Resistant AI: Document Fraud Detection for Finance

On This Page #

A Prague-founded forensics vendor that scores whether a submitted document is authentic, sitting alongside an extraction pipeline rather than replacing it.

Overview #

Nearly every entry in this IDP vendor directory is judged on how accurately it reads a document. Resistant AI answers a different question: whether the document deserves to be believed. Its Document Forensics product examines a file's structure, metadata, fonts and rendering rather than the meaning of the text inside it, and returns a trust verdict that a decision engine can act on. The two questions are independent: a forged bank statement generated cleanly in a PDF library extracts perfectly, because there is nothing wrong with its characters.

The operating company is Resistant AI s.r.o., registered in Prague on 23 January 2019 with share capital of CZK 364,000 and three managing directors: Martin Rehák, Martin Grill and Karel Bartoš. Its sole shareholder is Resistant AI, Inc., a Delaware corporation, so the Czech entity is the subsidiary of a US holding company. The about page names Prague, London and New York as offices and counts nine PhD founders, a wider group than the three officers on the register. Buyers who care about data residency or contracting jurisdiction should establish whether they are contracting with the Czech s.r.o. or its Delaware parent.

Funding is substantial for a company in this niche. Four disclosed rounds total roughly $55 million: a $2.75 million seed in 2020 led by Index Ventures and Credo Ventures, a $16.6 million Series A led by GV in October 2021, an $11 million extension led by Notion Capital in June 2023, and a $25 million Series B led by DTCP Growth in October 2025. Experian took part in the Series B, a credit bureau buying exposure to the question of whether the documents behind a credit decision are real.

How Resistant AI checks documents #

The product runs four stages the vendor names Quality, Class, Trust and Decide. The product page puts quality assessment under three seconds and classification under four, then gives the trust verdict and the policy decision under 20 seconds each, while advertising under 20 seconds per document as the headline figure. Those two statements do not fit together, and the page does not say which one is the end-to-end number. Accepted formats are PDF, JPEG, PNG and TIFF, and the vendor states the checks are language-independent because most of them never read the text.

The detection surface is described as more than 500 checks per document. Publicly, these fall into three families: file-level forensics on metadata, producer strings and internal PDF structure; visual forensics on fonts, rendering artifacts and pixel-level inconsistency; and template comparison against how a given issuer's documents are supposed to look. None of these depend on the extracted values being wrong, which is why the layer composes with rather than duplicates data extraction.

A further claim concerns cross-document analysis rather than single-file inspection. The vendor's 2026 fraud report states that 23% of high-risk documents showed signs of serial fraud, meaning the same underlying template or generator reappearing across submissions, and that 98.3% of that serial fraud carried no prior attribution to a known template farm. If that holds, the detection value comes from correlating traffic across many customers rather than from matching against a blocklist, and a buyer's practical question becomes how much volume the vendor sees in their specific document type and geography.

Integration is designed to sit inside existing pipelines. Google published a 2023 engineering post on running Document Forensics over Google Cloud Document AI, pairing Google extraction with a separate authenticity verdict, and Tungsten Automation documents a connector for TotalAgility workflows. The vendor's own partner page adds ABBYY on the document side and ComplyAdvantage and Lucinity on the transaction side, with Experian appearing on both that list and the Series B cap table. Camunda lists it as a technology alliance partner for orchestrated workflows, and in Germany the Düsseldorf integrator TeDeG lists it as a partner alongside Hyperscience and Infrrd. Every one of those routes reaches the buyer through someone else's document flow.

Use cases #

Lending and income verification

The core case is a lender reading bank statements, payslips and tax documents to make a credit decision. The vendor's published customer list names the mortgage broker Habito and the payments firm Payoneer. This is the workflow where automated document quality and verification checks have the most direct financial consequence, because the extracted income figure drives the underwriting model whether or not the statement is genuine.

Onboarding, KYB and account opening

Company registration certificates, proof of address and utility bills submitted during onboarding are checked before an account exists to monitor. The vendor's customer list names Finom, Holvi and Close Brothers, and its transaction monitoring product extends the same logic past onboarding into payment behavior.

Insurance claims and expenses

Claim documents, invoices and receipts are checked for manipulation before payout. Lemonade appears on the published customer list. Altered invoices and duplicated receipts are the classic cases here, and both are invisible to an extraction engine that reports high confidence on the amount it read.

What is not independently verified #

The headline accuracy figure is unaudited. The product page states 99.2% accuracy without a test set, a document mix, a definition of what counts as correct, or a corresponding false-positive rate. For a detection product the false-positive rate is the number that determines operational cost, because every wrongly flagged genuine document becomes manual review work, and it does not appear anywhere on the vendor's site. That absence should be the first item on a procurement checklist, and any pilot should measure it on the buyer's own traffic. The one assurance on the site carrying an outside signature is the SOC 2 Type 2 attestation issued by Sensiba San Filippo, which covers the security trust service criteria and says nothing about how well the detection works.

The fraud statistics come from the vendor's own customer traffic rather than a representative sample of documents in circulation. The 2026 report is more careful than most vendor research, disclosing methodology limitations around detection maturity and sample concentration, but the population it measures is the flow arriving at companies that already bought fraud detection. That population is not the general population of documents, and the direction of the bias is unknown.

The document counter does not reconcile across the vendor's own pages: the 2026 report cites more than 170 million documents processed since 2019, the about page says 200 million, and the Series B announcement carries both "180,000,000+ docs verified" and "over 150 million documents verified" in the same post. These are most likely snapshots taken at different dates and left un-updated, but the figure should be treated as approximate.

Independent user reviews are effectively absent. The Capterra listing has zero reviews as of 27 July 2026, and no Gartner Peer Insights entry was found. The recognition on display is a count rather than a test: the about page claims "19+ Awards" in financial crime, AI and startup categories and shows a Chartis FCC50 2026 Market Disruptor badge, none of which involves a measured comparison against another product. The patents line on that page reads "120+ AI & ML patents based on the founders work", which credits prior careers rather than claiming a company portfolio, and no patent numbers are given. No published third-party benchmark of financial-document fraud detection turned up for any vendor in this category, so the gap looks like a category problem rather than a mark against this vendor specifically.

Technical specifications #

Resources #

Resistant AI websiteDocument Forensics product pageGlobal Document Fraud Report 2026Google Cloud engineering post on Document ForensicsCzech commercial register entry

Company information #

Resistant AI s.r.o.

Lazarská 13/8, 120 00 Praha 2, Czech Republic

IČO 07825439, share capital CZK 364,000

Managing directors: Martin Rehák, Martin Grill, Karel Bartoš

Parent: Resistant AI, Inc. (Delaware, United States) resistant.ai