ai-automation-consultant / Failure Lab

Hallucinated AI Output Schema

The model returns confident prose where the workflow expects strict JSON.

Category

ai-output

Difficulty

standard

Last Run

none

Incident Brief

Diagnose the AI integration failure and protect the downstream workflow.

Broken State

The LLM occasionally returns a paragraph instead of the required JSON object.

Symptoms

  • Most runs pass, but ambiguous inputs fail unpredictably.
  • Downstream field access throws when output is prose.

Trace

llm.output=This looks urgent because the customer mentions tomorrow...
validator.error=Expected object, received string

Diagnosis