Executive Briefing: How Microsoft, Bayer, and Discovery Use AI on the Data You Can’t Upload

Source: Nate’s Newsletter Date: 2026-07-19 URL: https://natesnewsletter.substack.com/p/run-ai-offline-private-files

Summary

Nate’s Newsletter profiles three enterprises running AI on data that can’t leave the building: Bayer fine-tuned a Microsoft Phi model on proprietary crop-protection and regulatory data (cutting agronomic Q&A from weeks to under 30 seconds), Discovery Bank fine-tuned Azure OpenAI 4o-mini/4.1-mini variants on bank-specific terminology and SQL (5–6s down to 1.5–2s response times), and Microsoft reiterates that customer prompts/fine-tunes aren’t used to improve the base foundation model without permission.

Implications

• Local models thread: concrete enterprise evidence that small fine-tuned models (Phi-class, 4o-mini-class) beat generic frontier models on narrow, sensitive-data tasks — reinforces the local-first, fine-tune-not-prompt pattern this landscape favors.
• AI economy/business thread: these are named production deployments with measured latency wins, not pilots — a signal that the “data too sensitive to upload” objection is being resolved by on-prem/private fine-tuning rather than by convincing enterprises to trust cloud providers more.