Consilient

The models are what Consilient sells: federated anti money laundering, correspondent banking, typology, jurisdiction and risk rating models, synthetic training data for rare risks, and an analyzer that uses large language models to mine suspicious activity report narratives. Remove them and nothing remains to buy.

The models score and a person decides: a high risk score raises an alert for the institution's compliance team to investigate, and each alert carries the key factors and transactions behind it. Models are deployed locally under the institution's own model risk management, so the bank controls where in its monitoring they act.

Validation reports, accuracy beyond the headline figures and a mapping to model risk guidance are not published. Each flag comes with the key factors behind it, the models are built on XGBoost with feature importance and overfitting controls, and institutions deploy them locally under their own model risk management.

Each synthetic data package is tested for behavioral plausibility, distributional overlap, leakage, realistic class ratios and whether it is too easy to model, and the narrative analyzer logs how large language models are used. Consilient's own study across 12 typologies found detection keeps improving up to roughly 3,500 to 5,000 labeled examples, and models trained only on synthetic data ranked an unnamed bank's confirmed suspicious accounts 10 to 15 times better than random, adding a lift over the bank's own model that passed a formal statistical test.

Customers go unnamed, and the figures are Consilient's own: an 88 percent cut in false positive alerts, and three times the efficiency with 10 percent more cases found than a model trained on one company's data alone. The core model is described as developed with leading banks and the jurisdiction model as trained with more than 20 local banks in high risk jurisdictions, none of them named.

The cross client question is answered by design: models train on each institution's data inside its own secure infrastructure, raw data is never pooled centrally, and only aggregated model updates travel between participants. Ensembling then combines local models into the shared one, synthetic data supplies rare cases without anyone's customer records, and the narrative analyzer keeps suspicious activity reports inside the institution's environment.

Consilient, Inc.'s privacy policy covers visitor and service user information in general terms, without processing terms for the transaction data models are trained on. The stronger privacy story is architectural: under federated learning, raw data stays inside each institution and only aggregated model updates are shared.

Security certifications, audit reports and a trust center are absent from the product material, so the security of the platform that distributes models and aggregates updates has no independent assurance on offer.

Consilient states no regulatory admission of its own, selling models to regulated institutions and naming regulators and financial intelligence units as potential federated participants rather than supervisors.

Fairness or disparate impact testing of the customer risk scores is not published. Consilient's design choice on labels is stated: models are built from high risk behaviors and typologies rather than from suspicious activity reports, many of which are filed defensively.

A guarantee, indemnity or correction commitment behind a score is absent. Each alert shows the key factors and transactions that drove it, which lets an investigator check and challenge the model before acting.

Hosting for the federated platform and any subprocessors go unnamed. The modeling chain is set out: XGBoost gradient boosting, ensembling of local models and federated aggregation of model updates, built and run by Consilient itself. For the narrative analyzer the institution picks the language model, from Amazon Bedrock and locally deployed models to the OpenAI and Anthropic interfaces, and sanctions checks draw on OpenSanctions data.

Specific monitoring, case management or core banking systems are not named. Models are brought in from Consilient's platform and deployed locally into an institution's existing transaction monitoring and due diligence processes.

Models are deployed locally inside each institution's own infrastructure, and federated training keeps raw data where it already sits, with only aggregated model updates leaving the institution.

Price, licensing terms and the unit of charge for the models are unpublished, and the route in is a contact form.

Banks are the main buyers, with regulators, jurisdictions and financial intelligence units named as federated participants, a dedicated model for correspondent banking and synthetic data aimed at institutions with too few confirmed cases to train their own models. Material is organized by model rather than by institution type.