US hospitals wrote off an estimated $262 billion in claim denials and underpayments last year. Centers for Medicare & Medicaid Services (CMS) data shows roughly 65% of those denials are never resubmitted despite about 90% being recoverable with timely follow-up. The standard revenue cycle response is more billing staff, more denial management specialists, and a larger team working the queue.
The investment doesn't close the gap. Denial rates in US healthcare have been rising for more than a decade, and adding denial management capacity hasn't reversed that trend. What most revenue cycle teams are treating as a staffing or coding problem is often something more upstream: medical claims processing systems are feeding adjudication engines data that never should have left intake in that condition. Denial management doesn't fix a data quality problem. It just processes it more expensively.
The Document Problem Behind Every Denied Claim
The documents that medical claims processing systems handle aren't built for standard OCR. They're built for standardized form submission, a different requirement entirely.
The CMS-1500 (also called the HCFA 1500) is the standard form for professional claims. It has 33 box-bound fields arranged in a grid layout. Box 21 holds up to 12 ICD-10 diagnosis code pointers. Box 24 is a multi-row service table where each row contains date of service, place of service, procedure code, diagnosis code pointer, and charge amount in adjacent columns. The fields are dense, often separated by printed grid lines that scanned documents don't reproduce cleanly. Nothing flows left-to-right the way a paragraph does.
Hospital billing uses the UB-04, an 81-field institutional claim form with revenue codes alongside procedure codes, multi-page service-line tables for complex inpatient stays, and fields that mix printed and handwritten values. A single inpatient encounter can generate a UB-04 with dozens of service lines across multiple pages.
Revenue cycle teams also process Explanations of Benefits (EOBs) from payers. Paper EOBs have no standardized layout. Format varies by health plan. The electronic equivalent, the ANSI X12 835 Electronic Remittance Advice (ERA), is machine-readable but structurally different from the PDF EOBs that still arrive from smaller payers and government programs. Reconciling remittance data against submitted claims requires parsing adjustment reason codes from tabular formats that change by payer.
Prior authorization forms add another layer. Each health plan designs its own prior auth templates. Conditional logic fields (where a "yes" checkbox opens additional required fields) aren't handled by standard OCR without custom configuration per payer. Many still include handwritten physician attestation sections.
Medical claims software that promises to automate revenue cycle operations has to handle all of this simultaneously. In any real billing processes environment, you're processing a mix of scanned paper forms, portal-exported PDFs, faxed prior auths with handwritten sections, and EOBs from dozens of different payer formats, not a clean stream of digital CMS-1500s. Intake accuracy on every document type feeds directly into adjudication outcomes.
How Standard OCR Turns Extractable Data Into Claim Denials
Most off-the-shelf billing software processes claims documents as pixel-to-text conversion. It has no knowledge of what the extracted text represents or what rules govern valid values for a given field. On the grid-based, code-dense forms that revenue cycle automation depends on, that limitation produces specific, predictable failures. Not random noise. The same failures, in the same fields, on every similar document.
ICD-10 and CPT Codes: When a Single Character Triggers a Denial
ICD-10 has nearly 70,000 diagnosis codes, and many differ by a single character. The difference between I10 (essential hypertension) and I11.0 (hypertensive heart disease with heart failure) carries clinical and reimbursement weight. The characters OCR confuses most reliably (zero versus the letter O, one versus lowercase l, S versus 5) appear throughout medical coding notation.
A misread in box 21 of a CMS-1500 submits the claim under a wrong or nonexistent code and adjudication rejects it. The billing team gets a denial notice citing a coding error, but nobody in that workflow sees "your OCR misread the diagnosis code." The team investigates the coding, but the root cause stays invisible.
CPT modifier misreads compound the problem. Modifier -25 (significant, separately identifiable evaluation and management service) and modifier -26 (professional component) send a claim down different reimbursement pathways. A single character confusion routes it wrong. Standard OCR has no mechanism to validate that extracted values match expected coding formats. ICD-10 follows a predictable structure (letter + two digits + optional decimal extension), and a value that doesn't match that pattern is wrong. The pipeline passes it downstream anyway.
Grid Fields, Service Tables, and the Column Attribution Problem
Standard OCR reads left-to-right, top-to-bottom. When it processes box 24 of a CMS-1500, it doesn't recognize box boundaries. It reads the service table as a continuous text stream, merging adjacent field values. Date of service from box 24A bleeds into place of service from box 24B. If scan quality degrades slightly, box boundaries disappear entirely and field values merge into unusable concatenated strings.
UB-04 service table extraction fails the same way. Column misattribution (assigning a procedure code to the wrong service date, or a revenue code to the wrong service line) produces line-item errors that look like coding mistakes at adjudication but trace back to extraction misalignment.
EOB adjustment reason codes map codes like CO-45 (charges exceed fee schedule) or PR-2 (coinsurance) to dollar amounts in tabular format. A column misread produces a reconciliation mismatch between the paper EOB and the ANSI X12 835 ERA that forces manual resolution where automation should have closed the loop.
The STP Bottleneck Is an OCR Problem in Disguise
Typical straight-through processing rates for medical claims management with traditional systems run between 60 and 75%. That means 25 to 40 percent of claims require manual intervention before submission or after denial. Processing manually costs $15 to $25 per claim. Automated processing costs $3 to $5. Reprocessing a denied claim adds another $25 to $50 in billing staff time.
Most revenue cycle teams frame this as a staffing problem. The manual review queue and the billing staff time is real, but the queue grows from an extraction accuracy floor that doesn't improve when you add more headcount. More denial management specialists catch extraction errors after adjudication. They don't reduce the rate at which extraction errors occur. More staffing, same error rate, bigger queue.
Payer rejection patterns, when traced from adjudication back to the submitted claim and then to the extracted claims data, consistently reveal field-level extraction errors that surface as coding denials at the billing desk. Medical billing teams that audit denial reasons against source documents regularly find that a significant portion of coding-category denials trace back to character-level misreads at intake, not incorrect code selection by the coder.
Fixing the 65% STP rate means reducing extraction errors at the source. Claims processing software built on traditional OCR architectures has a structural ceiling on that improvement, regardless of the workflow layer built on top.
What Agentic OCR Does to a CMS-1500 That Standard OCR Can't
The architectural gap between standard OCR and agentic OCR on claims documents comes down to one thing: whether the system understands document structure before it starts reading.
LlamaParse applies layout-aware computer vision before extraction begins. On a CMS-1500, this means identifying box 21 and box 24 as discrete bounded fields (spatial zones with specific content types), rather than reading the page as a continuous text stream. The grid layout that causes standard OCR to merge adjacent field values is, for a layout-aware system, structural information that determines how extraction proceeds. The agentic document extraction approach treats each field zone as a typed slot with validation rules, not a region to transcribe.
| Document Zone | Standard OCR Failure | Agentic OCR Approach | Claims Consequence |
|---|---|---|---|
| Box 21 (ICD-10 pointers) | Merges codes with adjacent field values; no format validation | Bounded field identification; validates against ICD-10 format pattern | Misread codes → wrong/nonexistent code submitted |
| Box 24 service table | Column misattribution; date of service bleeds into place of service | Reads table structure; routes each column to correct field type | Service line errors → line-item denials |
| CPT modifiers | Character confusion (-25 vs. -26); no modifier validation | Extracts modifier chains; validates against known modifier set | Wrong reimbursement pathway → underpayment or denial |
| Handwritten prior auth fields | High error rate; no confidence flagging | Routes to VLM; returns per-field confidence score | Silently wrong auth data → prior auth denials |
Agentic orchestration routes ICD-10 and CPT fields to code-structure-aware validation. Extracted values are checked against known coding formats before they reach the clearinghouse. A value that doesn't match the ICD-10 pattern gets flagged, not silently submitted. When adjacent box values have merged into a concatenated string, the system recognizes the merged output doesn't match the expected pattern for either field and flags both for targeted review.
The result is real time confidence scores at the field level, not just document level. A queue where billing staff see specifically which fields are flagged (box 21 line 3, box 24D service line 7) is structurally different from a denial management workflow where staff re-enter entire claims. Targeted field correction runs roughly 4 to 6 times faster.
Here's the actual difference. Structured output from a processed CMS-1500 versus the flat text dump a standard OCR pipeline produces:
html
{
"box_21_diagnosis_codes": [
{ "pointer": "A", "code": "I10", "validated": true, "confidence": 0.99 },
{ "pointer": "B", "code": "E11.9", "validated": true, "confidence": 0.97 },
{ "pointer": "C", "code": "Z79.4", "validated": true, "confidence": 0.98 }
],
"box_24_service_lines": [
{
"date_of_service": "2026-04-01",
"place_of_service": "11",
"cpt_code": "99213",
"modifier": "25",
"diagnosis_pointer": "A",
"charge_amount": 185.00,
"confidence": 0.96
}
]
}Compare that to the raw output from a standard OCR pass on the same service line: 04012026 11 99213 25 A 18500. This has no field attribution, validation, or signal that anything is wrong.
Extraction Accuracy at Scale: Revenue Cycle Applications
Higher STP rates don't create equal value everywhere. The downstream impact depends on claim volume, code complexity, and the cost profile of the denial workflow. The agentic document workflows guide covers the broader architecture. The claims-specific payoff breaks down by vertical.
Hospital Revenue Cycle Management
Institutional billing at volume runs on UB-04 forms, and each payer has its own adjudication requirements around revenue code and procedure code combinations. Hospital revenue cycle departments processing tens of thousands of claims per month absorb the STP gap as a structural labor cost. The denial management team scales with billing volume rather than shrinking as automation improves.
Higher extraction accuracy on UB-04 service tables directly reduces the institutional denial rate. A five percentage point STP improvement on a system processing 50,000 claims per month eliminates 2,500 claims from the manual review queue per cycle. At $15 to $25 per manual claim, that labor cost compounds quickly.
Specialty Practices: Radiology, Behavioral Health, Anesthesia
Specialty billing concentrates the accuracy problem for healthcare organizations running high-volume CPT-based workflows. Radiology practices submit high volumes of standardized CPT codes with complex modifier chains: technical component (-TC), professional component (-26), laterality modifiers, bilateral procedure codes. A single modifier misread changes the reimbursement category for the entire claim. At radiology volumes, modifier extraction accuracy is the primary STP determinant.
Behavioral health billing often includes handwritten progress notes and non-standard DSM/ICD crosswalk formats attached to claims. Payer-specific prior authorization requirements vary more in behavioral health than almost any other specialty, and misread prior auth data can delay patient care in ways that have consequences beyond the billing department. The combination of handwritten content and variable payer formats is precisely where standard OCR fails most consistently.
Anesthesia uses time-based billing: base units plus time units, where the time block on the claim determines the reimbursement amount. A misread time field is an extraction error that changes the payment calculation. Anesthesia groups seeing unexplained variance between submitted charges and received payments should audit time field extraction accuracy before investigating coding.
Payer-Side Adjudication and EOB Reconciliation
Payers processing high claim volumes face the same extraction accuracy problem in reverse. Accurate extraction of adjustment reason codes from paper EOBs (CO-45, PR-2, OA-23) enables automated reconciliation against ANSI X12 835 ERA files. EOB misreads make manual reconciliation the fallback. The medical insurance claim estimation work with Scaleport AI shows what this looks like on real payer-side workflows at scale.
Downstream fraud detection on claims data also depends on extraction accuracy at intake. Anomaly detection on billing patterns is only reliable when the extracted fields accurately reflect what was submitted.
The Denial Problem Starts at Document Intake, Not at the Coding Desk
Real-time adjudication requirements are expanding, and high-deductible plan growth is putting more scrutiny on each submitted claim. Both trends increase the volume of claims that need accurate extraction to process correctly on first submission. The revenue cycles that improve STP rates in this environment won't do it by adding denial management capacity. They'll do it by fixing what's happening at intake. That's where the problem actually lives.
The $262 billion in annual claim denials represents a ceiling that staffing investments have consistently failed to break through. The medical claims processing systems with the best financial performance will be the ones that replaced their document intelligence layer before the denial management queue became load-bearing infrastructure.
LlamaParse handles the full range of medical claims software document types (CMS-1500, UB-04, EOBs, prior authorization forms) with layout-aware extraction and field-level validation that standard OCR architectures can't match. Healthcare revenue cycle teams ready to move the STP needle rather than expand the denial team can explore our healthcare and pharma implementations as a starting point.