For insurers · Deutsche Fassung · Calculator (German)
Claims fraud: the leak below the waterline.
Fraud is counted where it is caught. What is not caught is paid, and appears in no statistic. This audit shows which part of it a photo sealed at live capture and signed agent records shut out, which part they do not, and the threshold we ask your pilot to measure us against.
As of 5 November 2026. Every figure on this page has a source or is marked as an assumption.
1 · The problem is large, and the largest part is invisible
> €6 bn
Damage from insurance fraud per year in Germany, as estimated by the German insurers' association GDV, up from an earlier four to five billion.
Source: GDV press release, 2 May 2024
one in ten
property and casualty claims is suspicious enough to investigate, unchanged over the long term. Half of the dubious claims are motor.
Source: GDV special analysis of over 600,000 claims from three years, GDV press release, 2 May 2024
300 %
Rise at Allianz UK in claims backed by manipulated images, videos and documents, from 2021/22 to 2022/23.
Source: The Guardian, 2 May 2024, reporting Allianz UK figures
What other sources say
Europe
Detected and undetected fraud together are estimated by Insurance Europe at around 10 % of all claims payments in Europe. Note: "suspicious" (GDV) and "fraud" (Insurance Europe) are not the same thing.
Source: Insurance Europe, Annual Report 2023-2024, Fraud chapter, 11 June 2024
The GDV itself, October 2026
Generative AI makes it easier to fake images and invoices. The association names image and document forensics as a countermeasure. The rise to over six billion comes from higher cost per claim, not from more suspicious claims.
Source: GDV, insurance fraud dossier
United Kingdom
UK insurers detected £1.34 bn of fraud in 93,900 claims in 2025, £14,300 per case on average. That is only the part that was found.
Source: ABI, reported by Insurance Business, 6 October 2026
The checkers are unsure
In a Verisk survey, 99 % of the insurers asked had seen manipulated or AI-altered documents; only 32 % felt confident they could reliably spot deepfakes. 36 % of the US consumers asked would edit claim images or documents.
Source: Verisk survey (1,000 consumers, 300 claims professionals), reported by Claims Journal, 17 March 2026. Self-reported, not measured on claims.
What nobody has measured, ourselves included: how often real claims use edited, borrowed or old photos, and how much live in-app capture changes that. We found no study with a measured reduction, including from vendors of similar methods. That number is what the pilot produces.
Why it is a blind spot
Every fraud figure is an estimate. That is in the nature of it: fraud paid without suspicion leaves no trace. You do not notice a leak below the waterline by the water; you notice it because the ship sits lower. Three zones:
Stopped
Fraud that is detected and not paid. The only part the statistics count.
Suspicious
The roughly one claim in ten that has to be investigated. Most customers are honest (the GDV says so too): here you pay investigation time and adjusters for claims that turn out clean.
Paid without suspicion
The dark field. A well edited photo, someone else's picture, an old damage: paid, booked, forgotten. There is no figure here, only an estimate. This is the bottomless pit.
Today a claim photo is judged by eye, or by a detector that estimates a probability. Neither tells you what you missed. A seal at capture turns the question around: your checker no longer has to find the fake; the file has to prove that it was captured live and has not changed since.
2 · Audit: fraud patterns against Fortress building blocks
The first three patterns are the warning signs the GDV itself names. The "effect" column says what the building block does technically, not how much fraud disappears from your book.
| Fraud pattern | Building block | Effect | Limit |
|---|---|---|---|
| Manipulated purchase receipts (GDV) | RECEIPT | partly: the receipt photo is sealed at capture; edits and reuse of the same receipt show | A paper receipt that was forged from the start is sealed correctly. |
| Story does not match the damage (GDV) | CHAIN · Car damage claim | partly: photos, AI steps and decision in one checkable file; six guided views, all different, one phone, one visit | Matching story and picture remains your adjuster's work. |
| Contradictory statements (GDV) | CHAIN | partly: every statement stays unchanged in the file; later changes show | No record says whether a statement is true. |
| Claim photo edited (damage enlarged, retouched) | ProofCam · | shut out: the check fails and marks the changed area | Any re-compression counts as a change; the original must arrive. |
| AI-generated claim photo | ProofCam, live capture only | shut out as a file route: an uploaded image never gets a seal | A photo of a screen gets a valid seal; we do not detect that reliably today. |
| Someone else's photo (internet, another claim, another car) | Car damage claim · | shut out as a file route: only photos taken in the app at reporting carry a seal | As above: photographing a screen. |
| Old photo of an earlier damage | ProofCam | shut out as a file route: a gallery photo cannot be sealed | An old damage photographed again today passes. The seal dates the photo, not the damage. |
| Amount, payee or approval changed afterwards in your own process (insider, AI agent) | POAW · REPLAY · | shut out: the check fails and names the step | Shows what was recorded, not whether the decision was right. |
| Fraud rings learn your checking rules through a chatbot or portal | VAULT | partly: fixed rules in front of the model, an optional leak judge | No detection rate published. |
| Staged accident, deliberately caused damage, staged scene | none | not addressable | No seal says what was in front of the lens. |
The addressable part is fraud that works with a file rather than with a damage: edited, generated, borrowed and old photos, plus edited or reused receipt photos and changes inside your own process. How large that part is in your book, nobody knows exactly, you included. That is the blind spot.
3 · The open model: why at least half, and of what
We do not have a measurement at an insurer yet. So there is no promise here, but a model whose assumptions you can see and replace, and a threshold your pilot measures us against.
- The route changes, not the hit rate. Today your checker has to find a fake. With live capture required, the fraudster has to get past a seal that breaks on any pixel change. A file without a seal does not go into the fast lane but into normal review. For the file routes this is a yes/no check, not a probability.
- Three routes remain, and we name them. (a) photograph a screen, (b) cause or stage the damage for real, (c) refuse the app and go through normal review. Route (b) means real damage and real criminal risk. Route (c) ends up with your checker, as today. Route (a) is the open gap; your SIU team attacks it on purpose in the pilot.
- Why the threshold is half. The model holds as long as fewer than one in two file-based frauds successfully switches to a remaining route. We consider that cautious, because every remaining route costs more effort or more risk than uploading an edited file. This is an assumption, not a measurement.
- What half does not mean. Half of the addressable part is not half of all fraud. Share of all fraud = addressable share × share of sealed claims × reduction. An example with assumptions (addressable 30 %, sealed 60 %, reduction 50 %) gives 9 % of all fraud. Applied to the GDV estimate of €6 bn, that would be around €540 m a year, if those assumptions hold.
How the pilot measures the dark field
The largest benefit may not be the saving, but that the dark field gets a number for the first time.
- Two groups: some claims in one line run with the live seal, the others as today. The allocation is fixed in advance.
- What is counted: photos whose check fails; claims withdrawn after a live photo is requested; SIU hits per hundred claims; average payment per claim; on-site adjuster visits.
- The threshold, in writing before the start: file-based fraud in the sealed group falls by at least half against the comparison group. If the pilot misses it, we say so on this page.
- The limit: claim frequency and severity fluctuate. How many claims a robust result needs, we work out with your actuaries before the start.
4 · Why loss adjuster costs can fall
An adjuster drives out because a photo cannot be trusted. A photo that proves it is live, complete and unchanged removes that reason in part of the cases.
Photo assessment instead of a visit
Six guided, sealed views from one visit with one phone: the adjuster assesses remotely where today an appointment with travel is needed. Which claim sizes and lines qualify is your decision.
Fewer re-inspections and counter-reports
The argument "is the photo genuine, is it old?" disappears when the seal is intact. What remains is the argument about the amount, and that is shorter.
Faster review when in doubt
The heatmap marks the changed area in blocks of 16 × 16 pixels. It does not replace forensics, but it tells the adjuster where to look.
Recovering costs
Where fraud is proven, insurers can reclaim adjuster costs. A failed seal check and a signed process are evidence that a court can follow without our software.
Source on recovery: GDV, reported by Cash, 2 May 2024
That photos can replace the visit has been tried: in 2018 Allianz and VW Leasing reported that 40 % of claims were approved by photo and 16 % handled by livestream instead of an on-site adjuster, with a result after two hours instead of one to two days; 2 % still needed a visit. These figures are old and concern claims without tamper evidence. A seal is meant to make the photo route hold where today distrust leads to an appointment.
Source: AUTOHAUS Schadenforum, 12 November 2018. For scale: the average car property damage in German motor liability was €4,250 in 2024 (GDV, 1 October 2025). We found no reliable average for the cost of a motor damage report; enter your own in the calculator.
5 · Where costs fall: eight levers
| Lever | Where it lands | Building block | How measured |
|---|---|---|---|
| Fewer payments on file-based fraud | Claims incurred | ProofCam, car damage claim, RECEIPT | Pilot, comparison group |
| Less review effort for honest customers in the suspicious pile | Claims handling | Seal as exoneration | Hours per suspicious claim |
| Fewer on-site adjuster visits | Claims handling | Car damage claim, ProofCam | On-site visits per hundred claims |
| Shorter disputes | Claims handling, legal costs | POAW, CHAIN | Calculator part A |
| Fast lane for honest customers | Cycle time, retention | Claim photo check | Days to payment |
| Recovery and subrogation with evidence | Income | POAW, seal | Amounts recovered |
| Manipulation inside your own process | Operational risk | POAW, REPLAY | Failed checks |
| The dark field gets a number | Pricing, reserves | Pilot design | Difference between the groups |
What we do not claim
- That Fortress halves claims fraud overall. The pilot threshold applies to file-based fraud in the sealed group.
- That a seal shows whether a damage is real. It shows that the photo was captured live and has not changed since.
- That we reliably detect a photographed screen. That is open.
- A measured effect in an insurer's book. That exists only after your pilot.
Next step
One line, one region or one channel, two groups, the threshold in writing before the start. Pilot · Measurement plan