Switching AI Providers Without Changing the Product

The challenge

AI providers change their prices, models and terms often. A product wired to a single provider inherits all of that risk, and moving later means rewriting the parts that matter most. For a product handling financial advice conversations, where the AI runs is also a governance question. A business may want its models inside its own Microsoft Azure environment, under the agreements it already has with Microsoft, instead of calling a public API.

What we built

Two independent switches in FinPlan, each behind its own interface:

• Speech engine: Azure AI Speech, with speaker separation, is the default. OpenAI Whisper is the alternative. Both implement the same transcription interface, and the application starts with exactly one of them.
• Language model host: transcript clean-up and report drafting run on either OpenAI or Azure OpenAI. One setting picks the host, and the matching authentication, endpoints and response handling follow from it.
• Fail fast on bad configuration: if Azure is selected but its key or endpoints are missing or malformed, the application refuses to start. The problem surfaces at deployment, well before an adviser's first report. An unrecognised provider name falls back to OpenAI with a warning in the logs.

Technology

Kotlin and Spring Boot, with conditional beans selecting the transcription service and a provider abstraction that supplies clients, endpoints and authentication for OpenAI or Azure OpenAI. When Azure AI Speech is selected, speech processing runs in Azure's Australia East region. Choosing and changing providers is part of how we run An AI Platform That Turns Client Meetings Into Finished Reports in production.