How the agent recommends an outcome
The agent proposes an outcome and names the findings behind it. It never inherits a decision from an earlier case.
At the end of an investigation the agent recommends one of the five outcomes. The recommendation sits above the decision form. Nothing applies it until you do.
The counterparty is on a sanctions list and the customer's declared wallet funded it directly. Direct exposure, not traced through an intermediary.
The data here is invented.
Accept and Reject are both still open, which is how you can tell nothing has been applied.
What a recommendation contains
- The proposed outcome
One of the same five you can choose from.
- The reasoning
A short note. It leads with the deciding fact and the conclusion. Where the outcome keeps the case open, it ends with the one thing needed next.
- The findings it rests on
Numbered, and cited by number in the reasoning. Each one opens, so you can check the claim against its sources.
- The typology, where one applied
Named only when it informed the outcome.
- The SOP sections it relied on
Cited by document, version and heading, so a reviewer can see which rule the agent applied.
What each outcome obliges it to say
| If it recommends | It must |
|---|---|
| Escalate to a reviewer | Name what the reviewer needs to confirm |
| Request for information | Supply one to four questions the customer can actually answer |
| Hold | Name the external thing the case is waiting on |
A recommendation that keeps a case open without saying what would move it is of no use to whoever picks the case up next.
Two things it will not do
It will not recommend closing as a false positive where a sanctions screen in this run came back unavailable. No list was checked, so the case is not cleared.
It will not let verification clear onchain risk. A verified customer in good standing does not explain away exposure to a sanctioned entity, and the agent still escalates.
It does not inherit an earlier decision
The agent can see what a person decided on this case before, and what was decided on related cases. It treats both as context and assesses again from the evidence. Run an investigation twice and the agent reaches the second recommendation independently.
Related
Last checked against the product on .