Signzy

First party models do substantial work: optical character recognition on documents, biometric face matching, liveness and deepfake detection, forgery identification, and Trust Scores computed from more than 200 device, transaction and identity signals to predict mule account behaviour. Against that, the removal test returns a large product.

A no code journey builder, an API marketplace of more than 340 endpoints and the orchestration layer routing a customer through configured checks would all remain and would still be useful, because much of what a bank buys is the ability to assemble a compliant onboarding flow without engineering. This is the Alloy and Duna position, where models sit on top of routing and orchestration rather than constituting the product.

The product is built to remove the human from onboarding and the published framing is explicit, with verification that took days now completing in under a minute and fraud stopped at the door rather than discovered later. Risk based journeys give an institution real configuration control, letting it route by risk level, geography and product type, which is a genuine oversight lever at design time. What is absent is anything at decision time.

No review queue, escalation threshold, manual adjudication surface or alternative evidence path is described for a person the system cannot verify, and that omission is the specific contrast with Persona, which holds an A on this axis solely because a failing applicant is routed into a step up path instead of being rejected.

The published number measures the wrong thing. A 99 percent success rate across ten million monthly onboardings is a completion statistic, telling a buyer how often a journey finishes rather than how often the verification was correct, and the two diverge exactly where it matters. No false match rate, false rejection rate, presentation attack detection accuracy or demographic breakdown was located, and no accredited third party evaluation appears.

Training scale is offered in place of accuracy, with models described as trained on millions of real verification attempts, which is the same substitution recorded for CLARA Analytics where corpus size stands in for measurement.

More than 540 financial institutions worldwide, including every one of India's largest banks, a top three United States acquiring bank, more than ten of the Fortune 30 and named global fintechs, with reach stated across 180 countries. Volume is given in an operational unit rather than a marketing one, at more than ten million customer and business onboardings a month at a 99 percent completion rate, alongside a stated reduction in time to market from six months to three or four weeks.

Independent corroboration is unusually strong for an Indian vendor: recognition by a major analyst firm among notable innovators in know your customer automation as the only Indian platform listed, awards from the country's central bank, multiple granted patents in two jurisdictions, and a 4.7 average rating across public enterprise reviews.

The training arrangement is stated plainly and never governed. The company says its models have been trained on millions of real verification attempts from actual banking customers, covering every document type, lighting condition and edge case encountered when real people open real accounts, which is an explicit description of customer data improving models used by every other customer. That is the same disclosure Codat makes in passing.

Nothing addresses consent, whether the institution or the verified individual was told, whether biometric captures form part of the training corpus, or whether a bank can opt its customers out while still using the product.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located, and the payload is at the sensitive end even for this lane. Biometric capture through face matching and liveness checks runs at more than ten million onboardings a month, alongside identity documents, device fingerprints and behavioural signals, and biometric templates are the category of personal data that cannot be reissued once compromised.

Nothing published states how long a facial template is retained after a verification completes, whether it is stored at all, or how the home market's data protection statute and the various biometric regimes across 180 countries are each satisfied.

No attestation, certification, trust centre or enumerated framework was located, and no accredited presentation attack detection certification appears, which is the specific assurance a buyer compares in this lane and which two peers publish.

For a platform processing biometric captures at ten million onboardings a month for more than 540 supervised institutions, the absence of any published assurance set is the most conspicuous gap in an otherwise strong profile, and it is the first document a bank security review requests.

The regulatory position is grounded in named national infrastructure rather than asserted, which is the pattern now recorded across every Indian vendor in this index. The platform automates search, validation, upload and monitoring against the central know your customer registry, which is supervised national infrastructure a vendor must be admitted to.

Video based verification, a specifically authorised modality introduced by the central bank, is a core product, and the company's founding thesis dates from the central bank opening digital onboarding to banks. It has received awards from that central bank directly, holds granted patents in two jurisdictions, and publishes a reference library of the laws and compliance frameworks its buyers operate under. Sanctions, politically exposed person and watchlist screening are named functions. Fifth A on this axis after OnFinance AI, Akur8, NICE Actimize and Perfios.

Two exposures stack and neither is addressed. Biometric face matching and liveness detection carry the best documented demographic error differentials in applied machine learning, varying by skin tone, age and gender, and the consequence of a false rejection is a person denied a bank account.

No participation in an accredited presentation attack detection scheme or a government face recognition evaluation was located, which is precisely where Sumsub and Incode earn their standing, and Incode holds the only B on this axis in the lane specifically for submitting to adversarial demographic testing.

The second exposure is newer: Trust Scores derived from device and behavioural signals import the accessibility problem recorded for BioCatch, since atypical interaction reads as risk, and a mule prediction is an accusation attached to a named person. No demographic error rates, per market accuracy or appeal route was located.

No guarantee, indemnity or falsifiable accuracy commitment was located. The institution retains real control through risk based journey configuration, so a bank can widen or narrow its own thresholds, which is challenge capability for the buyer. The party with no route is the applicant.

A person whose face fails to match, whose liveness check fails, or who is assigned a mule risk Trust Score is not told which system reached that conclusion, cannot see the score or the signals behind it, and has no described path to an alternative evidence route or a human review. The consequence is exclusion from a bank account, and it falls hardest on the same groups the underlying techniques serve least well.

The commercial chain is named at its most significant points, with a global card network, a major enterprise software vendor and a hyperscale cloud and software provider all identified as simultaneous investors, distribution partners and customers, which is an unusual triple relationship and is disclosed rather than obscured.

The architecture is also honest about its dependencies, since an API marketplace of more than 340 endpoints is by definition an aggregation of external verification and data services. What is not published is the enumeration: individual data providers behind those endpoints are not listed the way Alloy lists its 270 partners, and no model provider or subprocessor is named for the biometric and generative components.

Integration is the company's strongest published property and it is evidenced from two directions. Technically, more than 340 application programming interfaces and low code widgets drop into an existing workflow, a no code builder assembles journeys from modular components without engineering, and institutions are stated to go live within days against a six month norm.

Commercially, the distribution reach is something almost no vendor here can match: a card network rolled the company's video verification engine out to banks worldwide across its own global network, and two major enterprise software providers act as partners alongside it. Being installed through the rails a bank already uses rather than sold into them is a structurally different integration position.

No hosting provider, region selection, residency commitment or private deployment option was located. Reach across 180 countries with offices on three continents means biometric captures and identity documents cross many transfer regimes, and the home market imposes its own localisation requirements on financial data, so the question is unusually consequential here. Nothing published tells an institution where a facial template or an identity document processed through the platform comes to rest, or whether the location can be constrained by contract.

The charging model is disclosed even though the rate is not: subscription access to the platform combined with per transaction fees on application programming interface usage. That is the structure a buyer needs to model cost against onboarding volume, and it warns plainly that expense scales with success, which is the same useful disclosure that earns Glia and UPTIQ their grade.

Integration effort is also stated concretely, with most institutions going live within days against a stated six month norm for traditional implementations. The disclosure reaches the reader through independent review rather than the company's own pricing page, and no rate, tier or minimum appears anywhere.

Coverage is wide on all three dimensions the axis measures. Institution types run from banks and non bank lenders to payment providers, acquirers, global fintechs and large non financial enterprises. Geography spans India, the United States, the United Kingdom and the Middle East with offices on three continents and stated reach across 180 countries, which matters in this category because document types and verification sources differ by jurisdiction.

Journeys covered include account opening, loan disbursement, credit card issuance and co lending, with risk based configuration segmenting customers by risk level, geography and product type, so one platform serves materially different onboarding obligations.