Why this decision matters
Stable AI integration isolates provider variation, controls timeout and retry, protects core transactions, monitors behavior and preserves disable or rollback paths.
Treat the AI module as an external capability that can time out, fail or return uncertain output; core business must not depend unconditionally on one model call.
Conditions to confirm before development
- Isolate model, prompt and provider changes behind a service boundary
- Configure timeout, rate limits, retry, circuit breaking and caching
- Use validation, idempotency and human approval for critical writes
- Monitor volume, latency, errors, cost and business failures
Implementation and delivery approach
Begin with shadow or read-only operation, then gradual traffic. Regression-test and rehearse rollback for model, prompt, API and configuration changes.
Test provider outage, slow response, abnormal output, duplicate requests and network failure to confirm continuation or safe degradation.
Acceptance boundary
Third-party availability is not controlled solely by the developer. Service levels, fallback and additional resource cost require explicit agreement.