Prior Authorization, Handled End to End by an AI Agent.

A human reviews every packet at launch. The agent drafts; the PA team or clinician approves. We only consider auto-submission after the agent has been in production for months with sustained accuracy on a specific category — and even then, only on low-clinical-risk PA types (durable medical equipment, routine imaging) where mistakes are recoverable. Cardiology, oncology, and complex specialty PA stays human-in-the-loop indefinitely.

EHRs: Epic, Cerner, athenahealth, NextGen, eClinicalWorks via FHIR R4 — and any system with a documented FHIR endpoint. Payers: direct API integration with Availity, CoverMyMeds, Surescripts, and Cohere Health covers most US national payers. For payers without API access, we automate the portal. For UAE deployments, integration with Daman, Thiqa, ADNIC, AXA, and eClaimLink for DHA-DRG.

Two layers. First, the policy corpus is re-ingested on a schedule — daily for payers with frequent updates, weekly for stable payers. Second, drift monitoring watches for sudden denial-pattern changes that imply an undocumented policy shift. When a payer changes their PA criteria, the agent flags affected templates within 24-48 hours rather than silently submitting under old rules.

Three mechanisms. (1) The agent uses RAG with citation-required output — every clinical claim in the packet must trace to a specific chart section or payer policy clause. (2) Generation is constrained to extract-and-rephrase, not free synthesis. (3) Every draft is reviewer-approved before submission. We have never shipped a PA agent without human approval in the loop, and we would not recommend one.

Every PA the agent drafts is tracked through to its final outcome — approved, denied, pended, withdrawn. Denial reasons are classified and fed back into the eval set. Quarterly, we retrain the criteria-matching models on the latest outcomes and update payer-specific templates that have shifted. We share accuracy dashboards by payer, service line, and reviewer so your team sees the same numbers we do.

A focused deployment — one EHR, top 5 payers, top 20 PA types — typically takes 12-16 weeks: 4 weeks scope and policy modeling, 6-8 weeks shadow mode and clinician tuning, 2-4 weeks controlled rollout. Multi-payer, multi-specialty enterprise rollouts run 6-9 months with phased expansion. We do not promise faster timelines than the clinical sign-off cycle allows.

Build cost typically runs $250K-$700K for a focused deployment, scaling with payer count, PA type breadth, and EHR complexity. Run cost is dominated by LLM tokens — for a typical mid-size health system processing 5,000-10,000 PAs/month, infra and model costs land in the $4K-$12K/month range. ROI usually shows up within 6-9 months on labor savings alone, with the bigger win being the recovered revenue from PAs that previously slipped through the cracks.

The architecture is the same; the integration layer is different. For UAE deployments we run the agent inside UAE data residency — AWS UAE or G42 Cloud — with no PHI leaving the country for inference. Payer integration uses eClaimLink and direct TPA APIs (Daman, Thiqa). NABIDH/Malaffi integration covers chart retrieval. All access is ADHICS-aligned with audit trails ready for DHA review.

Off-the-shelf PA platforms work well for high-volume, standardized PA types and standard payer integrations. They are weaker on: (1) complex specialty PA where your clinical workflow is unique, (2) payers and TPAs they do not cover (especially regional/UAE), (3) integration with your specific EHR customizations and downstream RCM workflow, (4) ownership and audit trail — your data, your models, your eval. The right answer is often a hybrid: use a vendor for the standard 60%, build the agent for the high-value 40%.