AI agents

Agents that prepare the case, and people who decide it

Ten agents, one per stage of the case. They read, check, screen, rate and draft. They don’t clear matches, waive checks, send to customers or decide.

The ten agents

One stage each, and a line none of them crosses

Every agent writes what it found and why. When a person needs to act, the agent prepares the decision and waits.

AgentStageWhat it preparesWhat it never does
IntakeIntakeWhat is on file and what is missing; a request for the missing documentsSend the request
Document readerIntakeFields from each document, quoted from their lines; the ones it couldn’t readInvent a field
OwnershipVerificationBeneficial owners from the registry; a request about undisclosed owners or officersInvent an owner or a percentage
VerificationVerificationIdentity, address, MRZ and company checks; a request to fix what failedPass a check without evidence, or waive one
ScreeningScreeningA proposed disposition for each match, with the reasonClear a strong sanctions match, or dispose of anything itself
RiskRiskThe rating with every factor’s points and reasonOverride a rating
EDDRiskWhether enhanced due diligence is needed; the request for source of funds and wealthRecord EDD as complete
Case writerDecisionThe case memo, from the checks, screening, ownership and riskAdd a fact that isn’t on the case
SupervisorDecisionWhether the case is ready for a decision, and what is in the waySubmit or decide the case
MonitoringCustomer bookAlerts from transactions against each customer’s declared profileClose an alert or open a review
Autonomy

Autonomy is earned, one level at a time

Each agent starts at L1. An administrator can raise it once people agree with what it prepares; the tenant ceiling caps every agent at L2 (Ask first). An agent moves up only after people agreed with at least 90% of 20 answers, and drops back if agreement falls below 70%.

LevelNameWhat the agent may do
L1DraftPrepares a draft; a person does the work
L2Ask firstPrepares the decision and asks a person before it acts. The ceiling for every InsightKYC agent today
L3Act and tellActs, tells the person, and lets them pull it back. Locked
L4ActActs on its own. Locked
Grounding

The model explains. It never supplies the facts.

Checks, ownership arithmetic, match scores and risk points are deterministic code. Where a language model is used, it reasons over the case’s facts and phrases what the agent found; an answer that cites something not on the case is rejected and the agent falls back to its rules.

Deterministic

Arithmetic is code

Effective ownership, check digits, match scores and risk points are calculated, not generated.

Grounded

Facts come from the case

Every line an agent writes refers to a document line, a registry record, a list entry or a transaction.

Fallback

Model off is a supported mode

With the model switched off, every agent runs on its rules and says so on each step.

Prepared decisions

Accept, reject or ask

What an agent prepares lands on the case and on the analyst’s home. Accepting carries it out; rejecting records that a person disagreed; asking sends a question back. Each answer is kept on the decision record and counts toward the agent’s agreement rate.

Watch the agents prepare a case

We run them on the demo bank’s cases and walk through every prepared decision.

Request a demo