How to Use Jev with Claude Code and Codex for KYC, KYB, and AML

Jev is a model from TypeSafe AI that reads text, like an onboarding application, an adverse media article, or an alert summary, and returns typed decisions instead of prose: a category, a score, or the probability that a statement is true. It answers in well under a second at a fraction of a cent, so you can run it on every case that enters a compliance queue.

Compliance work has a volume problem. Most screening hits and monitoring alerts close without action, and every one of them still waits in line with the few that matter. An analyst opens them in the order they arrived, not the order of risk.

This is a guide to fixing that with two layers: Jev as a pre-trigger that sorts and ranks every onboarding application, KYB and KYC review, sanctions and adverse media hit, and AML alert, and a Claude Code or Codex agent running in a Runtime sandbox that prepares the evidence for the cases Jev puts first. Your compliance team makes every decision.

The short answer. Send each case to Jev with a few questions your compliance team wrote: what kind of case is this, how much review does it need, is a specific red flag present. Code applies thresholds and decides which cases an agent prepares first. The agent gathers evidence from your verification, screening, and monitoring tools and drafts a packet. Jev never approves, declines, clears, or files anything.

What it does. High-risk applications and alerts reach an analyst first, already prepared, instead of sitting behind a hundred name-match false positives.

What it is not. It isn't a decision engine, a screening vendor, or a replacement for your procedures. Every material improvement comes from a human questioning a premise. Plan for that.

What Jev returns

Jev answers three kinds of questions: a choice from a list you define, a score across levels you describe (numbered from 0), and a noul, the probability that a statement is true. Here is the difference in practice:

reply = llm.generate(
  "Does the website match the "
  "application? " + application
)

Mostly no. The site seems to
sell CBD, which may be restricted
depending on your policy.

result = client.system_one(application, {
  "mismatch": Noul(instructions=
    "Website and application differ"),
})

{
  "mismatch": {"noul": 0.89}
}

The rule that matters most

Jev ranks. It never disposes. This is stricter than in a support queue, for three reasons.

No rationale. Jev returns probabilities, not reasons. A closed alert, a cleared sanctions match, or a declined application needs a documented reason that a reviewer, an auditor, and an examiner can follow. A score of 0.91 isn't one.

Fair treatment. Onboarding decisions affect who gets access to financial services. Simon Willison's early review of Jev raised bias concerns for exactly this kind of sensitive ranking. Keep Jev's output on questions about the business and the activity, never on personal characteristics, and test for skew across segments before you rely on it.

Model governance. In April 2026, US bank regulators replaced SR 11-7 and the 2021 BSA/AML model statement with revised model risk guidance. It excludes generative and agentic AI from its scope. Jev returns probabilities rather than text, so your model risk team may treat it like a classification model. If you run a program with a sponsor bank, expect the bank's own model governance to apply. Either way, document how you tested it.

A ranking mistake delays a case. A disposition mistake can become a finding. Keep Jev on the ranking side.

How the pre-trigger works

Jev reads every case and orders the queue. The agent prepares the cases that cleared your threshold. The compliance officer sees a finished packet and makes the call.

The stack

Job	Tools
Fast decisions	Jev by TypeSafe
Investigating agent	Claude Code Codex
Onboarding and KYB	Alloy Middesk
KYC and identity	Persona
Sanctions and adverse media	ComplyAdvantage LexisNexis
Transaction monitoring and cases	Unit21 Hummingbird
Activity data	Snowflake Or scoped Postgres views
Agent infrastructure	Runtime

Runtime is an operating system for coding agents. Each agent gets its own sandboxed computer, a credential vault, network rules, skills that persist between sessions, and a record of what it did. For compliance, that last part matters: you can show which records the agent read for every packet it prepared.

Where Jev fits in compliance

The last row is the second use for Jev. TypeSafe positions it for scoring and checking other models' output, not just routing. A draft narrative that's missing the "why" goes back to the agent before it reaches an analyst.

Step 1: Write the questions (compliance)

The compliance team writes these, not engineering. Jev answers exactly what you ask, literally, and TypeSafe's guidance is to ask one specific thing per question and combine the answers in code.

Compute numbers and dates in code. TypeSafe's documentation says Jev struggles with both. Don't ask whether deposits fall just under a reporting threshold or whether a company is less than a year old. Calculate that in code and pass the result as a sentence in the state: "11 cash deposits in 9 days, all between 90% and 99% of the reporting threshold."

Don't use Jev as a name matcher. Name matching across transliterations and aliases is what your screening vendor is built and tuned for. Jev ranks what the vendor returns.

Step 2: Set thresholds, then prove them

Your compliance team writes the routing rules as plain policy that an auditor can read.

Keep the sensitive thresholds low. A prohibited-industry false alarm costs an officer two minutes. A missed one costs far more.

Then run it in shadow mode. Send several weeks of historical cases through Jev, compare its ranking with what your analysts actually escalated, and pay most attention to the high-risk cases Jev ranked low. That comparison is your validation evidence. Keep it.

Sample the bottom of the queue every week once it's live. The risk with any ranking system is that important cases sink quietly.

Step 3: Build the triage with Claude Code or Codex

Put your TypeSafe API key in an environment variable and ask your coding agent to build it. TypeSafe has Python and TypeScript SDKs and publishes an agent skill, and community projects like jev-code register Jev as a tool inside Claude Code and Codex.

"Build a triage service for our compliance queues. For each new onboarding application, adverse media hit, and monitoring alert, compute the numeric facts in code, then send the text facts to Jev using the TypeSafe SDK with the questions in compliance-questions.md. Apply the rules in routing-policy.md. Log the input, every Jev answer, the model version in the response, and the rule that fired. Never close, clear, or approve anything. Write tests from the labeled cases in /fixtures and report every high-risk case that would have been ranked in the bottom half."

Two habits make these prompts work harder:

• Say why, not just what. "Log the model version, because we have to re-validate when TypeSafe ships a new one" gets you an audit trail instead of a debug log.
• Name the failure you're preventing. "A sanctions potential match must never leave the human queue, even with a very low score" exists because that's the finding nobody wants.

Step 4: Create the Runtime template

The template is what the investigating agent boots into when Jev hands it a case: the procedures, tools, and credentials it can use, and the hosts it can reach.

"Create a Runtime template for compliance case preparation. Connect Alloy, Middesk, Persona, ComplyAdvantage, and Unit21 with read-only access. Add a read-only connection to the masked activity view. Allow network access only to those hosts. Load every credential from secrets. The agent can save draft packets but can't change a case status."

Add credentials through the template's secrets in the dashboard, never by pasting them into chat. Give the agent its own identity in each vendor so your access logs show exactly what it read. Then test the limits: ask it to close an alert or approve an application and confirm the request fails.

Attach your procedures as skills: one per typology, one for your restricted industries policy, one for EDD. When Jev says "restricted industry," the agent loads that policy first.

Decide what reaches Jev separately from what reaches the agent. Jev needs a business description, alert facts, or article text. It rarely needs identity documents or account numbers, so leave them out and treat TypeSafe like any other third-party processor in your vendor review.

Step 5: Let Jev start the agent, then check its work

When a rule says "prepare," the triage service starts a Runtime session with the case, Jev's answers, and the rule that fired.

"When the routing rule is 'prepare now,' start a session from the compliance template with the case, the Jev scores, and the matched category. Load the skill for that category. Gather evidence with read-only access and cite every record. Separate facts, supported explanations, open questions, and missing information. Save a draft packet to the case and post a summary to #compliance-onboarding. Don't recommend a disposition. List the questions the reviewer needs to answer."

Before the packet reaches a person, send the draft narrative back to Jev with a short rubric: are who, what, when, where, and why each stated, and does every claim cite a record. Drafts that fail go back to the agent. Analysts spend their time on judgment, not on returning incomplete work.

What stays human

Where it breaks

Numbers and dates. Structuring patterns, company age, and date-of-birth checks all fail if you ask Jev directly. Compute them first.

Applicants write the input. A business description is written by the applicant, and TypeSafe notes Jev can be misled by adversarial content. That's why the website noul matters more than the stated description, and why no Jev score should ever shorten a review.

Literal reading. "Is this a money services business" and "does this business transmit money for others" get different answers. When routing looks wrong, fix the question before blaming the model.

New model versions. Every response includes a versioned model ID. When it changes, run your shadow set again before trusting the new rankings.

Nothing explains itself. Log every input and answer next to the case. When an examiner asks why a case waited three days, you need to show what Jev said and which rule applied.

Measuring it

• Time to first review for high-risk cases. The number a pre-trigger exists to move.
• High-risk cases found in the bottom of the queue. From your weekly sample. The target is zero.
• Analyst minutes per case. Split by queue, since EDD packets and name-match hits aren't comparable.
• Packets returned for missing evidence. Before and after the narrative check.

Measure the current queue first, or you won't be able to show what changed.

A realistic timeline

The point

Compliance doesn't lack judgment. It spends it in the wrong order. When sorting a case costs almost nothing, every application and alert can be ranked the moment it arrives, and the analyst's first hour goes to the case that deserves it.

Jev ranks, the agent prepares, and your compliance team decides. Keep those three jobs separate and the whole thing stays defensible.

Back to the case.

For the same pattern in support and payment ops, see how to use Jev with Claude Code and Codex for triage. For the investigation agent itself, see the compliance alert and KYB and KYC review guides.