Insurance document fraud detection that tests every claim document for edits before it is paid
Every claim document over a set amount tested for edits before it is paid.
See one case, screen by screen ↓Testing the documents a claim is paid on — invoices, medical bills, repair estimates and receipts — for signs they were edited or made up, before the money goes out. Document Fraud Detection reads each document, inspects the file itself, redoes the arithmetic, checks the vendor or provider and searches past claim documents for the same template.
How to choose AI solutions for insurance →
Updated
Invoices, medical bills, repair estimates and receipts are easy to edit, and an edited one now looks clean to the eye. The edits that matter are small: one digit in a total, a re-typed hours cell, a moved date. They pass a busy desk adjuster and the quick arithmetic of whoever is in a hurry.
Checking a document properly means calling the vendor, redoing the arithmetic, opening the file properties and comparing it against past bills — 10 to 15 minutes a document, which a desk adjuster with 140 open claims does not have. So almost nothing is checked. And serial fraud cannot be seen claim by claim at all: the same invoice template arrives under several business names, on claims handled by different adjusters in different towns.
Estimated split in minutes for one invoice checked by hand, and with the solution — where the adjuster reads the pinned findings instead.
Six agents test the page, the file and its sources in under a minute; the adjuster decides and SIU accepts.
On Monday Oct 12 at 08:12, the insureds on a water loss upload a mitigation invoice through the claim portal. Here is what happens next, screen by screen, in the working solution.
Since 06:00 the agents have tested 186 claim documents, each before anyone could pay on it, in about 38 seconds each: 171 pass, 12 question, 3 refer, with $14,749.39 held. At 08:12 a new one lands — Clearwater Dry-Out & Restoration LLC, mitigation invoice, $9,812.60, on homeowners claim HO-26-31847.
Intake keeps the original and its fingerprint. The reader places 31 fields, 9 lines and 3 totals on the page. File forensics measures every glyph, the consistency checker redoes the sums and checks the register, and the look-alike finder compares the layout with 1.84 million claim documents. Each step shows what it found and how long it took.
“Refer — 3 refer-level, 3 question-level findings · SIU referral drafted.” Then: “Marisol Quintero decides.”
Marisol opens the invoice on the light-table. Six numbered pins sit on the page. She clicks pin 1, and the lens flies to the total in the Ink weight view: “$9,812.60” is set in Arial Bold at a 2.3 px stroke, while every other figure on the page — 31 of them — is Inter at 1.7 px.
“3 refer-level findings. The fixed rules say any one is enough to refer; the decision is yours.”
The nine lines add to $4,812.60, which is the printed subtotal. Tax is $0.00, so the total should be $4,812.60. It reads $9,812.60 — exactly $5,000.00 more. Pin 3 adds that the leading “9” sits 1.4 px above the baseline of the other digits.
The file was made by an online invoicing app on Sep 22 at 18:04, then opened and saved again in a desktop PDF editor on Oct 11 at 21:47 — 19 days later, and ten hours before the upload. The only object added was the total, on a second text layer. Nothing on the claim explains a change on Oct 11.
File forensics reports evidence with measurements — it never calls a document fraudulent.
Clearwater is not on the Larkspur vendor register and has no state contractor registration under its name. Its EIN 84-3301927 is on a Northfork Emergency Services invoice paid on claim HO-26-26390. Unit 4 is a mailbox store, the phone is a prepaid mobile activated in June, and remit-to PO Box 4471 is on three other invoices. The insured and the work dates do match the claim.
“Claims are linked on shared identifiers and template traits, never on a name alone.”
The same invoice template is on four unrelated claims under four other business names: Summit Peak Water Restoration, Rapid Dry Pros LLC, Brightline Restoration Group and Northfork Emergency Services LLC — handled by three other adjusters. $40,365.00 across them, $19,150.00 already paid.
Marisol slides the overlay between her invoice and the Northfork one, then switches to Differences: the only fields that differ are names, numbers and amounts. The trait grid marks what each invoice shares with hers: all five have the same layout grid, the footer “we keep your home dry!”, the misspelling “Dehumidifer” and a total box set 6 px low; some also share the PO box, the EIN or the heavier total.
The note writer has drafted the referral from the six pins — every sentence points at one. Marisol reads it, signs, and the payment on HO-26-31847 is held. She sends it to SIU with the original file, the lens views, the file history and the look-alikes attached. At 09:21 Owen Achterberg accepts it as SIU-26-0418, for all five claims.
“Your signature confirms the facts above; it does not accuse anyone.”
4,972 documents tested: 4,511 pass, 389 question, 72 refer — $412,860 held on refer and $199,620 asked about before paying. Edit history is the busiest test with 214 findings. Template clusters show the water mitigation invoice as new today, beside a tow and storage ticket already with SIU since Sep 24.
The same invoice, seen by the four people who carry it — what their day looked like, and what it looks like now.
Six agents test the page, the file and its sources in under a minute; the adjuster decides and SIU accepts.
Picks up every claim document over the threshold from the portal, email, vendor uploads and bill feeds, and sorts invoices, medical bills, estimates and receipts.
Reads every field, line and total with its position on the page, so each finding can be pinned where it is.
Inspects the file itself: fonts and stroke weight, text layers, edit history, digit alignment and re-encoded image regions.
Redoes the arithmetic and checks the document against its own sources: the claim, the regional price list, the vendor and provider registers.
Fingerprints the layout and searches every past claim document for the same template, tax ID, PO box or bank details on unrelated claims.
Applies the fixed verdict rules — pass, question or refer — and drafts the claim-file note or the SIU referral with every finding cited to its pin.
Decides every question and refer; SIU accepts referrals. The agents propose; a named person decides.
Adjusters, bill reviewers and SIU can ask about any document, finding or vendor — or tell it what to change. Every answer names the agent and the pin it rests on.
Why is the Clearwater invoice a refer?
Three findings are refer-level on invoice 20931. The total “$9,812.60” is Arial Bold at a 2.3 px stroke; every other figure is 1.7 px. The lines and subtotal give $4,812.60, so the total is $5,000.00 more. And the same layout is on 4 unrelated claims under 4 other names.
Show me the other claims with this template
Four, all water-mitigation invoices from the same template: HO-26-29104 Summit Peak ($11,460.00, paid), HO-26-30552 Rapid Dry Pros ($8,975.00, open), DP-26-07718 Brightline ($12,240.00, open), HO-26-26390 Northfork ($7,690.00, paid). $19,150.00 already paid.
Has Clearwater billed us before?
No. Clearwater is not on the vendor register and has no state contractor registration. But its tax ID 84-3301927 is on a Northfork Emergency Services invoice paid on HO-26-26390 on Jul 24, and its remit-to PO Box 4471 is on three other invoices.
Draft the SIU referral for all five claims
Drafted. It covers the lead claim and the four linked ones ($50,177.60 invoiced, $19,150.00 paid), cites each of the six pins, and says what we have and have not done. It needs your signature.
13 screens from the working solution, on its sample data. Pick one to see it large.
Which documents are tested, how hard each test bites and who decides are settings the fraud manager owns.
| Setting | Default | Choose from |
|---|---|---|
| Test any claim document over | $1,000 | $0 – $5,000, in $250 steps |
| Test medical bills over | $500 (raised from $250 on Oct 5) | $0 – $2,000, in $50 steps |
| Every document on a claim already flagged | On | On · Off |
| Ink weight, arithmetic, look-alikes | Refer | Refer · Question · each test on or off |
| Text layers, edit history, baseline, claim facts, register | Question | Refer · Question · each test on or off |
| Look-alike match | Score ≥ 0.95 on 2+ unrelated claims | Same template or identifiers |
| Second approval to pay against a refer | Fraud manager | Named role |
| Changes | Published as a new version | Apply to documents received after that |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample claim data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (Coalition Against Insurance Fraud, The Impact of Insurance Fraud on the U.S. Economy, 2022, via Triple-I — $45 billion a year in property-casualty fraud; Inscribe — 10–15 minutes to review a document by hand, vendor-reported). People, companies, claims and documents named on this page are fictional — characters and sample data in the working solution.
Testing the documents a claim is paid on — invoices, medical bills, repair estimates and receipts — for signs they were edited or made up, before the money goes out. Document Fraud Detection reads each document, inspects the file itself, redoes the arithmetic, checks the vendor or provider and searches past claim documents for the same template.
File forensics measures the fonts and stroke weight of every glyph, lists text layers and when each was added, compares the edit history with the claim timeline, and checks that digits share a baseline. On the Clearwater invoice it found the total alone in a heavier font, on a layer added the night before the upload.
Yes. The look-alike finder fingerprints the layout and searches every past claim document for the same template, tax ID, PO box or bank details on unrelated claims. It never links claims on a name alone, and shows each shared trait — layout, footer wording, a misspelling, a total box set low.
Any claim document over a set amount — $1,000 in the working solution — every medical bill over $500, and every document on a claim already flagged. The fraud manager sets those thresholds.
No. The agents test and explain; they never hold a payment, refer, contact the insured or vendor, or call a document fraudulent. The adjuster decides every question and refer, SIU accepts or declines referrals, and paying against a refer needs a second approval.
A referral drafted from the pinned findings, where every sentence points at a pin, signed by the adjuster. It attaches the original file with its fingerprint, the lens views, the file history and fonts, the look-alikes and the vendor check — and can cover several linked claims as one case.
Yes. Besides invoices it tests medical bills, repair estimates, tow and storage tickets, auto glass invoices and purchase receipts — checking, for example, billed hours against the approved estimate, or a VIN against the policy vehicle.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your registers, price lists and a sample of past claim documents — and it goes live once every quality gate has passed. We will run it on your own documents first.
Referrals in days, not weeks, after first notice, each with its evidence.
1-min demoFraud Network AnalysisOrganized rings seen across claims, not missed one claim at a time.
1-min demoContractor Fraud DetectionContractor clusters raised before payment, not found after it.
1-min demoWe’ll run Document Fraud Detection on a sample of your own paid and open claim documents.