SentiLink

Each score is a machine learning model trained on fraud labels, and the scores are what institutions buy: synthetic, identity theft, first party and contact risk models run on every application, while Facets and Intercept repackage the same features for in house models and analysts. eCBSV, ID Complete, watchlist matching and the Treasury check service are data services that work without models, and they sit around the scores as treatments and add ons.

SentiLink scores and the institution decides. Every score returns reason codes for its high and low risk factors, and clients set their own thresholds and design step up treatments around them, such as an eCBSV check for a suspected synthetic or a one time passcode sent to whichever of the applicant's phone or email scores as lower risk.

Analysts can escalate hard cases from the dashboard to the Fraud Intelligence Team, which also reviews a subset of each day's applications to keep the models current, and Intercept gives investigators a single view for the cases that need a human call.

Each model learns from hundreds of thousands of labels assigned by SentiLink's analysts, with a clear definition for each fraud type, and returns a 0 to 1000 score with reason codes, while Insights adds a plain description of each flag; accuracy, false positive rates and model versions are not published.

Analysts review cases daily so the models keep pace with new tactics, and the first party model rests on a stated definition, accurate core personal details paired with misleading or incomplete financial information, rather than on borrower intent, which SentiLink calls impossible to measure.

Cherry's senior manager of fraud strategy says ID Complete filled in identity data for 92 percent of incoming applications and Insights cleared underserved applicants at a match rate above 80 percent, which grew Cherry's application funnel by 5 percent. Fraud and data science leaders at Citizens, Concora Credit, Current and Cardless speak to the partnership without figures.

For the fraud scores, retrostudies across more than a dozen client companies found 2.89 percent of funded loans and booked accounts high risk for first party fraud, and first party fraud made up 72 percent of the preventable losses across the three fraud types.

Learning runs across the network. The first party score draws on more than 3 million applications a day from hundreds of institutions, Facets carries network intelligence from account opening activity, and the synthetic score blends curated sources with SentiLink's internal consortium data. The boundary it states is that an institution's own performance data is not shared with other institutions. How long application data is held for training is not set out, but customer data is deleted when a client leaves and processing stays on AWS in the United States.

The privacy policy is written for users of SentiLink's website and services rather than the consumers its scores assess. It says machine learning may process personal information under accuracy and security safeguards and that SentiLink does not sell or share personal information, and it carries California consumer privacy disclosures, while the Gramm Leach Bliley Act goes unaddressed. eCBSV checks run only with the consumer's consent, and the trust center lists AWS as the only subprocessor, with data retention procedures among its controls.

SentiLink's trust center lists a SOC 2 Type 2 report, from an auditor it does not name, an attestation of compliance under version 4.0.1 of the payment card industry standard and a penetration test report, along with its security policies, the controls behind them and the subprocessor list, released on request.

The security program is modeled on frameworks from the National Institute of Standards and Technology (NIST), and its listed controls include cybersecurity insurance and tested disaster recovery plans.

SentiLink was the first provider to offer eCBSV, the Social Security Administration's consent based verification service, and as its longest tenured user it works under the eCBSV User Agreement, which the agency enforces through periodic audits of sampled transactions. Its terms state that it is not a consumer reporting agency under the Fair Credit Reporting Act and that its services are not consumer reports and are not meant to decide eligibility for credit, insurance or employment. Customer identification and watchlist matching help institutions meet their own onboarding rules.

Bias results by demographic group are not published, and no governance process for the models is described, although inclusion is part of SentiLink's pitch. It markets the synthetic score as telling real synthetics apart from groups other models wrongly flag, such as recent immigrants and students, and eCBSV as clearing young people, immigrants and people new to credit at a high match rate.

Recourse is limited to reasoning. SentiLink publishes no service terms or remedy for a wrong score, and because its terms say the scores are not consumer reports, it describes no route for an applicant to dispute one. What a client gets instead is the basis for every result, reason codes on each score and a plain description of each Insights flag, so a decision built on a score can be explained and reviewed.

The licensed commercial data behind the fraud scores goes unnamed. The models are SentiLink's own, trained on its analysts' labels and hosted on AWS, and the government sources it draws on are named where they apply: the Social Security Administration for eCBSV, state motor vehicle records for driver's license checks and the US Treasury for check verification.

Institutions call the scores through an API or work in SentiLink's dashboard, and the API documentation sits behind a client login. Partners carry the scores into their own platforms: Alloy for onboarding, Persona for identity orchestration, FiVision for account opening, Lendflow for small business lending and Scienaptic for credit decisioning. Intercept and the Treasury check service connect to official sources, from state motor vehicle records to the US Treasury's check verification interface.

An on premises or dedicated deployment is not described. The hosted service runs on AWS in the United States, the only subprocessor listed in SentiLink's trust center, and clients sign in through single sign on with layered access controls.

Scores are priced through a demo with SentiLink's team, with no published fee per score, tier or volume rate. Intercept is sold on predictable pricing so analysts are not penalized for each extra lookup, and the Treasury check service is free to banks, credit unions and other financial institutions.

Banks, credit unions, banking as a service providers and fintechs run the scores on deposit accounts, credit cards, personal, auto and small business loans, and buy now, pay later financing, and SentiLink counts 13 of the top 15 banks, 8 of the top 10 credit unions and more than 50 fintech unicorns among its clients.

Captive auto finance companies, patient financing providers, telecom and energy utilities and tenant screening firms use the same scores, and in SentiLink's own data first party fraud runs highest for auto lenders and telecoms. All of it is in the United States and keyed to the Social Security number.