- H2O Driverless AI trains alert-scoring models on your own past screening dispositions, with reason codes on every score.
- H2O works beside the name and payment screening engine you already run: your lists, matching rules and thresholds stay in place.
- Enterprise h2oGPTe drafts an alert summary that cites the list entry, the customer or payment record and the policy it relied on, for your analyst to accept or change.
- Every model ships with documentation and runs under monitoring, in four deployment options.
Screening alerts outpace the team that clears them
Your screening engine compares customer names and payment parties against sanctions lists. To avoid missing a true match it has to match loosely, so common names, transliterations and partial addresses fill the queue.
A payment alert holds a payment until someone clears it. Every hour an analyst spends on a clear false positive is an hour a legitimate customer waits, while a real match sits in the same queue. Lists change often, and each change sends a new wave of alerts through your customer and payment records, and each disposition has to stand up later when an examiner or internal audit asks why an alert was closed.
How AI sanctions screening works
H2O adds a scoring and summary layer on top of the screening you already run. Your screening engine still generates every alert.
Step 1: start from your alert history
Export past screening alerts with their final dispositions, the match details your engine records, and the customer, counterparty and payment attributes behind them.
Step 2: learn which alerts turn out to be true matches
H2O Driverless AI trains classification models on which past alerts were confirmed and which were cleared. Automatic feature engineering builds features from the match details, such as the match score your engine assigns and whether date of birth, nationality and address agree. Reason codes show which factors drove each score.
Step 3: prioritize and route every alert
Each new alert gets a score and reason codes. Your policy decides the routing: likely true matches go straight to senior analysts, and no alert closes without the review your policy requires.
Step 4: draft a cited alert summary
For each alert, Enterprise h2oGPTe pulls the list entry, the customer or payment record and your screening policy, then drafts a short summary with a suggested disposition. Retrieval-augmented generation ties each statement to its source, and guardrails keep the draft in scope. Your analyst reads, edits and decides.
Step 5: deploy, monitor and document
H2O MLOps deploys the scoring pipeline, watches for model drift after list updates or new customer segments, and keeps lineage and an audit trail.
Working beside the engine you already run
Your sanctions lists, matching rules and thresholds stay in place. Customer and payment data stay in the environment you choose, in any of four deployment options.