# ADR-004 — Deployment-selected AI provider

- **Status:** Accepted
- **Date:** 2026-07-30
- **Phase:** 5 (AI improvements)

## Context

The AI sidecar supports Ollama, Gemini, and Groq. Earlier documentation described a per-request provider abstraction with OpenAI, Claude, administrator controls, and user choice; none of those capabilities exist. Adding them now would increase credential handling, privacy exposure, testing, and billing complexity without demonstrated customer demand.

## Decision

Each deployment selects exactly one provider through `AI_PROVIDER`. The .NET application continues to call the private `ai-service` through `ISummarizerService`; application code does not branch on provider. Ollama remains the local privacy-first option. Gemini and Groq are operator-selected cloud alternatives.

Provider choice is not exposed to users or administrators. Revisit this only when a customer requires provider choice and the deployment has explicit credential, privacy, quota, and audit rules for every enabled provider.

## Consequences

- One provider configuration and failure mode per deployment.
- No OpenAI or Claude support is implied.
- AI results remain suggestions that require user review.
- Usage metering belongs at the application interaction boundary, independent of provider.
- Cloud-provider privacy and cost are deployment responsibilities until SaaS quotas are introduced.