Their PA review team was buried under 1.4 million prior-auth requests a year. We built a HIPAA-eligible four-agent AI platform on CrewAI + Anthropic Claude 3.5 Sonnet — and unlocked 85% automation, a 9-minute-to-38-second turnaround, and $6.2M in annualised savings inside 12 months. Faster approvals. Happier physicians. Patients back to care.
Discuss Your Project
Skim the case in 60 seconds — what was broken, what we built, what changed in year one. Then dig deeper below.
Meet the network. A regional integrated health system — 18 hospitals, 142 outpatient clinics, 4,800 physicians and 28,000 staff — moving 1.4M prior authorisations a year through a single, increasingly painful review pipeline.
Five forces were grinding the prior-authorisation desk down at the same time — volume, cost, payer complexity, EHR sprawl, and a physician-satisfaction score that wouldn't stop falling.
Five steps, end to end — from the moment an order needs prior-auth to the moment the decision is logged back in the EHR. Most of those steps, the AI now runs entirely on its own.
Four architectural pillars carry the load. Each was picked for a specific reason; together they're what turned prior auth from the slowest step in care to the fastest.
Instead of a single monolithic clinical chatbot, the system runs as a crew of four specialist agents — eligibility & benefits, medical necessity, documentation completeness, and submission — coordinated by CrewAI. Each agent has a narrow job, a tight prompt, and a focused set of tools. CrewAI routes the PA between them, holds shared state across handoffs, and decides when a human reviewer needs to step in. The result behaves like a senior PA reviewer with a perfect memory of every case the network has ever processed.
Claude 3.5 Sonnet does the actual clinical reading — the medical note, the imaging report, the prior-treatment history. We chose it for long-context reasoning across multi-page charts, native tool use for calling EHR APIs, and a writing style that produces PA narratives a payer reviewer can read without re-translation. Carefully designed system prompts encode the network’s PA-writing voice, the things the agent must never claim, and the explicit escalation rules. PHI is never used to train foundation models — Claude runs in a private deployment with zero-retention.
The crew’s institutional memory is an Amazon Bedrock Knowledge Base indexing 285,000 historical PA decisions across the network — every approval, every denial, every payer policy snapshot at the time the decision was made. Every PA the AI drafts is retrieval-augmented and cited back to the historical decisions and policy text that support it. Nightly refresh keeps the index in sync with new payer policy releases — no stale logic when a payer changes its medical-necessity criteria.
The agent crew lives inside the same EHR worklist reviewers already use. PA cases route through Epic, Cerner and Allscripts via FHIR R4 (US Core profile) plus HL7 v2 where required. When confidence drops or the case is novel, the conversation is handed to a human reviewer with the full agent transcript, the drafted PA package and every cited payer-policy clause attached — no "tell me the case again" moment. Every reviewer override is captured and fed back into the weekly retraining pipeline, so the crew gets measurably better month over month.
Year-one results, measured the way a health-system CFO measures them — automation rate, turnaround, denial outcomes, physician satisfaction. No vanity metrics.
Prior-auth automation rate — eight in ten PA reviews close without a human reviewer
Average PA turnaround, cut by roughly 93%
Annualised savings unlocked in the first 12 months of production
Physician PA-process satisfaction lift — from 18% to 65%
Annual PA requests now flowing through the four-agent crew
HIPAA, SOC 2 and internal-audit findings in year one
From AI driven platforms to enterprise software and custom applications, businesses across industries rely on DreamzTech to accelerate innovation, improve operations, and drive measurable growth. Let’s discuss how we can help your business next.
Six purpose-built AI agent platforms — one for every operations team. Pick a use case, see exactly what we ship.
Explore how DreamzTech helps businesses automate document heavy workflows across legal and insurance sectors with faster processing, improved accuracy, and measurable ROI.
Their Tier-1 support desk was drowning in 140,000+ monthly tickets. We rebuilt it as an autonomous multi-agent AI on LangGraph + Anthropic Claude 3.5 Sonnet — and unlocked 75% ticket deflection, $2.1M in annual savings and a 9-point CSAT lift inside 12 months. Faster answers. Happier customers. Zero added headcount.
An 85,000-employee global retail bank automated its IT service desk with a LangGraph hierarchical multi-agent platform on OpenAI GPT-4o + Microsoft AutoGen specialists with deep ServiceNow integration. Year one: 18,000 L1 tickets auto-resolved monthly, $1.8M direct cost saved and zero SOX 404 ITGC findings.
A Series D B2B cybersecurity SaaS replaced manual lead qualification with a four-agent CrewAI sales crew on Anthropic Claude 3.5 + OpenAI GPT-4o with native Salesforce + HubSpot + Apollo + ZoomInfo + 6sense integration. Year one: 4.2× lead-to-SQL conversion lift, $14.2M Q1 pipeline and 67% SDR productivity gain.
A $1.2B-revenue specialty chemicals manufacturer cut month-end close from 11 working days to 3 with a LangGraph + Temporal multi-agent finance platform integrating SAP S/4HANA + Oracle E-Business Suite + NetSuite. Year one: $850K annualised savings, 9,400 monthly journals automated and zero SOX 404 ITGC findings.
A regional integrated healthcare network with 18 hospitals and 142 outpatient clinics automated 85% of prior-authorisation reviews with a HIPAA-eligible four-agent CrewAI platform on Amazon Bedrock + Anthropic Claude 3.5 (fine-tuned). Year one: $6.2M annualised savings, physician satisfaction lifted from 18% to 76% and 64% drop in cancelled procedures.
Explore our comprehensive AI solutions tailored for different industries, including AI agents, multi agent systems, automation platforms, consulting, and enterprise AI integrations.
The questions health-system CIOs, CMIOs and revenue-cycle leaders ask us most when scoping a PA-automation build like this one.
Tier-1 routine categories — high-volume imaging (MRI, CT, PET), DME, common specialty drugs, elective procedures with well-defined criteria, behavioural-health authorisations following published payer guidelines, and any PA where the medical-necessity criteria are explicit and the documentation already exists in the EHR. Anything novel or high-cost still routes to a clinician reviewer.
The case is handed to a human reviewer inside the EHR worklist with the full agent transcript, the drafted PA package, every cited payer-policy clause and the agent’s confidence score attached. The reviewer either approves, edits or denies — and the override is captured for the next retraining cycle.
Yes. HIPAA BAA in place with the cloud provider, customer-managed encryption keys, TLS 1.3 in transit, immutable audit trails in Amazon CloudWatch for HIPAA and SOC 2 evidence. PHI is never used to train foundation models — Anthropic Claude 3.5 Sonnet runs in a private deployment with zero retention.
Fourteen weeks to a single-hospital production pilot. Twenty-two weeks for full network rollout across all 18 hospitals plus the 142 outpatient clinics. The health system in this case went from kickoff to year-one outcomes — 85% automation, $6.2M savings, +47 physician-satisfaction points — inside 12 months.
Yes. We integrate via FHIR R4 (US Core profile) and HL7 v2 where required, with native support for Epic, Cerner (Oracle Health), Allscripts, Athena, Meditech, eClinicalWorks and NextGen. Worklist round-tripping into the reviewer’s EHR queue keeps the existing clinician workflow intact.
Every reviewer override is captured and fed into a weekly retraining job against the 285,000-decision history and the latest payer-policy snapshots. Per-payer approval and denial rates are monitored on dashboards so confidence thresholds can be tuned where the crew is over- or under-approving. Most engagements see automation rate climb another 5–10 percentage points in the first six months after launch.
