherence, distress, or access problem.
The person requests a warm human introduction.
A required navigator contact is itself the service; software may schedule it but cannot replace it.

Appointment reminders ordinarily do not need a fresh HIPAA authorization; HHS treats them as part of treatment. GUIDE separately requires consent before SMS is used for supplemental communication, so the software needs the recorded channel permission—not a human approval for every message.

Deep dive B: education, videos, and AI conversation#

“Education” is not one labor category. Most content delivery can be automated; specific GUIDE services still have a human delivery requirement.

Educational work	Proxi automation band	What automation can do	Human residual
Approved text, video, audio, translation, accessibility formatting, and reminders	Definitely software	Deliver the right approved content in the person's language and format; record delivery and use.	None on the clean path.
In-scope text or voice chatbot	Deterministic + AI	Explain approved content conversationally, retrieve the supporting source, adapt reading level, ask comprehension questions, repeat, and route out-of-scope questions.	Exception review only.
Knowledge or comprehension check	Deterministic + AI	Ask approved questions, distinguish “not understood” from “disagrees,” and offer another explanation.	Human when comprehension remains uncertain or authority/capacity becomes an issue.
Required caregiver skills training	Gray area at the CMS-acceptance boundary; settled Proxi design under `D-019`	Deliver the approved ten-topic course, short issue-specific mini-series, scenarios, bounded questions, response capture, and reinforcement.	One permitted L2 care-team trainer or reimbursed vendor owns a focused aggregate application/correction touchpoint. AI-only completion is not proposed; compliance confirmation remains open under `O-036`.
Dementia-diagnosis information at program entry	Unquestionably human for the required conversation	Prepare personalized plain-language material, anticipated questions, translation, and the written packet.	A care-team member must provide the one-on-one conversation; AI supplements it.
Support group	Unquestionably human for facilitation	Schedule, present approved material, capture questions, summarize, and provide follow-up resources.	A trained facilitator works with the group.
Ad hoc one-on-one caregiver support call	Unquestionably human	Prepare context, prompts, relevant approved material, and the note.	A care-team member provides the support call.
Clinical education or treatment advice	Unquestionably human for the judgment	Present the clinician-approved plan and reinforce it.	Qualified clinician determines and explains the individualized clinical content when required.

The practical service design is therefore not “human educator versus chatbot.” It is:

Software delivers and reinforces the standard curriculum.
AI handles ordinary questions inside the approved knowledge base.
For accepted caregiver skills training, one permitted L2 trainer conducts a 5 / 10 / 20-minute low / typical / high plan-back or show-me application touchpoint across the course; this is one human event, not one review per topic.
A human also enters for services CMS expressly assigns to a person, failed application, relationship support, and clinical or safety questions. Clinical and physical-technique questions route beyond L2.

The public RFA says caregiver skills training may be virtual or in person and may be provided by the care team or a reimbursed vendor or community organization. It does not clearly establish that an AI alone may count as the provider. That exact point remains gray until the executed Participation Agreement and current GUIDE instructions are available.

Deep dive C: practical fit, referral, and outreach#

The prior phrase “navigator validates practical fit” assigned too much routine work to the navigator.

Software and AI should do the ordinary fit work#

Collect the service seeker's stated goal, location, language, accessibility, schedule, transportation, technology, price, cultural, and other constraints.
Maintain dated facts about service area, eligibility, capacity, waitlists, price, languages, accessibility, referral requirements, and contact routes.
Remove known mismatches.
Rank realistic options and explain which facts support the match.
Show stale, missing, contradictory, or unverified facts rather than hiding them.
Ask the beneficiary or caregiver to choose among the viable options.
Prepare the referral, scheduling, benefit, transportation, or information packet after that choice.

That is Deterministic + AI, not navigator labor.

What “authorization for outreach” should actually mean#

There is no useful generic action called “the navigator authorizes outreach.” One of these specific conditions applies:

Control	Who supplies it	Clean-path labor consequence
The beneficiary or caregiver selected the service or requested contact.	Beneficiary/caregiver	Software records the choice; this is not Proxi labor.
The recipient, purpose, permitted information, and channel are verified under the participant's treatment, payment, care-coordination, or other permitted process.	Participant policy and established facts	Software prepares and validates the communication. Under the current V1 boundary, an authorized human releases an outbound PHI packet with one click; future automatic release remains open.
A special authorization or restriction applies because of state law, psychotherapy notes, substance-use information, beneficiary restriction, contract, or participant policy.	Beneficiary or authorized representative supplies any required choice, authorization, objection, or restriction; privacy/legal/compliance determines whether the basis is required, valid, sufficient, and effective	Software blocks release until the exact required permission or subject-matter decision is present. A clinician separately owns clinical necessity or judgment.
A human must attest that a factual or clinical statement is true.	Authorized participant employee or clinician	Present the exact statement and evidence for a narrow review and click.
The outreach contains a sensitive choice, unresolved conflict, or requested warm introduction.	Navigator or other named human	Human performs that relationship act; software prepares and documents it.

HHS states that many treatment and care-coordination disclosures between providers are permissible without a patient authorization, and that appointment reminders do not need one. Participant policy, reasonable safeguards, beneficiary communication restrictions, state law, and special-category information still apply. The operating question is therefore “what exact permission or attestation is missing?”—not “can a human approve every routine message?”

AI is not itself the authorizer. AI may interpret the request, prepare the communication, and explain why the facts appear to satisfy a pre-approved rule. The underlying authority comes from the recorded beneficiary choice, the participant's policy or permitted disclosure basis, or an authorized signer. The deterministic executor acts only when those conditions are present.

The public RFA says the care navigator shall refer and connect the beneficiary and caregiver to relevant community resources. That does not justify manual navigator searching or routine fit validation. The current conservative posture is Deterministic + AI for matching, preparation, and tracking, with the navigator's accountability or narrow release captured as part of the required relationship until the current Participation Agreement clarifies how much of the formal referral act may execute automatically.

Precise split between AI, VEA, navigator, and clinician#

Actor	Actual job	Review
Software/AI	Intake constraints; maintain inventory; rank and explain options; prepare outreach; send non-PHI routine communications through an established route; prepare current V1 PHI packets for one-click release; monitor; retry; summarize; identify exceptions.
Proxi VEA	Call unconnected organizations; verify dynamic administrative facts; pursue records and responses; schedule manually; capture answers verbatim; resolve routine communication failures.
GUIDE navigator	Receive the prepared choice/coordination packet; own the required beneficiary/caregiver relationship; clarify or confirm a preference-sensitive choice when human help adds value; handle distress, disagreement, or sensitive barriers; provide a requested warm introduction; connect routine work back to the care plan.
Qualified clinician	Receive the prepared clinical packet; perform clinically necessary history/examination; decide clinical appropriateness, safety, urgency, referral need, reconciliation, diagnosis, or treatment. Routine retrieval and organization happen before clinician entry.
External organization	Accept the referral, determine its own eligibility, provide the appointment or service, and report the result.

The Filipina VEA should not be labeled as the relationship owner or as the GUIDE navigator. Her high-value lane is administrative execution against the outside world when software cannot complete the transaction.

Deep dive D: enrollment and program operations#

Most program operations should not require a person to retype, relay, or inspect a clean case.

Program work	Software/AI work	Narrow customer or human act	Band
Explain the standard GUIDE model	Deliver approved text, video, voice, translation, and in-scope chatbot Q&A; record what was presented and questions asked.	Human only on request, failed comprehension, capacity/authority concern, or a current contractual conversation requirement.	Deterministic + AI / Gray pending participant documents.
Obtain a beneficiary choice	Present the choice and consequences; capture the person's response, speaker, time, and scope.	Beneficiary or authorized representative chooses. This is not Proxi labor.	Deterministic + AI.
Document consent or refusal	Populate the record and preserve evidence of what was presented and chosen.	Participant employee only if policy requires acknowledgment or an exception must be resolved.	Definitely software.
Clean eligibility and service-area checks	Retrieve authoritative facts, compare dates and criteria, and show the failed fact.	Participant staff handles contradiction or a required attestation.	Definitely software.
Clinical or residence attestation	Assemble the exact source facts and present the precise statement.	Authorized participant clinician or program signer reviews and clicks.	Deterministic + AI; signer is customer labor.
Populate and submit the PAAF	Pull accepted facts, populate the current template or FHIR payload, validate completeness and timing, submit, and retain receipt.	Customer data custodian holds any required CMS identity or performs a narrow release if the portal requires it.	Definitely software or Deterministic + AI.
Retrieve preliminary and final CMS results	Retrieve, reconcile, explain status, open only genuine exceptions, and update the working record.	Participant staff resolves rejected, contradictory, or disputed cases.	Definitely software.
Generate and deliver notices	Select the correct CMS template, populate it, prepare delivery through the approved channel, and retain proof.	Current V1 outbound PHI uses an authorized release click; a future standing automatic-release policy remains open.	Deterministic + AI under current V1; potentially Definitely software later.
Maintain rosters, training evidence, and partner status	Compare current people and organizations with requirements, dates, approvals, contracts, and portal records; flag gaps.	Participant leaders credential, contract, attest, or resolve exceptions.	Definitely software / Deterministic + AI.
Care-delivery, performance, and health-equity reporting	Assemble source evidence, calculate permitted measures, draft narrative responses, and identify unsupported claims.	Authorized customer owner approves goals, interpretations, and final submission.	Deterministic + AI.
Monthly claim readiness and clean claim	Verify alignment, service occurrence, tier, rendering practitioner, dates, duplicate restrictions, and required supporting facts; construct the claim and prepare transmission under the customer's release policy.	One customer biller click under the current case-by-case release policy; no additional substantive review on a class that the approved authority matrix designates as a clean path. Any mandatory case-specific review still occurs.	Deterministic + AI under current case-by-case release; potentially Definitely software later.
Denial, contradictory evidence, privacy exception, authority dispute, or other non-standard case	Assemble the case and relevant evidence.	Named billing, compliance, privacy, clinical, or legal human resolves the exception.	Unquestionably human for the exception.
Partner contracting and material performance dispute	Track requirements, draft from approved terms, compare changes, and monitor obligations.	Authorized people negotiate, execute, and resolve the dispute.	Unquestionably human for negotiation and execution.

CMS's current 2026 Payment Methodology Paper expressly warns participants developing automated PAAF entry to update their solution when the template changes, and it permits PAAF exchange through FHIR. That is direct evidence that PAAF preparation is intended to be automatable, not a permanent data-entry job.

Requirement-based review ownership#

There are three review patterns, selected by the exact act and controlling requirement rather than by a global “compliance approves” switch:

Pattern	Meaning	Human labor
Approved clean path	The act, purpose, recipient, information, evidence, and permission fit an approved policy and no case-specific subject-matter determination is required.	Software performs the substantive work. Current V1 outbound PHI still receives the narrow authorized release click under `D-007`; the click is not automatically a compliance-person review.
Mandatory subject-matter act	Law, the executed Participation Agreement, contract, or approved policy requires a particular privacy, compliance, legal, billing, program, or clinical reviewer/attestor for every occurrence in that class.	The named owner performs the precise act. A clinician performs the clinical act; compliance cannot substitute. These minutes must be added once the `O-045` action inventory is approved.
Defined exception	The routine evidence is missing, stale, contradictory, restricted, unsupported, or outside the approved class.	Software assembles the issue; the lowest authorized subject-matter owner resolves only that exception.

This means the current 4-minute portfolio case-exception allowance is not an estimate for mandatory per-occurrence review. Until O-045 names those act classes, incidence, minutes, role, and rate, the mandatory-review lane remains explicit but unpriced rather than silently assumed to be zero.

Which humans would actually cost Proxi#

The commercial delivery model is settled by D-012: Proxi fulfills all eight GUIDE service domains in scope and respite alone is excluded. Proxi supplies the software, Philippines first-answer/admin team, Puerto Rico dementia care navigators, caregiver-service labor, program operations, and required licensed clinical capacity through employees or contracts. The Participant retains only accountability and acts that cannot legally or contractually be delegated; outside providers and community organizations retain the endpoints only they can furnish.

The task-level matrix records:

the automation band for Proxi's part;
the residual Proxi-paid role under the full-service offer;
the narrow nondelegable Participant action;
the beneficiary, caregiver, or external-party dependency.

This allows Proxi to price the service by occurrence and role-minute without pretending that every required professional must be a direct employee or that external-provider treatment time is Proxi labor.

What is unquestionably human on Proxi's side#

Under the settled full-service offer, the irreducible paid-human work is concentrated in:

the Alzheimer’s Association® essentiALZ®-certified Puerto Rico dementia care navigator providing required contact, the primary relationship, nonclinical assessment/care-plan work, caregiver education/support, and sensitive coordination;
the self-staffed Philippines human first-answer and administrative pool;
required one-on-one dementia-diagnosis information and caregiver support calls provided by a care-team member;
trained facilitation of required support-group work;
live or hands-on caregiver skills correction where the accepted delivery model requires it;
requested warm introductions and genuinely relationship-sensitive conversations;
VEA pursuit where no connected route exists or routine automation fails;
participant-facing exception resolution, contracting, negotiation, privacy disputes, authority conflicts, and non-routine billing work included in Proxi's full-service scope;
beneficiary-location-authorized assessment pre-review, clinical examination, interpretation, risk, urgency, treatment, medication reconciliation, prescribing, medical orders, and specialty assessment furnished by Proxi-supplied or Proxi-contracted professionals.

Participant actions remain separately named because legal accountability and external completion ownership do not disappear when Proxi sells the full service. That separation is an authority boundary, not a commercial option to leave GUIDE service labor unstaffed.

Source boundary#

The current public anchors are the GUIDE RFA, especially Appendix B §§3, 4, 6, and 8; the PY 2026 Payment Methodology Paper; the current GUIDE FAQ; AHRQ's current teach-back guidance; and HHS guidance on treatment, payment, and health-care-operations disclosures, minimum necessary, and appointment reminders.

The executed PY 2026 Participation Agreement, contractual Appendix D, GUIDE Connect instructions, live participant manuals, participant privacy policy, applicable state law, and special-category information rules remain necessary before gray items or release policies are treated as finally cleared.

The current repository also has a settled V1 product boundary: outbound PHI packets are human-released (D-007). Whether policy-approved routine outbound transmissions may become automatic is explicitly open (O-008). This classification treats the present clean path as software preparation plus a narrow authorized release click, while preserving automatic release as a later decision.

Task-level matrices#

The complete 603-row classification is divided only to keep the review readable:

reviews/automation-bands/assessment-care-plan-program.md
reviews/automation-bands/coordination-referral-medication.md
reviews/automation-bands/monitoring-access-caregiver-enhanced.md

Each matrix preserves the exact task name and records the band, software/AI work, residual Proxi-paid human, participant-customer action, beneficiary/caregiver/external action, boundary, and named Proxi human role.