Arva AI

The removal test leaves the manual process the product exists to eliminate. Agents conduct advanced web due diligence to establish what a business genuinely does, read and extract from formation, ownership and banking documents, detect document fraud, screen for adverse media, enrich entity and individual records from online sources including professional networks, adjudicate screening alerts and handle transaction monitoring queues.

The company describes its end goal as a suite of artificial intelligence workers doing compliance work, which is a product definition rather than a feature list, and nothing underneath it survives the models being taken out.

The most aggressive autonomy position recorded in this index, and it is stated rather than implied: the agents completely replace human analysts, and the company describes replacing human compliance analysts as the product. Every other vendor in this lane hedges toward augmentation, and this one does not.

There is a real boundary inside the claim, since replacement is scoped to low and medium risk cases with higher risk presumably retained by people, and that scoping is the difference between an aggressive product and a reckless one.

What is missing is everything that would make it assessable: no threshold defining low, medium or high risk is published, no escalation rule is described, no confidence exposure accompanies an agent decision, and nothing states who reviews the agents themselves in a function where independent testing is a supervisory expectation.

The published figures measure automation rather than correctness, and the distinction matters more here than almost anywhere in this index. Discounting 91 percent of screening alerts tells a buyer how many alerts the system closed, not how many of those closures were right, and a false negative in sanctions screening is a missed designated party rather than an inefficiency. The same applies to 92 percent of reviews automated. Finding risks humans miss is asserted without a comparison.

For a product replacing the analysts who currently catch those cases, the precision and recall of the agents against human adjudication is the number a model risk function and an examiner would both ask for first, and none is published.

No customer is named anywhere, which leaves a set of striking claims resting entirely on the vendor's own account: 92 percent of financial crime reviews automated, 91 percent of screening alerts discounted, operational spend cut by 80 percent, onboarding a hundred times faster.

What does carry weight is the founding credential, since the chief executive previously led the financial crime product team at a large European neobank's business arm, which is the buyer's seat rather than the seller's. A three million dollar seed was led by a major technology company's early stage artificial intelligence fund, joined by a well known accelerator and several fintech angels.

The problem framing is also specific and checkable, that a human analyst is required on more than 40 percent of business onboardings and that onboarding delay causes up to half of applicants to drop out.

No data boundary statement was located. The agents are described as analysing business websites on an ongoing basis, which means an accumulating picture of entities across every institution using the platform, and financial crime typologies learned at one bank are directly valuable to another.

Nothing states whether findings, risk assessments or fraud patterns are contained to the institution that generated them, whether a customer can decline to contribute, or how one bank's onboarding queue is separated from another's.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The payload extends well past corporate records: agents enrich individual as well as business details from professional networking sites and social media, screen adverse media on named people behind a business, and process formation, ownership and banking documents.

That means personal information about directors and beneficial owners is gathered from public online sources without those individuals initiating anything, and nothing published describes what is retained, for how long, or what happens to the profile of a business that is refused.

No attestation, certification, trust centre or enumerated framework was located. For a young company that is expected, and it is also the gate this particular product must pass, because a bank contemplating handing its onboarding and alert adjudication to an external agent will run a third party risk assessment before any pilot, and the absence of a published control set is the first obstacle a founder led sales motion encounters at enterprise scale.

Compliance is claimed insistently, with the agent described as highly compliant and the platform as strengthening compliance controls, and no regulator, statute, rule or guidance instrument is named anywhere. The omission is sharper here than for most vendors because of what the product does: anti money laundering programmes carry specific requirements about qualified personnel, documented procedures and independent testing, and a platform whose central proposition is removing the analysts who perform those functions engages those requirements directly. Naming the regime it operates inside would be the single most useful addition to its published material.

The subject is a business, which lowers protected characteristic exposure, and three inputs still carry uneven error profiles. Adverse media matching generates well documented false positives on common names and non Western naming conventions, and it is applied here to the individuals behind a company.

Website analysis to determine what a business actually does will read a polished English language site more accurately than a sparse one, so small, informal and non English businesses are assessed on thinner evidence. And coverage across more than 150 countries cannot be uniform, meaning registry quality varies by jurisdiction in ways that fall on businesses in less documented markets. The consequence is refusal of a bank account. No error rate, per market accuracy or appeal route was located.

No guarantee, indemnity or falsifiable accuracy commitment was located. One feature genuinely serves the applicant rather than the institution: the platform requests fewer input fields and gives feedback to a business while it is onboarding rather than after, reducing the back and forth of information requests, which turns a silent wait into a conversation. That is process fairness rather than recourse.

A business refused after agent review is not told a machine decided, cannot see what the web analysis concluded about its activities, and has no described route to contest an adverse media match against its owners.

Sources are described by category rather than by provider, covering company registries, social media, websites, documents and a named professional networking site for enrichment, with incorporation data across more than 150 countries implying substantial registry aggregation behind the scenes. Not one of those aggregators or data suppliers is identified, which matters because registry coverage and freshness vary by provider and determine what the agents can actually see. No model provider is named for the generative components, and no subprocessor list or hosting arrangement was located.

The adoption path is designed around the obstacle that usually blocks compliance technology, offering a modular deployment that combines with an institution's preexisting compliance stack or replaces it entirely as an end to end platform, so a bank can take a single product without ripping out what it already runs. Data reach is genuine, with entity and incorporation records across more than 150 countries and enrichment from professional networking sources. What is not published is any named system, with no case management platform, onboarding tool, screening provider or core banking system identified, and no developer documentation located.

No hosting provider, region selection, residency commitment or private deployment option was located. The question is live given entity coverage across more than 150 countries and a stated buyer base of banks and fintechs that will include European institutions with their own location requirements, and the material processed includes personal data about directors and owners gathered from online sources.

No pricing, packaging or basis of charge was located. The value case is quantified instead, at 80 percent lower operational spend against compliance teams the company says cost large fintechs millions annually, which frames the saving without indicating what the platform costs. Nothing states whether charge falls per review, per onboarding, per alert or as a platform fee, and for a product sold as replacing headcount the comparison a buyer needs is cost per review against cost per analyst.

Buyers are banks and fintechs, and the functional spread across screening, onboarding verification and transaction monitoring covers most of what a financial crime team does rather than one step of it, with the modular design letting an institution take one product or the whole stack.

Geographic coverage is the more substantive claim, with incorporation and entity data spanning more than 150 countries, which matters because business verification breaks down precisely where registry access is hardest and a platform covering only major jurisdictions forces manual work exactly where it is most expensive. What holds this at B is the absence of any evidenced deployment by institution type or market.