Intelligent document processing for insurance and claims
Turn claims correspondence, payment documents and forms into validated data inside the system you already use. Every value comes with a confidence score and the page it came from, and anything uncertain goes to your team.
- Built into a live claims platform
- Piloted on 212 real documents (820 pages)
- People review every exception
- ClaimantJ. Sample0.99Page 1
- Claim numberWC-0001230.98Page 1
- NDC00000-1111-220.97Page 2
- Date of service03/?4/20260.71Page 2
Your team shouldn't be retyping documents
Claims operations still run on paper and PDFs: correspondence from lockboxes, scanned payment documents, multi-page letters covering several claims at once. Somebody has to read each one and key the data in.
- Backlogsevery time volume spikes
- Typing errorsin claim numbers, drug codes and amounts
- Scrambled tablesfrom generic OCR tools that lose multi-column layouts
- Answers you can't checkfrom AI tools that auditors won't accept
Why we don't quote a single accuracy number
Accuracy depends on your documents: scan quality, layouts, handwriting. A vendor promising 99% before seeing them is guessing. We measure accuracy on your documents first.
- What we found
- Plain OCR text lost the layout of tables.
- What production does
- Sends the page images to the model as well.
- What we found
- Handwriting and poor scans are unreliable.
- What production does
- Gives them lower confidence scores and sends them to review instead of guessing.
The data your claims team actually needs
New fields are added by configuration, not by rebuilding.
- Claims correspondence
- Claimant name
- Claim reference numbers
- Prescription numbers and dates
- Drug names
- National Drug Codes (NDCs)
- Dates of service
- Amounts
- Response codes
- Payment documents (EOBs)
- Payer
- Payment details
- Claim lines
- Paid and denied amounts
- Other forms
- The fields you choose
From scanned page to checked data, in 7 steps
Shown on a synthetic sample document. Real documents never leave your environment.
- IngestBatches arrive from lockboxes, uploads or email: scanned or digital, one page or many.
- ReadOCR turns each page into text, and we keep an image of every page.
- SplitPages are grouped by claimant, including pages shared between claimants.
- ExtractThe AI model reads the text and the page image together, so tables and layout aren't lost.
- CheckRequired fields, formats and business rules are validated. Every value gets a confidence score and a page citation.
- RouteConfident records go straight into your system; the rest go to your team's review queue. You set the threshold.
- MeasureEvery run is logged with its cost, and accuracy is checked against verified samples so it keeps improving.
AI where documents are messy, structured data where it exists
When one payer started sending electronic payment files (835s), we processed those directly and switched AI off for that payer. It's cheaper and exact. We'll always tell you when you don't need AI.
- Scanned correspondenceAI extraction
- Payer sending 835 filesDirect processing, AI off
- Handwritten pagesLow confidence, to review
Built for audits, not demos
Every value can be explained, every decision can be traced, and every run has a cost.
Traceable: a confidence score and source page for every value
Controlled: you set the confidence threshold; people review everything below it
Auditable: a full log of every document, decision and change, with versioned AI instructions
Protected: safeguards against text in a document trying to steer the AI
Costed: AI cost recorded per document from day one
Private: processing runs in your cloud environment, with a BAA where health information is involved
No new app for your team to learn
We build the capability into the platform your team already uses. Results land in your database, your existing users and permissions apply, and allocation waits until a whole batch is processed, so nothing is decided on half the data.
- Your database
- Your users and permissions
- Allocation after the full batch
Live in a workers' compensation claims platform
For a workers' compensation pharmacy claims administrator, correspondence from multiple lockboxes now becomes claimant, prescription, drug-code and response-code data inside their claims platform.
Read the claims platform case study- Records above the admin-set confidence threshold are allocated to claims automatically
- Records below it go to a review screen where staff check, correct or complete them
- Letters covering several claims are allocated line by line
- The approach was proven on 212 real documents before the full build
Intelligent document processing FAQ
What is intelligent document processing?
IDP combines OCR, AI models and business rules to read documents, pull out the data you need and check it before it enters your systems. Unlike basic OCR, it understands layout and context, and it flags what it is not sure about.
How accurate is it?
It depends on your documents, so we measure it on them in a pilot before going live. The pilot has a fixed scope and fixed fee, credited toward the full build. Confidence scores and human review mean uncertain values are never passed through silently.
Do we need to replace our claims system?
No. We build IDP into your existing platform through its database or APIs.
How do you handle health information?
Processing runs in your own cloud environment, under a BAA where required. We never accept health information through our website, and our demos use de-identified documents.
What does it cost to run?
AI usage is recorded per document from the first run, so you can compare it directly with the cost of manual processing.
Can it handle handwriting?
Handwriting and poor scans get lower confidence scores and go to review, so nothing is guessed.
See it on documents like yours
In a 30-minute demo with our CEO or CTO, we'll walk through the process on de-identified sample documents and talk about what a pilot on your documents would look like.
- Extraction on de-identified samples, step by step
- Confidence scores, page citations and the review queue
- What a pilot on your own documents would involve