Flowvenue does not assume that the LLM is always correct. The model interprets a request; the runtime verifies whether that interpretation can be turned into a valid action.
Ambiguous requests
When essential information is missing or multiple interpretations are possible, the correct behavior is not to invent. The system should ask for clarification or request missing data.
Invalid parameters
Wrong types, invalid dates, unknown fields or incompatible values should be intercepted before persistence when possible, returning structured errors that support targeted correction.
Integration failures
Timeouts, authentication failures or external system errors should not be masked by plausible language. The process distinguishes completed, failed and recoverable actions.
Human-in-the-loop
When a step requires confirmation or approval, execution can pause and wait for human intervention.
Core principle
LLM interprets; runtime validates. Reliability does not come from promising that the model will never be wrong, but from preventing an interpretation error from automatically becoming a wrong action.