Cleared a prior-authorisation backlog with a healthcare-admin agent

The brief: A US revenue-cycle-management company processed prior authorisations on behalf of specialty clinics — gathering clinical documentation, matching it to each payer’s medical-necessity rules, submitting, and chasing. The work was a backlog machine: a 6,000-case queue, a four-day turnaround that delayed patient care, and a team that could not hire fast enough. They wanted the prior-auth service delivered by agents, with their nurses reviewing the hard cases.

What the audit surfaced

• The bottleneck was not the submission. It was assembling the clinical packet and matching it against payer-specific, frequently-changing medical-necessity criteria — slow, manual, and error-prone.
• Denials were mostly avoidable: missing documentation or the wrong criteria cited. A first-pass that caught those before submission would move the number more than anything downstream.
• This is healthcare — every action had to be HIPAA-compliant, logged, and reviewable, and clinical judgement had to stay with a licensed human.

What we built

• Wrote the eval set first. 3,500 historical authorisations — approved, denied, and appealed — re-graded by two utilisation-review nurses against the question that matters: was the packet complete and correctly matched before it went out.
• A documentation and criteria-matching agent. Pulls the relevant clinical record from the EHR via FHIR, extracts the evidence each payer’s policy requires, flags what is missing, and drafts the submission with the specific medical-necessity criteria cited. Payer rules encoded as a maintained, versioned ruleset — not left to the model.
• A nurse review layer. Anything involving clinical judgement, any low-confidence match, and a sample of clean cases route to a licensed reviewer. The agent never makes the medical-necessity call — it prepares it and shows the evidence.
• A defensible trail. Every case carries the records used, the criteria matched, and the rationale — built for the payer appeal and the compliance review before either is asked for. PHI handled inside the client’s own HIPAA-compliant environment throughout.
• Shadow then switch. Five weeks preparing every case in shadow alongside the human team, graded on completeness and match accuracy, before taking the live queue payer by payer.

What changed

• Packets prepared straight-through and cleared by a reviewer without rework: 0% → 78%.
• Median turnaround from four days to under one — clean cases same-day.
• First-pass denial rate (missing docs / wrong criteria): 23% → 9%.
• Backlog cleared from 6,000 cases to under 400 within ten weeks of full switchover.
• Cases per reviewer per day rose 3.4×, with the nurses spending their time on judgement, not data entry.

What we left behind

A prior-authorisation service the company runs as an AI-native operation: the agent prepares, the nurses decide, the trail satisfies the payer and the auditor. The payer ruleset is versioned and editable by the clinical team as policies change. The service scales by adding clinics and payers, not data-entry staff — and patients wait hours, not days. This was a “build it, we run it” engagement: Cravings operates the function against an SLA while the client’s clinical team owns the decisions.