Insurers and TPAs receive claims as whatever the claimant could photograph or fax. DocxIntel splits those packets, reads the handwritten Arabic forms and stamped garage invoices inside them, reconciles the claim against the policy schedule and flags the invoices that do not add up — all on your own hardware.
Accuracy is measured field-by-field against a human-adjudicated ground truth, not character-by-character. Your own threshold is set on your own claim documents during the Proof of Value, and it is the number the engagement is judged on.
Claims do not arrive as clean forms. They arrive as a WhatsApp photograph of a police report, a stamped garage estimate, a discharge summary and an IBAN letter, all in one email — and someone has to sort it before anyone can assess it.
Every page in an incoming packet is classified, then the packet is split into its component documents with the right pages attached to each. Handlers open a structured claim file rather than a 40-page scan.
FNOL forms, police reports, vehicle registration cards, driving licences, garage estimates and final repair invoices are read together, including stamps and signatures over printed totals.
Medical reports, discharge summaries, lab results, pharmacy invoices and reimbursement forms are extracted with diagnosis, procedure and amount fields intact, including handwritten physician entries.
Extracted claim values are checked against the policy schedule: cover in force on the incident date, sub-limits, deductibles, exclusions and the named insured. Mismatches are raised as specific findings.
Duplicate and near-duplicate invoices, amounts that disagree between documents, recycled repair photographs and totals whose ink does not match the surrounding print are flagged for investigation.
Reimbursement forms, receipts and bank details are captured for payment, and subrogation files are assembled with the recovery evidence already indexed and cited.
This is the workflow insurers ask us to prove first, because it is where the handler time actually goes. Nothing in it involves a page leaving your data centre.
Every value in that claim file can be clicked back to the pixels it came from, which is what makes a settlement decision defensible months later.
One line of business first, measured on your own claim files, before anything touches the rest of the book.
Motor or medical, whichever generates the most manual handling. We ask for the packets your team dreads — the photographed forms, the third-generation faxes, the handwritten Arabic FNOLs — because those decide whether the number holds in production.
Which documents you expect in a packet, which fields each one must yield, and which of those fields are payment-critical. Payment-critical fields get tighter confidence thresholds than descriptive ones.
Containers and open-weight models are deployed on your hardware — air-gapped, private cloud or managed single-tenant. Your infrastructure team holds root and the encryption keys; no claim data is visible to us.
Policy-schedule checks, sub-limit logic, estimate-versus-invoice tolerance and duplicate-invoice matching are configured against how your handlers already work, then tuned on historical claims where the outcome is known.
Your reviewers adjudicate field-by-field and the report shows per-field accuracy, confidence distribution and the exception rate you would actually staff for. Miss the agreed threshold and you keep the report and owe nothing further.
Adding health to a motor deployment, or reprocessing three years of settled claims for a subrogation review, is a configuration change and some GPU time. The licence does not move because the volume did.
An indicative schema. The final field list is agreed against your own forms during the Proof of Value, not taken from a template library.
Handwritten fields, stamped fields and fields recovered from photographed pages carry the same provenance metadata as clean printed text — a handler can always see what the model actually read.
Cloud IDP platforms handle insurance forms competently. The constraint is that every claim page has to reach them first, and claim pages are exactly the pages that contain health data and identity documents together.
Where claim packets are read
Health data in a claim leaves your control
Cost of a claims surge
Reprocessing settled claims for subrogation
Illustrative cost at 100,000 pages a month
Parsed claim data cached by the vendor
Handwritten Arabic claim forms
Self-hosting available on
Cost figures are our own arithmetic on publicly published list rates as of 2026, not vendor quotations. Caching behaviour and self-hosting availability reflect publicly documented behaviour of metered parsing services. Verify against current vendor documentation before making a decision — we would rather you checked.
The objections that decide whether claims automation gets approved, answered without hedging.
That is the normal case, not the exception. DocxIntel classifies page by page and splits the packet into its component documents — FNOL form, policy schedule, police report, garage estimate, invoices, medical report — then extracts the fields that belong to each. The claim handler receives a structured claim file instead of a single scan to scroll through.
Thirty to forty-five days, fixed fee, deployed in your environment on your own claim files. The accuracy threshold and the conversion price are both agreed in writing before we begin.
