Industries/Healthtech/Prior-authorization AI
Prior-authorization AI
AI prior-authorization decisions sit at the intersection of HIPAA audit controls, payer-network contract obligations, and increasing state-level transparency mandates (California SB 1120 utilization-review human-review rule, plus Texas SB 1188 for any diagnostic-AI use case). Promethean records every prior-auth recommendation as an L12 entry — the spec, the model, the rationale, the human reviewer sign-off — making contested denials auditable rather than reconstructed from logs.
What this is
Prior-authorization (PA) AI typically does two things: classifies medical-necessity criteria fit (approve / further-review / decline) from a submitted treatment request, and drafts the narrative justification accompanying the submission. Volume at healthtech-RCM SMB scale runs 10k–500k decisions per month across customers.
The regulatory perimeter is dense: HIPAA §164.312(b) audit controls for any system handling ePHI; payer contracts often require demonstrable consistency in PA criteria application; state laws (CMS Interoperability + Prior Authorization Final Rule effective 2026-2027, California SB 1120 2025, Texas SB 1188 2023) add transparency + appeals-rights requirements that fold back into the AI's operating envelope.
Promethean records each PA decision as an L12 entry. The closed-enum output schema (approve / further-review / decline + criteria-matched array) bounds the AI; the reviewer-gate primitive captures the revenue-cycle specialist or clinician sign-off; the chain export is what a provider's denial-management team needs when contesting a payer's adverse determination.
Regulatory pressure specific to Prior-authorization AI
HIPAA §164.312(b) — audit controls
Covered entities must implement hardware, software, and procedural mechanisms that record + examine activity in information systems containing ePHI. AI prior-auth decisions on ePHI fall under this.
Official source ↗CMS Interoperability + Prior Authorization Final Rule (CMS-0057-F, 2024)
Phased compliance through 2026-2027. Payers must publish prior-auth metrics, support FHIR-based PA APIs, and provide decisioning rationale. AI used in PA workflows is in scope for the rationale + appeals requirements.
Official source ↗California SB 1120 (2025) — Physicians Make Decisions Act
Prohibits algorithms from being the sole basis for denying or modifying medically necessary care in utilization-review decisions by health-plan insurers; requires licensed clinician review of AI-assisted UR decisions. Effective 1 Jan 2025.
Official source ↗Texas SB 1188 (89R, 2025) — practitioner AI-use disclosure + EHR localization
Healthcare practitioners using AI for diagnostic purposes must disclose AI use to patients, operate within scope of licence, and review AI-generated records per Texas Medical Board standards. General effective date 1 Sept 2025; EHR data-localization 1 Jan 2026.
Official source ↗How Promethean addresses this
Per-decision medical-necessity criteria evidence
The spec's outputSchema enumerates the criteria checked; each L12 entry's outputCanonicalHash binds the criteria-matched vector. A denial appeal can be answered with 'these specific criteria were evaluated, this was the outcome'.
Clinician / RCM-specialist sign-off evidence (CA SB 1120 utilization-review human-review rule)
Reviewer-gate set to always-human for any decline; the L12 entry records the reviewer identifier + verdict. Over the audit window, the override rate proves human review was meaningful. Texas SB 1188 (89R, 2025) layers on top for any diagnostic-AI use case, separately from the PA flow.
CMS PA Rule rationale requirement
Each PA decision is committed with the rationale field (length-bounded by schema); outputCanonicalHash binds the rationale to the entry. The rationale is retrievable for FHIR PA API responses + appeals workflows.
HIPAA Audit Controls (§164.312(b))
L12 chain provides hardware-and-software-enforced recording: the substrate emits per-decision entries; the operator cannot disable emission. Per-entry Ed25519 attestation provides integrity. Combined with operator-side SIEM (which Promethean doesn't replace), this satisfies the audit-controls implementation specification.
Multi-payer-customer tenant isolation
Healthtech-RCM SMBs typically serve many provider customers; R4 tenantId on each L12 entry lets you ship per-customer audit bundles filtered to their decisions only. Production tier and above.
What this looks like in practice
Hypothetical: state-AG inquiry on AI-driven PA denials
A state Attorney General announces an inquiry into AI-driven prior-auth denials following patient-advocacy complaints. The AG demands: 'Provide per-decision evidence for all AI-touched PA denials in 2026 affecting CA residents. Demonstrate clinician review per CA SB 1120. Demonstrate the criteria-matching logic was the documented one.' Without Promethean: months of cross-referencing claims-system exports with internal AI logs + a defensive narrative. With Promethean: chain-export filtered to recordedAtIso 2026 + outputCategory=decline + tenantId where the patient's payer matches the CA cohort; reviewer-verdict shows clinician sign-off on 100% of declines (CA SB 1120 satisfied); specHash sequence shows two documented model versions with documented transition dates. The AG accepts the chain output + handoff packet; inquiry concludes with corrective recommendations rather than enforcement action.
Frequently asked
We use multiple LLM providers — Anthropic for narrative drafting, OpenAI for classification. Does Promethean handle that?
Yes. Each spec declares its modelIdentity (provider/model/version) explicitly. The L12 entry records which provider+model decided which call — so a regulator can isolate 'show me every Anthropic-classified decline' vs 'every OpenAI-drafted narrative'. This level of granularity is exactly what state AI-in-healthcare laws (CA, TX, NY) are starting to demand.
How does Promethean fit with the CMS Patient Access API + FHIR PA API?
FHIR APIs are the data-exchange layer; Promethean is the audit-trail layer underneath. When a payer/RCM platform responds to a provider's FHIR PA request, the AI decisioning behind that response is what Promethean records. The FHIR layer reads the rationale + decision; Promethean records what model decided, under what spec, with what reviewer engagement.
What about the medical necessity criteria — are those proprietary?
Yes, and the substrate doesn't expose them. The spec's outputSchema enumerates criteria categories (e.g. 'meets-evidence-of-failed-prior-therapy', 'meets-imaging-criteria'); the underlying criteria definitions remain in your IP. The L12 entry's outputCanonicalHash binds which categories matched on the specific decision — verifiable by you, opaque to the substrate, queryable by the regulator if needed.
Can the L12 chain feed our appeals workflow directly?
Yes. When a provider appeals a denial, your appeals team queries the chain by inputHash (which is computable from the PA request identifiers). The entry returns the spec version, the criteria evaluated, the reviewer verdict, and the rationale. This is the foundation document for the appeal response — and it's verifiable end-to-end with verify.mjs by external auditors.