What the LLM Does

The LLM acts as an information extraction layer, not as the authority for workflow execution.

For example, a user might say:

“I noticed a charge of $249.99 that I don't recognize. It happened yesterday.”

The extraction layer might return a candidate structure:

{
  "intent": "report_unrecognized_transaction",
  "amount": 249.99,
  "currency": "USD",
  "transaction_date_reference": "yesterday",
  "customer_recognizes_transaction": false
}

This output is not automatically trusted. The system must validate the schema, resolve relative dates using a known clock and timezone, match the transaction against authorized records, and ask for clarification when details are missing or ambiguous.

The LLM should not independently:

• decide whether a dispute qualifies;
• skip required verification;
• call a financial action without permission;
• change workflow state outside approved transitions;
• invent facts that are not in the user's message or trusted data.

A useful design rule is: extract candidates, validate facts, then evaluate rules.