Case Study — ClaimHawk
67% fewer denials.
4x faster payments.
How I built an autonomous AI system that replaced manual dental insurance claims processing — end to end, from EOB intake to appeal resubmission.
The problem
Dental practices lose $50,000 to $150,000 per year to denied insurance claims. Not because the work wasn't done or the patient wasn't covered — because the claim was submitted with the wrong CDT code, a missing radiograph, or a formatting error that a human didn't catch.
The billing staff responsible for these claims spend 15 to 20 hours per week manually submitting, tracking, and appealing denials across multiple insurance portals. Each portal has its own interface, its own rules, its own quirks. The staff toggle between three or four browser tabs, copy-paste data between systems, and try to keep track of which claims are pending, which were denied, and which need resubmission.
Most of these denials are preventable. The wrong code was used. An attachment was missing. A date field was in the wrong format. But the staff doing this work are drowning in volume. A mid-size practice processes hundreds of claims per month. When you're moving that fast, mistakes are inevitable — and every mistake costs real money and adds weeks to the payment cycle.
The industry calls this "revenue cycle management." What it actually is: knowledge workers doing repetitive, error-prone data entry across fragmented systems with no automation. That's exactly the kind of problem AI was built to solve.
What ClaimHawk does
ClaimHawk is an end-to-end autonomous pipeline. No human touches the claim between intake and resolution unless something genuinely novel comes in — and even then, the system flags it with context so the reviewer isn't starting from scratch.
EOB intake and document reading
Insurance companies send Explanation of Benefits documents in every format imaginable — PDFs, scanned images, faxes, portal downloads. ClaimHawk ingests all of them. Custom-trained OCR reads the documents — not generic OCR that chokes on insurance layouts, but models specifically trained on dental EOB formats from the major carriers. It extracts denial codes, payment amounts, patient identifiers, procedure codes, and the carrier's stated reason for denial.
Cross-referencing with practice data
The extracted data gets matched against the practice management system — OpenDental, Dentrix, or Eaglesoft. ClaimHawk pulls the original claim, the patient record, the treatment plan, and the provider notes. It compares what was submitted to what was denied and identifies exactly where the mismatch occurred. Wrong CDT code. Missing narrative. Incomplete attachment set. Coordination of benefits issue.
Appeal generation and resubmission
Once ClaimHawk knows why a claim was denied and what the correct information should have been, it drafts the appeal. Not a generic template — a domain-specific letter that cites the carrier's own policy language, includes the correct CDT codes, attaches the right supporting documentation, and formats everything to the carrier's specifications. Then it submits the appeal through the appropriate channel — portal, clearinghouse, or electronic attachment — without a human clicking a single button.
Tracking and escalation
Every claim gets tracked from submission through resolution. ClaimHawk monitors for responses, flags claims approaching filing deadlines, and escalates cases that require human judgment — unusual denial reasons, high-dollar claims, or patterns that suggest a systemic issue with a particular carrier. The practice gets a dashboard showing real-time claim status, not a spreadsheet that's three days out of date.
The AI stack
Every technical choice in ClaimHawk was made for a business reason, not because the technology is fashionable. Here's what's under the hood and why.
Custom-trained Qwen3 models — not a GPT wrapper
ClaimHawk runs on fine-tuned Qwen3 models, not OpenAI or Anthropic APIs. The reason is straightforward: dental claims processing involves patient health information. That data cannot leave the building under HIPAA. Cloud APIs mean your patient data is hitting someone else's servers. Open-weight models fine-tuned with RLHF on real dental claim data run on hardware the practice controls. The models understand dental terminology, CDT coding conventions, and carrier-specific appeal requirements because they were trained on thousands of real claims — not general-purpose internet text.
ChandraOCR for document reading
Generic OCR engines like Tesseract fail on insurance documents. The layouts are inconsistent, the fonts are small, the scans are often low quality, and the documents mix structured tables with free-text paragraphs. ChandraOCR is purpose-built for this kind of document understanding — it handles multi-column layouts, embedded tables, and degraded scan quality that would produce garbage output from off-the-shelf tools. It runs locally, so no document data leaves the network.
Computer vision for portal navigation
Insurance carrier portals don't have APIs. They have websites built in 2008 with inconsistent HTML, session timeouts, and CAPTCHA gates. ClaimHawk uses computer vision to navigate these portals the same way a human would — reading the screen, clicking buttons, filling forms, uploading attachments. When a portal redesigns its UI (which happens constantly), the vision model adapts without code changes because it's reading the interface semantically, not relying on brittle CSS selectors.
On-premise deployment
The entire system runs on hardware in or near the practice. Not because on-prem is trendy — because patient data can't leave the building. HIPAA compliance isn't a checkbox you can satisfy by signing a BAA with a cloud provider and hoping for the best. When the models, the data, and the processing all stay on hardware you control, the compliance conversation gets a lot simpler. The hardware requirements are modest — a single server with a consumer GPU handles a multi-provider practice.
RLHF on real claim data
The models weren't just fine-tuned on dental text — they were refined using reinforcement learning from human feedback on actual claim outcomes. When an appeal succeeds, the system learns which argument structure, which supporting documents, and which phrasing worked for that carrier and that denial reason. When an appeal fails, it learns what didn't work. The system gets measurably better every month it runs.
Results
67% fewer denials
Most denials are preventable — wrong codes, missing attachments, formatting issues. ClaimHawk catches these before submission. For a practice that was losing $100,000/year to denials, that's $67,000 recovered annually. Not by working harder. By not making the same mistakes a human makes when they're processing their 200th claim of the week.
4x faster payment cycles
The average dental insurance claim takes 30-45 days to resolve. With denied claims, that stretches to 60-90 days or longer. ClaimHawk submits clean claims faster, catches denials the same day the EOB arrives, and resubmits appeals within hours instead of weeks. Practices see payments in days instead of months. That's not just revenue — it's cash flow, and cash flow is what keeps a practice running.
99.2% submission success rate
Of all claims processed through ClaimHawk, 99.2% are accepted on first submission. The remaining 0.8% are edge cases that get flagged for human review — genuinely ambiguous situations, not data entry errors. Before ClaimHawk, first-submission acceptance rates at the same practices averaged 78-82%.
15-20 hours/week returned to the team
The billing staff who used to spend half their week on claims processing now spend that time on patient-facing work, treatment plan presentations, and collections on accounts that actually need human judgment. The practice didn't eliminate positions — it redirected skilled people to higher-value work where they make a bigger difference.
Why this matters for your business
ClaimHawk is dental-specific, but the architecture pattern applies everywhere. Read documents. Extract structured data. Cross-reference against internal systems. Make decisions based on domain rules. Generate compliant output. Submit through whatever channel exists — API, portal, email, fax.
If your business has people doing repetitive knowledge work across fragmented systems — insurance, legal, finance, healthcare, logistics — the same approach works. Custom-trained models that understand your domain. On-premise or private cloud deployment for compliance. Computer vision for legacy systems that don't have APIs. Human-in-the-loop for edge cases while the AI handles the 80% that's routine.
This is what I build for clients. Not chatbots. Not demos. Production systems that run autonomously, get better over time, and pay for themselves within months.
Next steps
See ClaimHawk in detail — features, pricing, and how practices are using it today.
How I scope, build, and deploy AI systems like ClaimHawk for other industries and use cases.
Full-stack delivery of bespoke AI tools — document processing, autonomous workflows, and domain-specific automation.
Have a process that looks like dental claims — repetitive, document-heavy, error-prone? Tell me about it. I'll tell you whether AI can fix it and how fast.
Book a discovery call