Tookitaki

The removal test leaves exactly the product this company defines itself against. Federated machine learning models trained on real world typologies do the detection, an agentic copilot conducts investigations and drafts case narratives, network analytics surface relationships between customers and counterparties, and anomaly detection reduces false positives.

Strip all of it and what remains is a rules engine with static thresholds and a case queue, which the company names explicitly as the legacy failure mode it exists to replace, arguing that fixed thresholds cannot keep pace with laundering techniques that change faster than they can be rewritten.

Four mechanisms combine and the boundary is stated rather than implied. The company writes that machine learning will not replace compliance professionals but augment them, and that the goal is not automation for its own sake but better detection with less friction, which is an explicit refusal of the autonomy most vendors in this category advertise.

Every flagged transaction carries an explanation of the data points and logic behind the decision, so an investigator inherits reasoning rather than a score. Low code threshold controls let compliance teams adjust detection parameters themselves without engineering support, which puts calibration in the hands of the accountable function. And rules can be simulated for effectiveness before deployment, so an institution sees the operational consequence of a change before customers experience it.

The most complete model lifecycle position in this index alongside Upstart, and unlike most it is described as built into the product rather than performed around it. Performance metrics, drift detection and retraining workflows are stated as part of lifecycle management, which addresses the failure mode that actually degrades financial crime models, namely that criminal behaviour changes and a model tuned last year quietly stops working.

Independent validation comes through a national testing framework rather than from the vendor's own assessment. Every decision is explainable at the data point level, so an examiner can interrogate a specific alert. And simulation lets an institution measure the effect of a threshold change before it takes effect. What is still absent is a published accuracy or false positive rate, which would let a buyer compare rather than trust.

The collaborative network is the headline number, with more than 200 institutions contributing to it and over 1,200 risk scenarios accumulated, spanning investment scams, mule operations and cross border layering across Southeast Asia.

One customer is named directly and quoted, an all in one financial platform describing itself as the leading digital bank in its market, on the strength of the platform and the support behind it, and a second quoted customer speaks to reduced false positives without being identified. Pre configuration for four named supervisors is itself evidence of deployments in each of those jurisdictions.

What is absent is measurement: false positive reduction is described as significant without a figure, no transaction volume or alert count is published, and network participation is not the same as paying customers.

The most sophisticated answer to this axis recorded in the index, and it is the only one that enables sharing rather than preventing it. Every other vendor here either pools customer intelligence silently or forecloses pooling in a sentence. This company does a third thing: it pools deliberately, names the mechanism that makes it safe, and states precisely what crosses the boundary and what does not.

More than 200 institutions contribute anonymised typologies, red flags and fraud patterns to a common pool now exceeding 1,200 scenarios, and federated learning distributes the resulting model improvements without moving the underlying data, so a bank in one market benefits from a mule network detected in another without either seeing the other's customers. That is the sixth benchmark answer to the cross party boundary problem after Mitigram, DiligenceVault, Saphyre, DwellFi and Omnisient, and the first federated one.

The privacy commitment is architectural and sits at exactly the point where a shared intelligence network would otherwise be most exposed. Members benefit from insights across the whole network through federated learning, which means models travel between institutions while customer data does not, and the company states plainly that institutions gain shared intelligence without sharing sensitive data. Contributed material is anonymised typologies and patterns rather than records.

That is a genuine mechanism rather than an assurance. Held at B because no retention schedule, subprocessor list or data processing terms were located for the platform itself, and several markets served impose their own localisation requirements.

No attestation, certification, trust centre or enumerated framework was located. The national artificial intelligence testing framework validation is meaningful and it assesses model governance rather than organisational security controls, so it does not substitute.

For a platform holding transaction data and investigation case files on behalf of banks across several supervised markets, and operating a shared intelligence network between them, a published assurance set is what each of those institutions' examiners will expect to see.

Four supervisors are named and the platform is stated to be pre configured for each, covering the monetary authority in Singapore, the central bank in the Philippines, the financial intelligence agency in Australia and the central bank in Malaysia, which is operational alignment rather than a compliance claim because it means the detection scenarios, reporting formats and thresholds are built to those regimes rather than adapted afterwards.

Suspicious transaction report narratives are generated to the format one of those supervisors expects, the international standard setter is referenced, and the explainability design is tied directly to a named central bank's stated emphasis on governance. Models are additionally validated through the national artificial intelligence testing framework operated by the Singapore government.

Models are validated through a government operated national artificial intelligence testing framework, which is independent assessment of governance properties including explainability, robustness and fairness by a body with no commercial interest, and it is the same reasoning that earns Incode and HyperVerge their standing except applied to general model governance rather than to face recognition.

Universal per decision explainability supports it, since every flagged transaction exposes the logic behind it and a wrongly flagged customer's case can therefore be examined rather than merely appealed. Two qualifications belong on the record: that framework assesses process and documentation more than demographic outcomes, and no false positive analysis across customer populations is published, which matters because mule and scam detection in this region falls heavily on migrant workers and remittance corridors where unusual transaction patterns are normal behaviour.

No guarantee or indemnity was located, and what earns the grade is that error is findable by design rather than promised against. Because every flagged transaction carries an explanation of the data points and logic behind it, an institution can establish exactly why a customer was held and defend that to a supervisor, and a wrong decision can be traced to the signal that produced it rather than disappearing into a score. Automated narration produces the documentation trail alongside it.

What is missing is the customer facing half: a person whose account is frozen or whose transaction is blocked is not told which system reached that conclusion and has no described route to contest it, and a mule designation is an accusation that follows them.

The intelligence source is not merely named but quantified and explained, with the collaborative network identified as the origin of the typologies driving detection, its scale disclosed at more than 200 contributing institutions and over 1,200 scenarios, and the federated mechanism by which that intelligence reaches members described rather than glossed.

Sanctions lists and fraud typology databases are named as categories, and the national testing framework validating the models is identified. What remains undisclosed is the generative layer, with no model provider named for the investigation copilot or the automated narration that drafts regulatory filings, and no subprocessor list or hosting arrangement located.

The consolidation argument is the integration argument here, since the company identifies poorly integrated tools creating data silos as one of the central problems it solves, and unifies fraud prevention, transaction monitoring, screening and case management on one platform so a signal in one surfaces in the others.

Connections to fraud typology databases, sanctions lists and the community intelligence network are described, and a modular design lets an institution adopt components incrementally. What is not published is the list a buyer needs: no core banking system, payment platform or customer relationship system is named, and no developer documentation or interface reference was located.

No hosting provider, region selection, residency commitment or private deployment option was located. The federated architecture reduces what is at stake for the shared intelligence layer, since customer data is stated not to leave the institution for model training, but it says nothing about where the platform's own monitoring, screening and case management processing occurs. That question is live across a footprint spanning several jurisdictions with their own financial data localisation requirements.

No pricing, packaging or basis of charge was located. The modular structure implies component based commercial terms, since an institution taking only transaction monitoring has a different footprint from one running monitoring, screening, fraud and case management together, and participation in the intelligence network may carry its own arrangement. Nothing indicates whether charge falls per transaction screened, per alert, per institution or on assets.

Coverage is deep within one region rather than broad across the world. Institution types span traditional banks, digital banks, fintech platforms and payment institutions, and the functional spread is complete for financial crime, unifying transaction monitoring, fraud prevention, sanctions and adverse media screening, case management and investigation in a single platform rather than the siloed tools the company identifies as a core problem.

Jurisdictional depth is the differentiator, with pre built alignment to four supervisors across Singapore, the Philippines, Australia and Malaysia and typologies specific to Southeast Asian crime patterns that the company argues generic global platforms miss.