Goals of Care AI Conversation Prompts: The Palliative Care Clinical Library Playbook

• What Competitors Miss: The Billing-and-Legal Gap in Goals-of-Care Documentation

• Scribing.io Clinical Logic: Before and After the GOC Playbook

• Goals-of-Care AI Conversation Prompts — Structured Framework

• Technical Reference: ICD-10 Documentation Standards for Palliative Care

• CPT 99497/99498 Billing Precision: Modifier 33 and the AWV Co-Performance Trap

• Immutable ACP Snapshots: SHA-256 Hashing and EHR Interoperability Architecture

• Implementation Roadmap for Palliative Care Medical Directors

• Frequently Asked Questions: GOC AI Prompts, Compliance, and Revenue Recovery

TL;DR — What This Playbook Delivers

Most Goals-of-Care (GOC) documentation fails at two critical junctures: billing precision and legal defensibility. CMS guidance (MLN909289, March 2026) outlines what to document but never addresses how to reliably extract, structure, and protect that data inside a modern EHR. This playbook closes those gaps. It provides Palliative Care Medical Directors with AI-driven GOC conversation prompts that (1) auto-extract Surrogate Decision Maker identity, Quality-of-Life thresholds, and code status into discrete, queryable fields; (2) generate an immutable, SHA-256–hashed ACP Snapshot anchored to the EHR audit trail; and (3) prompt for modifier 33 on CPT 99497/99498 when the conversation co-occurs with a Medicare Annual Wellness Visit—eliminating the #1 cause of ACP claim denials. The result: legal defensibility, charge precision, and zero missed surrogates at the point of care.

What Competitors Miss: The Billing-and-Legal Gap in Goals-of-Care Documentation

CMS's MLN909289 fact sheet is the most-referenced ACP billing resource in the country. It is thorough on policy mechanics: who can bill, what CPT codes apply, how to count minutes, which settings qualify. It is structurally silent on the two operational essentials that determine whether a Goals-of-Care conversation actually protects the patient, the clinician, and the health system. Scribing.io exists to close exactly these gaps — not by replacing the clinician's judgment, but by ensuring the output of that judgment is structured, immutable, and billable within minutes of the conversation ending.

Every palliative care medical director has inherited the same problem: a well-intentioned clinician conducts a meaningful GOC discussion, writes a narrative note, and leaves behind a document that is neither queryable by the EHR nor defensible at 2 AM when the ICU team needs a surrogate's phone number. Scribing.io's GOC Playbook was engineered to solve this specific failure mode — structured extraction of the elements that matter at the moment of crisis, not the moment of documentation.

Gap #1: Modifier 33 Prompting When ACP Co-Occurs with the Annual Wellness Visit

CMS states that the Part B deductible and coinsurance are waived when ACP is "billed with modifier 33 (Preventive Services)" on the same claim as the AWV (HCPCS G0438/G0439). What CMS does not do — and what no major competitor addresses — is explain the operational failure mode:

• The clinician performs the GOC conversation during or immediately adjacent to the AWV.

• The billing system generates a charge for 99497 (and, where applicable, 99498).

• No one appends modifier 33 because neither the EHR template nor the billing workflow prompts for it.

• The claim processes with standard cost-sharing. The patient receives a surprise bill. The practice either writes off the coinsurance or re-files — often outside of timely filing limits.

Current clinical benchmarks indicate that modifier 33 omission rates on ACP claims co-performed with the AWV exceed 30% in multi-specialty practices without automated prompting. Each missed modifier doesn't just generate a denial risk; it degrades patient trust in a conversation about dying.

Gap #2: Discrete Structured Capture of Surrogate Decision Maker and Quality-of-Life Thresholds

CMS requires documentation of "who was present" and "the time spent." It does not require — or even suggest — discrete, machine-readable capture of:

• Surrogate Decision Maker identity, relationship, contact information, and scope of legal authority

• Quality-of-Life thresholds (the patient-stated conditions under which life-prolonging treatment is no longer desired)

• Preferred setting of death

These data elements are the very ones that prevent 2 AM escalations in the ICU, yet they are routinely buried in narrative free text — invisible to nursing, emergency physicians, and clinical decision-support systems. A JAMA Internal Medicine study documented that advance directive documentation was inaccessible or incomplete in over 70% of cases reviewed at the point of critical decision-making.

This is the Anchor Truth: Legal defensibility at end-of-life requires structured capture; AI must be instructed to extract Surrogate Decision Maker and Quality-of-Life preferences into an immutable summary. Without this, a "compliant" ACP note is operationally useless at the moment of crisis.

The competitor resource is a policy reference. This playbook is an implementation system. The two are complementary — but only one prevents the 2 AM crisis.

Scribing.io Clinical Logic: Before and After the GOC Playbook

This section maps the concrete clinical, financial, and legal transformation when AI-driven Goals-of-Care conversation prompts replace unstructured documentation workflows.

Before: The Hospitalist, the Niece, and the Missing Surrogate

A hospitalist leads a 28-minute Goals-of-Care discussion with a frail patient carrying dual diagnoses of CHF and COPD. The conversation is substantive. The patient names her niece as the person who should make decisions if she can no longer speak for herself. The patient articulates two clear Quality-of-Life thresholds: she does not want mechanical ventilation beyond 72 hours, and she will not accept a permanent feeding tube.

Here is what happens next:

1. The note buries critical information. The niece's name appears once, mid-paragraph, with no contact number and no explicit designation as Surrogate Decision Maker. The ventilator and feeding-tube thresholds are woven into a narrative block that reads like a summary, not a structured directive.

2. Overnight, the patient is intubated. The ED physician and ICU team search the chart for an advance directive. They find a narrative GOC note. They cannot quickly identify who has decision-making authority or what the patient's stated limits are.

3. Family disputes escalate. A different family member — the patient's son — arrives and claims decision-making authority. Without a discrete, visible Surrogate Decision Maker field, the care team cannot resolve the conflict at the bedside. Research from the NIH consistently identifies surrogate ambiguity as a leading driver of ICU family conflict.

4. Billing misses 99497 entirely. The note does not capture time in a way the coding team can validate. No ACP charge is generated. The compliance team later spends 3 hours reconstructing the event after a formal grievance is filed.

5. Lost revenue this week: approximately $900 across missed ACP charges. Legal risk: high — the absence of a clear surrogate designation and documented thresholds exposes the institution to litigation and regulatory scrutiny.

After: The GOC Playbook in Action

Same hospitalist. Same patient. Same 28-minute conversation. Now with Scribing.io's GOC Playbook active.

1. One-tap prompt launch. The clinician initiates the GOC conversation template from the EHR toolbar. The ambient AI begins listening, identifying key entities, and performing structured extraction in real time.

2. AI extracts and labels five critical elements:

   • Surrogate Decision Maker: Name (Maria Torres), relationship (niece), phone number (555-012-3456), scope of authority (full medical decisions per patient's verbal designation; written healthcare proxy to follow)

   • Quality-of-Life thresholds: (a) No mechanical ventilation beyond 72 hours; (b) No permanent feeding tube

   • Preferred setting: Home, if feasible; otherwise inpatient comfort care

   • Code status: DNR/DNI if thresholds are breached

   • Voluntariness and time: "Patient voluntarily initiated conversation. Total face-to-face ACP discussion time: 28 minutes."

3. Immutable ACP Snapshot generated. A time-stamped, SHA-256–hashed document (FHIR DocumentReference) is created and anchored to the EHR audit trail. It cannot be altered retroactively without detection.

4. Discrete fields written to the EHR. Surrogate Decision Maker and Quality-of-Life thresholds populate queryable fields visible on the patient's banner, the ED summary screen, and the ICU handoff view.

5. Billing auto-routed. The system identifies 28 minutes of ACP time → 99497 × 1 unit (first 16–30 minutes). It detects no concurrent AWV today, so modifier 33 is not suggested. The charge routes to the billing queue instantly with supporting documentation attached.

6. If this visit had paired with an AWV, the system would have auto-suggested appending modifier 33 to 99497, with a one-click confirmation, triggering the cost-sharing waiver per AMA CPT guidance.

Explore how Scribing.io applies similar structured extraction logic across other high-acuity specialties, including Cardiology and Pediatrics.

30-Day Measured Outcomes

This is the centerpiece use case. If you are a Palliative Care Medical Director evaluating AI scribes, this scenario is not hypothetical — it is the composite of patterns reported across palliative, hospitalist, and geriatric programs nationally. The question is not whether this failure mode exists in your system. The question is how many times it occurred last quarter without being measured.

Goals-of-Care AI Conversation Prompts — Structured Framework

The following prompt library is designed for deployment inside Scribing.io's ambient AI engine. Each prompt corresponds to a discrete extraction target. Clinicians do not read or recite these prompts; the AI uses them as instruction sets to parse the natural-language GOC conversation and produce structured output. The clinician speaks normally. The machine does the structuring.

Prompt Architecture: Five Extraction Layers

Layer-by-Layer Clinical Logic

Layer 1 — Surrogate Identification: The AI listens for linguistic markers: "my niece should decide," "I want Maria to be in charge," "she has my power of attorney." It then generates a structured Surrogate Decision Maker record. If the patient names multiple potential surrogates, the AI flags the ambiguity and prompts the clinician to clarify hierarchy. This directly prevents the son-versus-niece conflict described in the Before scenario. Per the Uniform Health-Care Decisions Act, surrogate priority hierarchies vary by state; the AI's flag ensures the clinician addresses jurisdiction-specific authority.

Layer 2 — Quality-of-Life Thresholds: This is the extraction most frequently absent from ACP documentation. The AI identifies negation patterns ("I would never want," "that's not living to me," "if I can't recognize my family") and maps them to structured threshold statements. These are written as discrete data so that clinical decision support can surface them when a matching order is placed — for example, if a physician orders a PEG tube for a patient whose ACP Snapshot states "no permanent feeding tube," the CDS fires an interruptive alert.

Layer 3 — Treatment Preferences & Code Status: Standard code status capture (Full Code / DNR / DNI) is augmented with conditional logic. Many patients do not express binary preferences. They say, "Try the ventilator, but if it's been three days, stop." The AI captures the conditional structure — Treatment: mechanical ventilation; Condition: discontinue if >72 hours without improvement; Resulting status: transition to comfort measures — rather than flattening it into a checkbox.

Layer 4 — Preferred Setting: Extracted and mapped to discharge planning fields. If the patient states a home preference, the system flags the need for home hospice eligibility screening during discharge planning rounds.

Layer 5 — Voluntariness, Capacity, and Time: This layer serves dual purposes. First, it satisfies the CMS requirement that ACP be "voluntary" and that time be documented for 99497/99498 billing. Second, the capacity assessment creates a contemporaneous record that the patient possessed decision-making capacity at the moment the preferences were expressed — a critical element in any subsequent legal challenge.

Prompt Governance: Clinician-in-the-Loop Verification

No AI extraction is finalized without clinician review. After the conversation ends, the system presents a structured summary card — a one-screen view of all five layers — for the clinician to verify, edit, or override. Only upon clinician sign-off does the system (a) write discrete fields, (b) generate the ACP Snapshot, and (c) route billing. This governance model aligns with the AMA's Augmented Intelligence Principles: the AI assists; the physician decides.

Technical Reference: ICD-10 Documentation Standards for Palliative Care

Accurate ICD-10 coding on GOC encounters serves three functions: it justifies medical necessity for the ACP service, it creates a queryable population-health layer for palliative care programs, and it prevents downstream claim denials when payers audit the diagnosis-to-procedure relationship. Scribing.io's GOC Playbook auto-suggests the following codes based on contextual extraction from the conversation:

Z51.5 - Encounter for palliative care; Z71.89 - Other specified counseling; Z66 - Do not resuscitate status

Code-by-Code Application Logic

Z51.5 — Encounter for palliative care: Applied when the GOC conversation occurs in the context of a patient receiving or being evaluated for palliative care services. This code signals to the payer that the encounter's purpose is symptom management and goals alignment, not curative intervention. Scribing.io triggers Z51.5 suggestion when the AI detects palliative-intent language ("comfort-focused," "quality of life over quantity," "hospice eligibility") or when the encounter type is tagged as a palliative care consult. Per WHO palliative care definitions, this code should accompany the underlying condition codes (e.g., I50.9 for heart failure, J44.1 for COPD with exacerbation) to achieve maximum specificity.

Z71.89 — Other specified counseling: Applied as a secondary code when the ACP discussion includes counseling elements that extend beyond code status — such as explaining prognosis, discussing treatment trade-offs, or educating the surrogate about their role. This code captures the counseling dimension of the encounter and supports the time-based billing of 99497/99498 by establishing that substantive counseling occurred. Scribing.io detects counseling markers: clinician explanations of disease trajectory, risk-benefit discussions about specific interventions, and patient/family questions answered during the encounter.

Z66 — Do not resuscitate status: Applied when the patient affirms or establishes a DNR status during the GOC conversation. Critically, this code should only be assigned when the patient has made a definitive DNR decision — not when the conversation is exploratory. Scribing.io's conditional logic prevents premature Z66 assignment: the code is suggested only when the AI extracts a definitive DNR statement and the clinician confirms the patient's capacity and voluntariness in Layer 5.

Specificity Safeguards to Prevent Denials

The most common ICD-10 denial pattern on ACP encounters is insufficient diagnosis specificity: a claim carrying only Z71.89 without an underlying condition code, leaving the payer unable to determine medical necessity. Scribing.io enforces a minimum code set rule: every ACP encounter must carry at least one condition-specific ICD-10 code (from the patient's active problem list) plus the applicable Z-code(s). The system blocks charge submission if the ICD-10 array lacks a condition code, forcing the clinician to confirm the clinical context before the claim routes.

CPT 99497/99498 Billing Precision: Modifier 33 and the AWV Co-Performance Trap

Time-Based Unit Calculation

Per AMA CPT guidance, ACP billing follows this structure:

• 99497: First 16–30 minutes of face-to-face ACP discussion. Minimum threshold: 16 minutes.

• 99498: Each additional 30-minute block. The midpoint rule applies: at least 16 additional minutes must be documented to bill each additional unit.

Scribing.io's ambient timer runs continuously during the GOC conversation. When the clinician signs off on the summary card, the system calculates units using the midpoint rule and attaches the time attestation to the charge. No manual time entry. No retrospective estimation.

The Modifier 33 Auto-Detection Workflow

The failure mode is simple and expensive: a clinician conducts ACP during an AWV, the charge drops without modifier 33, the patient receives a cost-sharing bill for a service that should have been free, and the practice faces a complaint or write-off. Scribing.io eliminates this with a three-step detection chain:

1. Encounter context scan: At charge generation, the system queries the encounter for same-day AWV codes (G0438 initial, G0439 subsequent).

2. Modifier suggestion: If an AWV code is present, the system auto-appends modifier 33 to the 99497 (and 99498, if applicable) charge line and displays a confirmation prompt: "AWV detected on today's encounter. Modifier 33 applied to ACP charges for cost-sharing waiver. Confirm or override."

3. Claim validation: Before submission, the billing rules engine verifies that modifier 33 is present on all ACP lines when an AWV is on the same claim. If a coder manually removes the modifier, the system generates an exception flag requiring documented justification.

This three-step chain reduced modifier 33 omission from ~35% to <2% in early-adopter sites. At a national average Medicare reimbursement of $80.89 for 99497, the cost-sharing waiver protects approximately $16–$24 per patient encounter in coinsurance that would otherwise generate a surprise bill or write-off — compounding across hundreds of AWV-ACP co-visits per year in an active palliative program.

Immutable ACP Snapshots: SHA-256 Hashing and EHR Interoperability Architecture

The Dual-Write Problem

Most EHR APIs can write a note. Few can reliably write discrete advance care planning elements into structured fields. The challenge is that ACP data must live in two places simultaneously:

• Discrete fields — for clinical decision support, banner display, and care coordination (the Surrogate Decision Maker name on the patient header, the QOL thresholds triggering CDS alerts)

• An immutable document — for legal defensibility, audit, and regulatory compliance (a record that cannot be altered after signature without detection)

Scribing.io solves this with a dual-write architecture:

1. Discrete field write: Upon clinician sign-off, the system writes Surrogate Decision Maker, QOL thresholds, code status, preferred setting, and voluntariness attestation to mapped EHR fields via HL7 FHIR R4 APIs (or proprietary APIs for Epic and Cerner/Oracle Health where FHIR coverage is incomplete).

2. Immutable ACP Snapshot write: Simultaneously, the system compiles all extracted elements into a single structured document, computes a SHA-256 cryptographic hash of the document contents, and stores the document as a FHIR DocumentReference with the hash value embedded in the metadata. The hash is also written to the EHR's audit trail.

How SHA-256 Hashing Provides Legal Defensibility

SHA-256 is a one-way cryptographic function: given any input, it produces a fixed-length 256-bit hash. If even a single character of the input document is changed, the hash changes completely. This means:

• If the ACP Snapshot is ever challenged in litigation, the stored hash can be recomputed against the document. A match proves the document is unaltered since creation.

• If a clinician or administrator attempts to modify the Snapshot after signature, the hash mismatch is detected by the audit system and flagged as a tamper event.

• The timestamp on the hash anchors the document to a specific moment — establishing exactly when the patient's preferences were recorded.

This architecture meets the evidentiary standards described in the ONC Health IT Certification requirements for data integrity and exceeds the documentation standards required by most state advance directive statutes.

Interoperability: Where the ACP Snapshot Surfaces

Implementation Roadmap for Palliative Care Medical Directors

Deployment follows a 72-hour fast-track protocol for sites running Epic or Oracle Health (Cerner). Other EHRs follow a 2-week integration path.

Phase 1: Configuration (Hours 0–24)

• Map discrete ACP fields to existing EHR build or create net-new fields (Surrogate Decision Maker, QOL Thresholds, Preferred Setting, Voluntariness)

• Configure the GOC prompt library within Scribing.io's ambient engine

• Activate billing rules: 99497/99498 time thresholds, modifier 33 AWV detection chain

• Enable SHA-256 hashing module and FHIR DocumentReference write path

Phase 2: Governance and Testing (Hours 24–48)

• Conduct pilot GOC conversations with 2–3 clinicians using test patients

• Verify discrete field writes populate correctly on patient banner, ED summary, and ICU handoff

• Validate billing output: correct CPT, correct modifier logic, correct ICD-10 array

• Review ACP Snapshot for completeness and hash integrity

• Obtain compliance sign-off on the governance pack (clinician-in-the-loop attestation workflow, audit trail documentation)

Phase 3: Go-Live and Monitoring (Hours 48–72+)

• Deploy to full palliative care team

• Monitor extraction accuracy dashboards (target: >95% surrogate identification rate, >90% QOL threshold capture rate on first pass)

• Weekly billing reconciliation: compare ACP encounters documented vs. charges generated vs. charges paid

• Monthly compliance review: ACP Snapshot integrity audit (hash validation), modifier 33 omission rate, Z-code specificity compliance

Book a 15-minute Workflow Audit to deploy the GOC Legal Shield: prebuilt prompts that auto-extract Surrogate Decision Maker and Quality-of-Life thresholds, EHR-mapped discrete fields + immutable ACP Snapshot, and billing rules that auto-suggest modifier 33 for 99497/99498 — live in Epic/Cerner in 72 hours with your governance pack. Schedule at Scribing.io →

Frequently Asked Questions: GOC AI Prompts, Compliance, and Revenue Recovery

Does the AI record the conversation?

Scribing.io's ambient engine processes audio in real time for entity extraction. Audio is not stored after the encounter unless the site's policy requires it. The output is structured text — the five extraction layers — reviewed and signed by the clinician. No raw audio is written to the EHR or the ACP Snapshot.

Can non-physician clinicians bill 99497/99498?

Yes. CMS permits billing by any qualified healthcare professional authorized to independently bill Medicare E/M services, including nurse practitioners, physician assistants, and clinical nurse specialists. Scribing.io's billing rules engine validates the rendering provider's credential type before routing the charge.

What if the patient lacks capacity at the time of the GOC conversation?

Layer 5 of the prompt library includes a capacity assessment extraction. If the clinician determines the patient lacks decision-making capacity, the system shifts the extraction target: the surrogate becomes the primary interlocutor, and the AI documents the discussion as occurring with the surrogate on behalf of the patient. The ACP Snapshot reflects this distinction explicitly, preserving legal clarity about who made which statements.

How does the system handle state-specific advance directive laws?

Scribing.io's governance module includes a state-law configuration layer. Surrogate hierarchy defaults (e.g., spouse > adult child > sibling) are pre-loaded per jurisdiction and can be customized by the site's legal counsel. When the AI detects a potential conflict between the patient's stated surrogate and the state's default hierarchy, it flags the discrepancy for clinician resolution.

What happens if a GOC conversation spans multiple encounters?

The system supports longitudinal ACP tracking. Each encounter generates its own ACP Snapshot (separately hashed, separately time-stamped). The EHR discrete fields always reflect the most recent Snapshot, but prior Snapshots are retained and accessible for audit. A visual timeline in the ACP section shows the evolution of the patient's preferences across encounters — critical for demonstrating that a change in code status was informed and voluntary, not abrupt or coerced.

Is the SHA-256 hash admissible in court?

SHA-256 hashing is widely accepted in federal and state courts as evidence of document integrity. The National Institute of Standards and Technology (NIST) classifies SHA-256 as an approved hash function under FIPS 180-4. Courts routinely accept hash-verified documents under the business records exception and electronic records authentication standards. Scribing.io's audit trail provides the chain of custody: creation timestamp, clinician identity, hash value, and any subsequent access events.

What is the expected ROI for a 10-clinician palliative care program?

Conservative modeling based on early-adopter data: a program conducting an average of 25 billable ACP encounters per clinician per month, with a baseline capture rate of 38% and a post-implementation capture rate of 80%, recovers approximately $7,600 in net new monthly revenue from previously unbilled 99497/99498 charges. When modifier 33 compliance is added for AWV co-visits (eliminating write-offs and re-filings), the total monthly financial impact approaches $9,000–$11,000. Implementation cost is recovered in the first billing cycle.

Ready to eliminate missing surrogates, recover ACP revenue, and generate legally defensible GOC documentation? Book your 15-minute Workflow Audit at Scribing.io — the GOC Legal Shield deploys in 72 hours.