An example of identity document verification as a governed process: agents do the 8 steps, and a person signs the gate before anything irreversible happens. Download it, import it into your workspace, and adapt it to your systems.
minctrl.process/v1 · hash 6a845a9bc036…059a
Document and liveness adjudication at onboarding scale means human reviewers eyeballing millions of ID scans and selfie-match results a year, a slow and costly bottleneck that throttles activation. Yet a wrongly auto-approved synthetic identity is a fraud loss and a wrongly auto-rejected real customer is an adverse-action and reputational hit nobody wants to own. Clean document-authentic, liveness-passing, selfie-matching submissions are auto-approved in seconds with a documented rationale and audit trail. Only ambiguous, deepfake-suspect, or synthetic-identity cases reach a fraud/ops officer, who reviews and signs every fraud disposition and every final REJECT that denies a customer access.
Rejecting a genuine customer (or approving a fraudulent identity) is consequential and can't always be cleanly undone, so borderline or failed identity cases route to a human adjudicator before the account is opened or denied.
| Rule | What it requires |
|---|---|
| KYC / Customer Identification Program (CIP) | Under the BSA CIP rule, an institution must verify enough identity information to form a reasonable belief it knows the customer's true identity before or shortly after account opening. |
| Liveness / presentation-attack detection | Document and biometric checks must detect spoofing (presentation attacks) — a genuine, live person matching a genuine, unaltered document — not just parse an uploaded image. |
| Fair-treatment & bias considerations | Identity decisions should account for demographic performance differences in face-matching, and a rejected genuine customer must have a path to human review. |
A summary for orientation, not legal advice.
A minctrl.process/v1 document: the graph and the hash of the process it becomes.
Import into Minctrl opens this example in your workspace; one click stores it as your first version. Import JSON does the same with the file, and a file changed after download is refused.
Every connector starts in shadow: steps run, nothing leaves Minctrl until you configure a connector and switch it live.
Change signers, time limits and steps on the canvas. Each save is a new version; a running case stays on the version it started with.
The example models the checks a Customer Identification Program calls for: it verifies identity information to support a reasonable belief in the customer's true identity, and records the document, the checks run and the decision with a tamper-evident audit hash. Whether that satisfies your CIP depends on your written program, your risk assessment and the verification vendors you bind. Minctrl holds no certification, and this is not legal advice.
The AI agent checks document authenticity and runs liveness / presentation-attack detection and face-matching, so a genuine, live person must match a genuine, unaltered document. Cases that don't clearly pass are escalated rather than auto-approved.
It routes to a human adjudicator before the account is opened or denied. Because wrongly rejecting a genuine customer — or approving a fraudulent identity — is consequential, borderline and failed cases don't auto-decide; a person owns the call and it's logged for review.