nickgallick SKILL 45: AI Provider Acceptable Use Policies
Purpose
Understand the TOS/AUP of every AI provider Perlantir depends on. Know what triggers account termination, what requires enterprise approval, and how to architect for provider redundancy.
Anthropic (Claude) — Primary Provider
Current AUP Restrictions (verify at anthropic.com/legal/aup before building)
- Gambling and betting: historically restricted/requires approval for gambling-related applications
- Financial advice: Claude should not provide personalized financial advice
- High-stakes automated decisions: using Claude as SOLE decision-maker for financial outcomes may require Anthropic approval
- Illegal activity facilitation: cannot enable prohibited activities in target jurisdiction
- Deceptive content: cannot generate misleading claims about real people/events
Pre-Launch Required Actions
- Contact Anthropic sales/partnerships BEFORE building gambling-adjacent or financial-decision features
- Explain the specific use case in writing
- Get WRITTEN approval — verbal "that should be fine" from a sales rep is not binding
- Enterprise agreement can include custom terms that override standard AUP
- Document the approval and keep it in your legal files
Risk of Termination
- Platform dependency on a single provider = existential business risk
- Termination scenario: entire platform goes dark, every agent stops working
- Anthropic has terminated accounts with limited notice for AUP violations
Mitigation: Provider Redundancy Architecture
- Design ALL AI calls behind an abstraction layer (single interface, swappable backends)
- Primary: Anthropic Claude (enterprise agreement, custom terms)
- Fallback: OpenAI GPT-4o (separate enterprise agreement)
- Emergency: self-hosted Mistral (Apache 2.0, no use restrictions)
- Target: swap to fallback provider within 24 hours without product rewrite
OpenAI (GPT) — Backup Provider
- Similar AUP restrictions on gambling, financial advice, high-stakes decisions
- Has been MORE aggressive about enforcement and account termination with less warning
- Key restriction: "activities requiring professional licenses (legal, medical, financial)"
- Prediction market scoring could trigger this
- Enterprise agreement available; negotiate custom terms explicitly authorizing use case
Google (Gemini) — Alternative
- Restricted uses: gambling, financial services without compliance
- Generally less restrictive than OpenAI, more than open-source
- Enterprise terms available via Google Cloud
- Vertex AI provides enterprise-grade terms with explicit compliance representations
Open-Source Models — Last Resort / Emergency Fallback
| Model | License | Commercial Use | Use Restrictions |
|---|---|---|---|
| Mistral 7B/8x7B | Apache 2.0 | ✅ Unlimited | None |
| Llama 3 | Meta Community License | ✅ (< 700M MAU) | Some restrictions |
| Phi-3 (Microsoft) | MIT | ✅ Unlimited | None |
| Falcon | Apache 2.0 | ✅ Unlimited | None |
- Self-hosting eliminates provider termination risk entirely
- Trade-off: lower capability than frontier models (narrowing gap as of 2025-2026)
- Best for: fallback/emergency use, deterministic scoring tasks, high-volume/low-complexity operations
Strategic Architecture Principle
Never build a business that dies if one API provider terminates you. This is as important as database redundancy and uptime SLAs.
This is legal research and intelligence, not legal advice. Consult qualified legal counsel before taking action.