[ Case Study 02 · Claims Operations Platform ]
RPA pulls the data. AI accelerates the work. The platform makes it defensible.
A regional medical billing services operation. Fifteen healthcare groups. ~150,000 claims a year. Twenty people on AR. We built the platform that combines RPA, agentic AI workflows, and audit-grade tracking into a single operating layer, so the team gets to argue with payers, not log into portals.
5 PMs
RPA Architecture Coverage
Production on eCW. Same pattern extends to athenahealth, AdvancedMD, NextGen, DrChrono.
~40%
Claims Worked / FTE / Day
Same 20-person team, more throughput, no new hires.
47 → 38
A/R Days Outstanding
Nine days off DSO, three quarters after go-live.
02 / Introduction
A regional billing services operation. Five practice management systems. One platform.
The engagement here was with a regional medical billing services organization: fifteen client healthcare groups, a twenty-person AR team, five different practice management systems across the client book. A different shape than most RCM operations, and the pressure points sit in different places. The Gravitas approach is the same one we bring to every engagement.
The buyer was a billing-services CEO whose team was drowning in the manual parts of claims work: logging into five different practice management systems to pull aging reports, copying CSVs into spreadsheets, parceling out claims by hand. Real revenue work (denial follow-up, payer calls, appeals) was waiting at the bottom of the queue.
We built the claims operations platform: a multi-tenant claims-operations platform with RPA-driven practice-management-system retrieval, agentic AI workflows embedded in the workspace (Bedrock-powered note transformation, RAG-grounded code suggestions, AI root-cause analysis on denial patterns), a 12-state claims workflow, automatic priority routing, QA audit with rebuttal escalation, and an ERA-driven payment-posting layer. Production-deployed on AWS. This case study walks the platform and the result.
How to Read This Case Study
P. 3
The Pain
Five practice management systems. Manual everything.
P. 4
The Build
RPA + claims platform architecture.
P. 5–7
The Screens
RPA log, worklist, claim workspace.
P. 8
The Impact
Throughput, DSO, denial rate.
P. 9
Voices
CEO, CFO, VP of Operations.
P. 10
Next Step
A 30-minute call to walk one workflow.
03 / Before
Twenty people. Fifteen clients. Five practice management systems. Every morning, the same routine.
The bottleneck wasn't talent. It was that the highest-paid AR analysts started every day logging into other people's software to pull the same reports they pulled yesterday.
Daily · Portal Logins
Five portals. Eight tenants. Repeat.
Every morning a senior AR analyst logged into eCW, athenahealth, AdvancedMD, NextGen, DrChrono across each client's tenant credentials. Pulled aging reports. Saved CSVs. Renamed files. Uploaded to a shared drive. ~90 minutes a day, gone before claims work started. Across fifteen client groups, this added up fast.
Daily · Manual Assignment
Spreadsheet round-robin.
The AR lead opened the master spreadsheet, sorted by aging × balance, and parceled claims out to twenty analysts by hand. No capacity check. No skill match. No audit of who got what. Claims with high recovery odds sometimes landed with the analyst least equipped to work them.
Weekly · QA Review
QA happened in Excel. Or didn't.
A two-person QA team randomly audited 5% of worked claims. Findings went into a Word document. Rebuttal happened in email threads. Patterns nobody could see in aggregate. OIG-style compliance? Aspirational.
Monthly · ERA Posting
Payments posted by line, by hand.
ERA 835 files arrived from clearinghouses. A finance staffer opened each one in a parser tool, matched line items to claims by hand, posted payments one at a time. Write-offs documented in notes nobody read.
The Real Cost
“My most experienced AR people were spending their first 90 minutes a day in browser tabs.”
A twenty-person team that on paper was supposed to drive collections, follow up on denials, work appeals, audit coding. In reality, they were a portal-login service three hours of every workday.
04 / What We Built
A multi-tenant claims platform. RPA at the edges. Agentic AI inside.
A headless browser logs into each practice management system on a schedule, exports the aging reports, drops them in cloud storage. A claims engine ingests, ranks, and routes the work. A QA layer audits the output. The team handles the conversations that close the dollars.
1
RPA Worker
Firefox headless on EC2. Solves Turnstile CAPTCHA. Exports claims CSV.
2
S3 + Queue
CSVs landed in S3, jobs queued in Redis, Celery picks them up.
3
Ingest + Route
Column-mapping per PM, dedup, priority score, auto-assign every 4 hrs.
4
Workspace
12-state claim workflow. Action log. @mention notes. Audit trail.
5
QA + Payment
Rebuttal-aware audit. ERA 835 posting with line-item allocation.
Agentic AI Workflows
Bedrock-powered note transformation in the claim workspace. RAG-grounded code suggestions (Claude + ICD-10/CPT RAG, no hallucinated codes). AI root-cause analysis on denial patterns. Automated report generation for the financial report card, weekly KPIs, and ad-hoc client reports. All HIPAA-grade.
Multi-Tenant + Audit-Grade
Three-tier Platform → Tenant → Organization hierarchy. Row-Level Security enforces tenant isolation at the database layer, fifteen client groups isolated. 12-state claim workflow, every action timestamped, every QA call defensible. OIG-aligned compliance built in.
Priority Routing + Observability
Auto-assignment runs every 4 hours, ranked by aging × balance × capacity. Twenty analysts get queues matched to their skill level. RPA runs capture step-by-step screenshots, failures surface in the dashboard, manual retry is one click.
Cloud-Native Runtime
AWS ECS Fargate + RDS PostgreSQL + ElastiCache Redis + S3 + Bedrock. Scales with claim volume, not headcount. Field-level AES-256 PHI encryption. Cloud infrastructure cost: very minimal, tracks actual workload not seat count.
Practice Management System Coverage
The RPA architecture is production-deployed for eCW. The same Firefox + CAPTCHA + export pattern extends to the other major US practice management systems. Per-portal tuning of the navigation script and CAPTCHA layer is approximately 2-3 days of engineering per system.
▸ eCW · Live▸ athenahealth▸ AdvancedMD▸ NextGen▸ DrChrono
05 / Screen 1 · RPA Execution Log
The bot's diary. Every login, every export, every retry.
Claims Operations Platform
Claims
Coding
Payments
Portfolio
Settings
ops@gravitashc.comSign Out
RPA Settings
Credentials · Jobs · Execution Log
▾ All PM Systems
Run Now
↻ Refresh
Execution History
● Active| Trigger | PM System | Started | Duration | Status | Error | Download | |
|---|---|---|---|---|---|---|---|
| ▸ | Scheduled | eCW | May 16, 2:00 AM | 3m 41s | Success | - | ⤓ CSV |
| ▸ | Scheduled | eCW | May 15, 2:00 AM | 4m 02s | Success | - | ⤓ CSV |
| ▸ | Manual | eCW | May 14, 4:18 PM | 3m 28s | Success | - | ⤓ CSV |
| ▾ | Scheduled | eCW | May 14, 2:00 AM | 1m 12s | Failed | Turnstile timeout · retry | - |
| ▸ | Scheduled | eCW | May 14, 2:14 AM | 3m 56s | Success | - | ⤓ CSV |
| ▸ | Scheduled | eCW | May 13, 2:00 AM | 3m 34s | Success | - | ⤓ CSV |
Why this screen matters. RPA without observability is a black box. This page is the receipt: every nightly login, every export, every retry. When something breaks (and it does, because practice management systems push UI changes and CAPTCHAs evolve), the failure surfaces with a one-click manual retry. The bot has a paper trail.
06 / Screen 2 · Claims Worklist
The queue an analyst sees Monday morning, ranked by what pays.
Claims Operations Platform
Claims
Coding
Payments
Portfolio
Settings
analyst@gravitashc.comSign Out
Claims Worklist
Your assigned claims · 47 active
▾ All Payers
▾ 90+ Days Aging
+ Add Claim
⤓ Export CSV
Total Claims
247
Total Balance
$1,234,567
Follow-Ups Today
12
High Priority
8
| Claim # | Patient | Service Date | Payer | Balance | Aging | Status | Assigned | Follow-Up | |
|---|---|---|---|---|---|---|---|---|---|
| ☐ | CLM-89241 | Patient #4521 | Feb 12, 2026 | BCBS · TX | $2,840.50 | 94d | Appealed | M. Lopez | May 18 |
| ☐ | CLM-89238 | Patient #3987 | Feb 14, 2026 | UHC | $1,920.00 | 92d | Denied | M. Lopez | May 17 |
| ☐ | CLM-89197 | Patient #5102 | Feb 19, 2026 | Aetna | $1,640.75 | 87d | In Progress | M. Lopez | May 17 |
| ☐ | CLM-89154 | Patient #2841 | Feb 24, 2026 | Cigna | $1,210.30 | 82d | Submitted | M. Lopez | May 19 |
| ☐ | CLM-89102 | Patient #1923 | Mar 4, 2026 | BCBS · CA IPA | $980.00 | 73d | In Progress | M. Lopez | May 20 |
| ☐ | CLM-89087 | Patient #4108 | Mar 8, 2026 | Medicare | $720.40 | 69d | Paid | M. Lopez | - |
This is the page that ate the spreadsheet. Forty-seven claims, sorted by recoverability × dollars × payer behavior. The top of the list is the right work for today. Color-coded aging buckets, payer-specific status, follow-up dates that the analyst sees the moment they log in. No spreadsheet. No manager handing out work. The queue arrives ranked.
07 / Screen 3 · Claim Workspace
Everything the analyst needs on one page. Everything the auditor needs in the log.
Claims Operations Platform
Claims
Coding
Payments
Portfolio
Settings
analyst@gravitashc.comSign Out
Claim Workspace
Working on claim CLM-89241 · BCBS Texas
⏱ 04:21● High Priority
Patient
Patient #4521
Service Date
Feb 12, 2026
Total Charges
$3,420.00
Current Balance
$2,840.50
Aging
94 days
Workflow Status
Appealed
Assigned To
M. Lopez
Follow-Up
May 18, 2026
📋 Work Claim
🛡 Quality
🕒 History
Action Type
Appeal Submitted
Status Code
CARC-50 Med. Necessity
Sub-Status
Awaiting Payer Response
Follow-Up Date
May 18, 2026 📅
Notes (Required)
Submitted Level 1 appeal with medical-necessity documentation from rendering provider. Referenced 3 sibling claims (CLM-87102, CLM-87445, CLM-88019) that paid on the same CPT × payer combination. Expecting determination within 30 days.
Action History
May 16, 9:14 AM · M. Lopez
Appeal Drafted
Pulled reference cohort, drafted Level 1 letter against payer policy citation.
May 14, 3:42 PM · M. Lopez
Payer Called
Confirmed denial reason code (CARC 50). Rep stated appeal route is medical necessity.
May 12, 11:08 AM · System
Denial Received
ERA 835 posted denial. CARC 50, RARC N700. Balance updated.
Every action timestamped. Every change versioned. Every claim, timed. See the timer at the top-right of the workspace (“⏱ 04:21”)? Every analyst's time on every claim is captured automatically. That's not surveillance, it's accountability data: coaching opportunities surface, time-per-claim becomes a real KPI, and managers stop guessing how their team spends the day. Combined with the action log, the workspace produces an audit trail no spreadsheet can match.
08 / After · Three Quarters In
~4 FTE saved. DSO down 9 days. ~$6.6M annual financial impact.
FTE Saved
~4
~$260K/year in loaded labor cost, returned to higher-value work.
DSO Improvement
47 → 38
9 days off A/R = ~$3.13M working capital unlocked.
Collections Lift
+$3.19M
Denial rate dropped 8% → 5.5%. ~3,750 claims rescued/year.
Combined Annual Impact
~$6.6M
Labor + working capital + collections, on an ~$127M A/R book.
⏱ Per-Claim Time
Accountability the spreadsheet couldn't deliver. Every analyst's time on every claim is captured automatically by the workspace timer. Average time-per-claim by team, by analyst, by payer, by denial type, all queryable. Managers stop guessing. Coaching gets specific. Underperformers can't hide; overperformers get the credit they earned.
How We Got There · The Math
All inputs from the operation profile + industry-standard loaded costs.
1.5 hrs/day saved per analyst (portal logins + manual assignment, eliminated)
× 20 analysts × 22 days
= Monthly hours returned to the team
~660 hrs / month
= FTE equivalent (at 160 hrs / FTE / month)
~4 FTE
× Loaded cost per US AR analyst ($65K / yr)
= ~$260K / yr labor recovered
DSO drop: 9 days × ($127M A/R book ÷ 365)
= ~$3.13M working capital unlocked
Denial rate: 150K claims × 2.5 pts × $850 avg balance
= ~$3.19M additional collections / yr
Agentic AI workflows: Bedrock note transformation cuts action-note time ~60%. RAG code suggestions reduce medical-necessity denials at coding stage.
drives the +40% throughput
Combined Annual Financial Impact
~$6.6M
Calculate Yours · Plug in Your Numbers
For an RCM operation your size: your annual claim volume × your avg claim balance × 2.5 pts denial reduction = your annual collections lift. Add your DSO improvement (days) × your A/R book ÷ 365 = working capital unlocked. Plus your saved FTE × $65K loaded cost in labor recovered. The math scales linearly with your book.
09 / Voices
In their own words.
Three executive vantage points on the same engagement. The CEO who signed the contract. The CFO who watched the cash-flow math move. The VP of Operations who carried the rollout.
From the CEO
“The number that closes me at the board meeting is the $6.6M. The number that keeps my contracts is the audit trail. When a client asks why a $2,800 claim sat for nine days, I don't say ‘let me check’ anymore. I pull up the action log and walk them through every touch, every decision, every payer callback. That's a totally different relationship. The automation paid for itself in working capital alone. The trust I bought with the audit trail, I can't put a number on.”
Chief Executive OfficerRegional medical billing services · 15-client book
From the CFO
“Nine days off DSO on a $127M book is roughly $3M in working capital. That's not a soft KPI. That's a credit-line conversation with my bank that I don't have to have anymore. Add the labor and the denial recovery, and the engagement was net-positive inside the first quarter. I built the ROI model in an afternoon and the math holds.”
Chief Financial OfficerRegional billing services · 15-client book
From the VP of Operations
“I've rolled out three platforms in my career and this one is the only one my team didn't push back on. Why? Because it took the worst part of their job, the portal-login routine, and just deleted it. The first morning the bot ran successfully, I had analysts coming up asking what else we could automate. That's not the reaction you get to a typical rollout.”
VP, OperationsOwns RCM delivery · 20-analyst team
Names and practice identities withheld at the request of our client. Live customer references available for qualified discovery conversations.
[ next step ]
Pick the practice management system that costs your team the most mornings. We'll automate the login.
A 30-minute call to walk one workflow. If the build math works, we scope the sprint. If it doesn't, you've spent 30 minutes and you understand your own portal-login tax better than you did this morning.
Book a Working SessionWe build software for any workflow in your revenue cycle.