A wraparound team meets in a living room in Fresno County: the youth, his grandmother, a care coordinator, a behavioral health clinician, a peer support partner, and a basketball coach the family asked to be there. Ninety minutes of conversation produce the raw material for a person-centered plan of care that a 1915(c) waiver program requires, an EVV-verified service schedule that Medicaid will audit, and a stack of wraparound documentation the care coordinator will type until midnight. The temptation arrives the next morning: feed the meeting notes to a model and let it write the plan, goals and all. That last step is the one this lesson exists to prohibit, because in Home and Community-Based Services, the participant's goals in the participant's voice are not a documentation style, they are the legal substance of the plan. By the end of this lesson you will be able to build the AI rollout pattern for HCBS and wraparound settings: drafting the structured plan from the team-meeting transcript, surfacing strengths-based language, mapping authorized service units against utilization, and holding the two hard limits that keep the whole thing lawful: AI does not author goals on behalf of the participant, and AI does not infer a participant's preferences from chart history.

The Waiver World: 1915(c), 1915(i), and Why the Plan Is the Law

Home and Community-Based Services exist because Medicaid's default architecture funded institutions, and the waiver authorities, 1915(c) for the classic HCBS waivers and 1915(i) for the state-plan HCBS option, let states pay for services delivered in homes and communities instead. The deal embedded in those authorities is person-centeredness: the state gets federal money to serve people outside institutions, and in exchange, the services must be organized around the person's own goals, preferences, and chosen life, not around what a facility or an agency finds convenient. The HCBS Settings Rule made that deal enforceable, requiring that settings be integrated in and support access to the community, and putting the person-centered plan of care at the center of compliance: the plan must be driven by the individual, reflect their preferences and goals, document their choices among services and providers, and be developed through a process the person directs.

This is why the person-centered plan is not a progress note with better adjectives. It is the legal instrument that authorizes services, the document the waiver case manager audits, and the evidence that the setting and the services honor the participant's direction. In wraparound, the high-fidelity team-based model used heavily for youth with serious emotional disturbance, the plan emerges from exactly the kind of meeting this lesson opened with: the youth and family at the center, a team they chose, strengths and needs surfaced in their words, and a plan that belongs to them. Documentation burden in this world is enormous, and the people carrying it, care coordinators and wraparound facilitators, are among the lowest-paid and highest-turnover roles in behavioral health. The case for AI relief is as strong here as anywhere in this program. The constraint is equally strong: the thing being documented is the participant's voice, and a tool that synthesizes voice can counterfeit it without anyone noticing until a settings-compliance review reads three plans from the same agency and finds the same beautifully phrased goals in all of them.

Hold one analogy through this lesson: the plan of care is a portrait the participant sits for, not a portrait painted from old photographs. AI can stretch the canvas, mix the paint, and clean the brushes; it can transcribe the sitting and lay out the composition the sitter described. What it cannot do is paint the face from memory, because the entire legal and ethical value of the portrait is that the person was present and directed it.

The Two Hard Limits, Stated First

Because everything else in this lesson depends on them, state the limits before the workflow. First: AI does not author goals on behalf of the participant. HCBS person-centered planning requires the participant's voice, not the LLM's. A goal is the participant's statement of what they want their life to contain, "I want to get my GED," "I want to live in my own apartment with my cat," "I want to play on the school basketball team." A model can transcribe that statement from the meeting, format it into the plan template, and connect services to it. A model that composes the goal, however plausibly, has replaced the participant in the one place the regulation puts the participant in charge. Plausible is precisely the problem: a model-written goal sounds like what a person in this situation typically wants, which is the opposite of person-centered.

Second: AI does not infer a participant's preferences from chart history. The chart records what happened; it does not record what the person wants now. A participant who attended a day program for two years may have hated it for the last one. A youth whose chart is full of basketball references may have quit and moved on. Inference from history bakes the past into the future and silently disenfranchises the person from their own plan. Preferences come from the person, in the planning conversation, at this planning cycle, and the documentation workflow must be built so the only path for a preference to enter the plan is through something the participant or their chosen representatives actually said.

Notice that both limits are the same principle the rest of this program applies to risk: the consequential determination belongs to the accountable human, and here the accountable human is not even the clinician, it is the participant. That is the inversion that makes HCBS distinct. In a SOAP note, the clinician's judgment is the protected center. In a person-centered plan, the participant's direction is the protected center, and the clinician, the care coordinator, and the AI are all in supporting roles.

The person-centered plan is a portrait the participant sits for; AI may prepare the canvas and transcribe the sitting, but it never paints the face from old photographs.

The Lawful Pattern: Drafting the Plan From the Team-Meeting Transcript

Now the work AI legitimately does, and does well. The wraparound or person-centered planning meeting, conducted with consent to record that covers every person in the room, including the grandmother and the basketball coach, produces a transcript. That transcript is the gold source: it contains the participant's actual words, the family's actual priorities, the strengths the team actually named, and the decisions the team actually made. The AI pattern is structured extraction, not composition: the model maps what was said into the plan template's architecture, goals as stated by the participant, quoted or closely paraphrased with attribution; strengths as named in the meeting; needs as the team articulated them; chosen services and providers as the participant selected; the risk and backup plans the team discussed; the responsibilities each member accepted.

The prompt makes the limits operational. A working version: "From the attached planning-meeting transcript, draft the person-centered plan of care using the attached template. For every goal, use the participant's own words from the transcript, quoted or minimally paraphrased, with a transcript citation. Do not write any goal, preference, or choice that does not appear in the transcript. Where the template requires content the transcript does not contain, insert [NOT DISCUSSED: bring to participant] rather than filling the gap. List strengths only as named by the participant, family, or team in the meeting." The [NOT DISCUSSED] device is the heart of the pattern: gaps in a person-centered plan are not for the model to fill; they are agenda items for the next conversation with the person. A plan that comes back with six [NOT DISCUSSED] flags is not a failed draft; it is an honest one, and it just wrote the agenda for the follow-up call.

The verification pass belongs to two parties. The care coordinator checks the draft against the transcript: every goal traceable to the participant's words, every service choice traceable to an actual decision, every strength actually named. Then, and this is the step that distinguishes HCBS verification from every other verification pass in this program, the plan goes back to the participant and family before finalization, in accessible language, with the question the Settings Rule implies: is this your plan? The participant's review and sign-off is not a courtesy; it is the completion of the person-directed process the plan exists to document.

Surfacing Strengths-Based Language Without Manufacturing It

Wraparound documentation lives and dies on strengths-based practice, and this is a place AI helps in a way that respects the limits. The legitimate use is surfacing and reformulating, not inventing. A care coordinator's raw notes read "youth refused to engage for first 40 minutes." The transcript also shows the youth eventually walked the team through his sneaker customization side business in detail. AI can be prompted to surface what is actually present: "Identify every strength, skill, interest, and relationship mentioned anywhere in this transcript, with citations." The sneaker business, the grandmother who shows up to everything, the coach who came on a Tuesday, those are in the record, and a model is excellent at making sure none of them fall out of the documentation.

The same tool can reformulate deficit-framed sentences the writer composed: "rewrite this progress summary keeping every fact, removing deficit framing where a strengths-based equivalent is accurate." What it may not do is manufacture strengths that exist nowhere in the record, "demonstrates resilience and a positive outlook," when nothing in the transcript supports it. Manufactured strengths are not kindness; they are inaccurate clinical records that erode the plan's credibility at exactly the review where credibility matters, and they crowd out the real, specific strengths that make a wraparound plan actually work. The test is the same citation test as everywhere: every strength in the document traces to something someone actually said or did.

There is a quieter benefit here worth naming for the agency director: strengths-surfacing is a training tool. New care coordinators learn strengths-based practice faster when the model shows them, citation by citation, how much strength material was already in the meeting they just sat through. The AI is not replacing the practice model; it is holding up a mirror to it.

EVV, Authorized Units, and the Utilization Map

The third lawful AI lane is administrative, and it is where HCBS programs bleed quietly. Home-delivered services in Medicaid programs operate under EVV, Electronic Visit Verification, the 21st Century Cures Act requirement that states electronically verify visits for covered home-delivered services: who delivered what service, to whom, where, when it started, and when it ended. EVV data is structured, voluminous, and chronically under-analyzed. Meanwhile, every participant's plan authorizes specific services in specific unit quantities, and the gap between authorized and delivered units is operationally invisible until it surfaces as a problem: unused authorizations that suggest the plan is not being implemented, over-delivery that will not be paid, visit patterns that contradict the schedule the plan promised.

AI's job here is the utilization map: ingest the authorization data and the EVV-verified delivery data, and produce the per-participant reconciliation, units authorized versus units delivered by service code, trends, and flags. Flag the participant whose personal care hours are at 40 percent utilization three months running; that is either an access problem, a preference change, or a plan that no longer matches the person's life, and every one of those possibilities is a human conversation, starting with the participant. Flag the EVV gaps, visits the schedule expected that no verification shows, before they become payback findings. The hard limits echo here in a subtler register: low utilization is a signal to ask the participant, never a basis for the model, or the agency, to conclude what the participant must want. The map routes conversations; it does not draw conclusions about people's lives.

Note the documentation discipline EVV adds to the progress-note layer: the wraparound or HCBS progress note for a home-delivered visit must align with the EVV record, the same start and stop times, the same service, the same location category. An AI-drafted progress note with plausible-but-wrong times is not just the program's standing hallucination problem; it is a note that contradicts the state's own verification data. The verifiable details AI cannot supply, actual times, actual location, what actually happened, are cross-checked here by a federal data system, which makes the verification habit non-optional in the most literal sense.

The HCBS Rollout Pattern: Consent, Roles, and the Order of Operations

Assemble the rollout. Phase zero is consent and governance with HCBS-specific texture: recording a planning meeting requires consent from everyone present, including family members and natural supports who are not clients; the consent process must be accessible, in plain language and the participant's preferred language, consistent with the population the waiver serves; and the BAA and vendor diligence from the procurement chapter apply unchanged, with the addition that wraparound records for youth may also intersect the minor-consent and education-records issues the next lesson covers. Guardianship and representative arrangements need mapping before recording: who can consent, for whom, on what basis.

Phase one is the progress-note and meeting-documentation layer: AI drafts wraparound progress notes and team-meeting summaries from transcripts, the writer verifies the EVV-aligned details and signs. Phase two is the person-centered plan workflow this lesson centers: transcript-grounded drafting with the [NOT DISCUSSED] discipline, coordinator verification against the transcript, participant review and sign-off as the final gate. Phase three is the utilization map: authorization-versus-EVV reconciliation feeding human conversations. The named owners follow the pattern of this whole chapter: the care coordinator owns the plan draft's fidelity to the transcript; the participant owns the plan; the program supervisor owns the utilization-map review; the agency's compliance lead owns consent architecture and the settings-compliance posture of the documentation.

And the training message for every coordinator and clinician in the program, in one sentence each for the two limits: if a goal in the draft is not something the person said, delete it and ask the person; if a preference in the draft came from the chart instead of the conversation, delete it and ask the person. The remedy for every gap in a person-centered plan is the same: ask the person. That is not an AI policy; it is the practice model, and the AI policy simply refuses to let the tool shortcut it.

The Applied Problem: The Person-Centered-Plan Drafting Protocol

Your artifact is the Person-Centered-Plan Drafting Protocol, a one-page standing procedure your agency adopts for every AI-assisted plan. Section one, preconditions: recording consent on file from every meeting participant, in accessible language; guardianship and representation verified; the vendor BAA and the plan template attached to the protocol. Section two, the drafting prompt, written out in full so no coordinator improvises it: the transcript-only grounding, the participant's-own-words requirement for goals with transcript citations, the prohibition on any goal, preference, or choice not in the transcript, the [NOT DISCUSSED: bring to participant] device for every gap, and strengths only as named in the meeting.

Section three, the verification ladder. Step one, coordinator pass: every goal, preference, choice, and strength traced to the transcript; every [NOT DISCUSSED] flag converted into a follow-up agenda item, never filled from memory or chart. Step two, the EVV-relevant details, service types, frequencies, and schedules, checked against what the team actually agreed. Step three, participant review: the draft goes to the participant and family in plain language, walked through by the coordinator, with explicit confirmation that the goals are theirs and the words sound like them, and changes made in their words, not wordsmithed afterward. Step four, signatures: the participant signs the plan as theirs; the coordinator signs the documentation as accurate.

Section four, the standing prohibitions, printed on the protocol: AI does not author participant goals; AI does not infer preferences from chart history; chart-derived content may appear only in sections that are explicitly historical, never in goals, preferences, or choices; and no plan finalizes without participant review. Then pilot it: run the protocol on three consecutive planning meetings, count the [NOT DISCUSSED] flags per draft, time the coordinator's documentation before and after, and, most importantly, ask the three participants whether the plan sounds like them. Done looks like a protocol your waiver case manager could read without finding a single step where the model speaks for the person, three pilot plans whose goals quote the people they belong to, and a coordinator who got home before midnight.

Key Takeaways

• HCBS programs under the 1915(c) waiver and 1915(i) state-plan authorities trade institutional care for community-based services on one condition: person-centeredness, made enforceable by the HCBS Settings Rule, which puts the person-centered plan of care, driven by the individual and reflecting their goals, preferences, and choices, at the center of compliance.
• Two hard limits govern every AI deployment in this setting: AI does not author goals on behalf of the participant, because person-centered planning requires the participant's voice, not the LLM's; and AI does not infer preferences from chart history, because the chart records the past while preferences belong to the person, now. The remedy for every gap is the same: ask the person.
• The lawful drafting pattern is structured extraction from the team-meeting transcript: goals in the participant's own words with citations, strengths only as named in the meeting, services as actually chosen, and the [NOT DISCUSSED: bring to participant] flag for every gap, which converts missing content into follow-up conversation instead of model invention.
• Strengths-based language is surfaced and reformulated, never manufactured: AI excels at finding every strength actually present in the transcript and at rewriting deficit framing where a strengths-based equivalent is accurate, but a strength with no source in the record is an inaccurate clinical record, not kindness.
• EVV under the 21st Century Cures Act makes home-delivered visit data a federal verification layer: progress notes must align with EVV times, services, and locations, and AI's utilization map, authorized units versus EVV-verified delivery with flags, routes human conversations that start with the participant, never conclusions about what the participant must want.
• The verification ladder is unique in this program because its final rung is the participant: coordinator pass against the transcript, EVV-relevant detail check, participant review in plain language confirming the goals are theirs, then two signatures, the participant's owning the plan and the coordinator's attesting the documentation.
• The deliverable is the Person-Centered-Plan Drafting Protocol: preconditions including all-party recording consent and guardianship mapping, the full drafting prompt, the verification ladder, and the printed prohibitions, piloted across three planning meetings and judged by the only reviewers who matter: whether the participants say the plan sounds like them.