The AI layer assists human decisions. It does not make them.
- Classify and triage support tickets (category + priority).
- Summarize recorded risk indicators for an opportunity.
- Explain an actor's computed career fit for an opportunity.
- (Extensible) opportunity classification, agency analysis, KB assistance.
- Never invent agencies, casting calls, production companies, or verification evidence.
- Never guarantee legitimacy, representation, or employment.
- Never provide legal advice.
- Never present assumptions as facts.
- Grounded inputs only. Prompts contain only the listing/ticket text that already exists; the model is asked to summarize or classify, not to research.
- Confidence on everything. Every result from
src/lib/ai.tsreturns asource(ai|rule-based) and aconfidence(low/medium/high), both rendered in the UI (e.g. the "AI-assisted · medium confidence" badge on a risk summary). - Human-in-the-loop. AI output is framed as a suggestion. Moderators make the final verification decision; analysts confirm ticket triage.
- Graceful fallback. With no
ANTHROPIC_API_KEY, every function returns a transparent rule-based result, clearly labeled as such. The product never depends on the model being available.
- Contract analysis — extract and explain commission, exclusivity, termination, likeness/AI-replica clauses. Issue-spotting and education only; always points to a qualified entertainment attorney where a clause carries real consequence.
- Digital likeness module — flag AI-training, voice-cloning, and synthetic- performance language for careful review.