Choosing the AI model
Select one of the model cards and the analysis runs on that model. If you choose nothing, the default model is used, and the edit screen shows that no AI model is selected so the default will be used.
Step 3 — AI model selection. Without a choice, the default model is used.
The models you can choose
The base rate for one analysis differs by model. The rates below include up to 5 steps; from the 6th step a surcharge applies — how credits are charged. Engine call credits are added on top of this.
If your data has to be processed only inside ARGOS infrastructure, use Gemma4 12B — ARGOS hosts it directly.
You can check what was selected in
aiModelId and aiModel from GET /workflows/:workflowId.
Engines are assigned automatically
When you save the policy, Omni reads it, builds the playbook actions, and attaches the engine each action needs. The execution order is decided at the same time, so you never configure dependencies yourself. A clause such as “run AML screening on the managing member’s name” becomes an action with the AML Search - Person engine attached. A clause that needs no external lookup, such as “if the ownership percentages total more than 100%”, produces an action with no engine. Action names differ every time the playbook is generated. You can see the result in two places.
To change an assigned engine, use the step engine editor in editing a workflow.
Engine catalog
The description of each engine is also visible in the step engine editor — editing a workflow.
The
- KOR engines support Korean businesses and accounts only, and Korea is not among the 40 regions Address Validation supports. When the target country does not match, the engine is either not assigned or the call fails — so state which country the verification subject is in when you write the policy.Text Verifier
An engine that decides from document text alone. It queries no external database and uses only the items uploaded to the profile. In most workflows it is the engine that runs most often.
It extracts values, compares fields across documents, checks that required entries are complete, and judges policy compliance. Results land in
extractedData (in the output-schema structure) and in agentAuditLog.steps[].
AML Search - Person
An engine that searches external watchlists. Tool calls are recorded inagentAuditLog.steps[].mcpcalls[].
Input — the agent pulls these from the documents and passes them in.
Match score — the average of the per-field scores. The name is scored with a string-distance algorithm and an alias library; date of birth and nationality score 100% for an exact match and 80% for a partial one.
Result status
Risk icons — matched subjects are classified with these codes.
The database sources and code details are in AML database sources and codes.
Where to see the results
Engine execution shows up in three places in the analysis response.
How to read them is covered in reading analysis results.