Sardine

The company describes itself as behaviour infused, and that is an accurate account of the architecture rather than a slogan. Proprietary device intelligence and behaviour biometrics are not a module sitting beside the other checks, they are combined into every other signal the platform produces, and the stated basis for a device network exceeding five billion profiles is exactly that continuous inference.

Detecting a user acting under social engineering pressure, or subtle behavioural drift over time that indicates emerging money laundering risk, cannot be expressed as a rule. Take the models out and the differentiating signal disappears entirely.

The oversight position is stated as a commitment rather than left to inference. The company writes that its models are designed to work alongside human judgement and not to replace it, and its agents are positioned to resolve noise and surface what matters so analysts can concentrate, covering onboarding review and investigation work rather than final adjudication. Rules and workflows are authored by the institution.

Investigation tooling assembles transactions, session behaviour, device signals, identities and network links so a reviewer sees the full context behind a flag, and rule performance analytics let a team judge how its own controls are behaving rather than trusting the vendor's word.

Three things put this above the category norm. The company states plainly that its models are transparent and explainable rather than treating opacity as proprietary advantage. Rule performance analytics expose how controls are actually behaving, including risk distribution, trends over time, payment type risk and detected anomalies, which is the raw material of ongoing monitoring.

And signals are consumable three ways, as a scored interface response, as machine learning features, or as raw signals, so an institution can feed them into models it validates itself rather than accepting a black box score. The formal package is still absent, with no model documentation, validation summary or stated drift monitoring cadence.

Scale is substantial and the named customers are unusually strong for a company of this age, including a core banking processor, a retirement services administrator and large technology platforms, with strategic investment from card network, credit bureau, ratings and core processing incumbents that also act as partners.

Published aggregates cover more than 400 enterprise customers across over 70 countries, payment volume screened above 1.3 trillion dollars, and roughly 985 million consumers. Two things hold the grade at B. There are no published per customer outcome figures of the kind competitors provide, so no institution's fraud reduction can be traced.

And the headline metrics move inconsistently across sources and dates, with device profile counts cited at 2.2 billion and later at 5.4 billion and customer counts at 250, 300 and 400, which makes the numbers hard to anchor.

Sardine handles the shared intelligence question better than the other network operators in this lane. Its consortium is named, described as sharing anonymised risk intelligence so members see mule networks and repeat abusers earlier, and explicitly framed as preserving privacy and member control, which at least acknowledges the boundary that others leave silent. The company also commits in writing to models that are transparent and explainable.

The remaining gaps are provenance and validation: the enrichment sources are named by category rather than enumerated, nothing states whether customer data outside the consortium trains shared models, and no adversarial testing of the behavioural models is described.

The data footprint is wide, combining continuous device and behavioural collection at a scale of billions of profiles with enrichment drawn from email, telecommunications, open banking and bank consortium sources, which means account level financial data enters the picture alongside behavioural telemetry. The consortium is described more carefully than most, with intelligence characterised as anonymised and members retaining control, which is a better disclosure than peers offer.

What is missing is the framework beneath it: no published lawful basis or permissible purpose position, no retention schedule, no subprocessor disclosure and no statement addressing the financial privacy service provider obligations a bank buyer must flow down.

This pass surfaced no trust centre, certifications page, attestation list or audit scope statement on the public site. Given customers including a core banking processor and a retirement services administrator, standard attestations would almost certainly be required contractually, so the published record is likely to understate the actual control environment.

The index grades what a buyer can verify without entering a sales process, so the grade records absent evidence rather than a judgement about the controls. Revisit if a trust surface is published or located.

Sardine supplies technology and holds no financial licence, the expected posture in this category. Product scope covers customer and business due diligence, sanctions and politically exposed person screening, transaction monitoring and case management, which map directly onto anti money laundering obligations, and credit underwriting brings lending regulation into scope as well.

Commercial relationships with a card network, a credit bureau, a ratings group and a core processor imply passage through the vendor risk diligence those institutions run, though no specific named authorisation is published.

Two distinct fairness exposures sit on this platform and neither is addressed publicly. Credit underwriting places the models inside Equal Credit Opportunity Act and Regulation B, where adverse action reasons must be specific and disparate impact is a supervisory concern, and there is no fair lending testing, adverse action reason code documentation or disparate impact analysis published.

Separately, behavioural biometrics infers risk from interaction patterns that vary with age, motor impairment, neurodivergence and assistive technology use, and no accessibility or demographic error analysis exists. The written commitment to transparent and explainable models is real and helps a reviewer trace a single decision, but explainability describes how a model reasons and says nothing about whether its outcomes fall unevenly across groups.

Sardine commits in writing that its models are transparent and explainable and designed to work alongside human judgement rather than replace it, which sets an expectation a customer can hold it to and implies a human is answerable for each consequential decision. It stops short of accountability. No accuracy guarantee, no remediation term, no published error rate, and no correction route for a consumer whose device or behavioural profile produced a decline or a credit denial.

Enrichment sources are named by category, covering email, telecommunications, open banking and bank consortium data, and the consortium itself is disclosed as a named product with an anonymisation and member control commitment, which is more candour about inter customer data flow than most peers manage. The specifics are absent: no individual providers named, no subprocessor list, and no disclosure of which model providers sit behind the agentic layer.

Consumption flexibility is the strength. The platform is delivered as modular building blocks behind one interface and one dashboard, and the underlying signals can be taken as a scored response, as machine learning features or as raw data, which lets an institution place Sardine anywhere from a turnkey decision layer to a signal supplier feeding its own stack.

Enrichment connects out to email, telecommunications, open banking and bank consortium sources, and the core processor relationship extends reach into institutions running that infrastructure. Public developer documentation, a service status page, a changelog and a named partner directory of the breadth seen elsewhere in this lane were not located in this pass.

Delivery is cloud hosted software as a service operating across more than 70 countries, which necessarily implies multi region infrastructure and cross border data movement. None of it is described. No hosting regions, in country residency options, transfer mechanisms or subprocessor list appear in public material, which matters given that device and behavioural telemetry plus open banking derived data are being moved across those borders.

No pricing is published at any level and every path terminates at a demo request. The one commercial structure that is disclosed is packaging rather than price, namely the claim of one dashboard, one contract and one interface in place of separately negotiated fraud, screening and monitoring tools, which does address the multi vendor cost stacking problem common in this category. Rates, billing unit and minimums remain undisclosed.

Coverage runs across banks, fintechs, merchants, marketplaces and commerce platforms in more than 70 countries, spanning the lifecycle from onboarding and account funding through payments and into credit underwriting, with all payment types supported rather than a card or transfer subset. Serving a core banking processor gives indirect reach into institutions well beyond the direct customer list.

The limit is the same one seen elsewhere in this lane: no dedicated material for credit unions, wealth management, insurance or capital markets, and a meaningful share of the positioning speaks to merchants rather than to regulated institutions.