AI Engineering Judgment/Compliance & Responsible AI
Advanced10 min

Vendor Risk & Exit Strategies

Your AI system depends on external APIs that can go down, deprecate models, change pricing, or alter terms of service. Learn to assess vendor risk, build abstraction layers, implement multi-provider failover, and plan exit strategies before you need them.

Quick Reference

  • Every LLM provider dependency is a single point of failure — plan for outages before they happen
  • Model deprecation is guaranteed: providers retire models every 6-12 months, breaking your system if you pinned to a specific version
  • Pricing changes can blow your budget overnight — build cost monitoring with alerts
  • Lock-in vectors: proprietary fine-tunes, custom embeddings, provider-specific features (assistants API, etc.)
  • Abstraction layers (LangChain, LiteLLM, or custom) let you swap providers without rewriting application code
  • Always have a tested fallback path: if your primary provider goes down at 3 AM, does your system still work?

Vendor Risk Categories

AI vendor risk is broader than traditional SaaS vendor risk. Beyond the usual uptime and pricing concerns, you face model deprecation (your fine-tuned model stops existing), capability changes (a model update changes behavior), and regulatory risk (provider's data practices violate regulations you are subject to). Understanding each risk category helps you plan mitigation.

Risk CategoryLikelihoodImpactExample
API outageMedium (quarterly)High — feature downOpenAI had 5+ major outages in 2024-2025
Rate limitingHigh (frequent)Medium — degraded serviceSudden traffic spike hits your tier limits
Model deprecationCertain (annual)High — breaking changeGPT-4o superseded by GPT-5.4, Gemini 1.5 shut down, Claude 3.5 now legacy
Pricing increaseMedium (annual)Medium-High — budget impactOpenAI raised GPT-4 prices 20% in early 2024
Behavior changeHigh (with updates)Medium — quality regressionModel update changes tone, accuracy, or format of outputs
Terms of service changeLowCritical — legal/complianceProvider starts using API data for training (opt-in → opt-out)
Provider shutdownVery LowCritical — total dependency lossSmall AI providers have shut down or been acquired
Regulatory actionLowHigh — forced migrationProvider banned in your operating region
Model Deprecation Is Not 'If' but 'When'

Every major provider has deprecated models: OpenAI deprecated GPT-3.5, GPT-4 variants, and code-davinci. Anthropic deprecated Claude 1, 2, and 3 series. Google deprecated PaLM, Gemini 1.5, and Gemini 2.0 Flash. If your system pins to a specific model version, you WILL need to migrate within 3-6 months of the deprecation announcement. Plan for this.