Vendor Lock-in and LLM Gateways

Most AI pilots start the same way: one engineer picks one provider, installs one SDK, and pastes the provider's client call into the place where the feature needs an answer. It works, and nobody questions it. Six months later, the pilot has become a product. The provider's client library is imported in forty files, prompts are tuned to one model's quirks, and the response parsing assumes one vendor's JSON shape. Then something changes: a model is deprecated, a price moves, latency degrades in your region, a legal team rejects a data-processing term, or a stronger open-weight model appears. The team discovers that "switching providers" is a quarter-long migration project, not a configuration change.…