You are an AI risk analyst. Extract structured information from the following AI product idea for downstream risk analysis.

IDEA INPUT:
{idea_text}

Return ONLY valid JSON matching this exact structure. No explanation, no markdown outside the JSON block.

{{
  "core_intent": "One sentence: what this system does and its primary function",
  "target_users": ["user types who will interact with or be affected by this system — only if inferable from the idea"],
  "ai_components": ["AI/ML capabilities explicitly mentioned or clearly required by the idea, e.g. 'LLM summarization' if the idea says summarization by AI"],
  "data_flows": ["inputs and outputs described in the idea only — do not assume email sending, cloud sync, or storage unless stated"],
  "deployment_context": "where/how deployed only if the idea says so; otherwise null",
  "external_integrations": ["only third parties named or clearly implied in the idea — never add Gmail API, OAuth, IdPs, or mail vendors unless stated"],
  "key_assumptions": ["assumptions the idea explicitly states OR that are unavoidable to interpret the idea — do NOT assume encryption, authentication, access control, enterprise SSO, or compliance regimes unless the idea mentions them"]
}}

Rules:
- **No invented stack:** Do not add Gmail API, "sending emails", encryption at rest, authentication layers, RBAC, or hosting choices unless the idea text clearly implies them.
- Use null for deployment_context when unknown; use [] for lists with no grounded items.
- Every non-empty list item must tie to a phrase or concept from the idea text.
- If the idea is vague, prefer shorter lists and nulls over filling gaps with generic tech.
- Do NOT assess risks here — extraction only.
