ElancoGPT used OpenAI GPT-3.5 + Google PaLM 2 behind a federated retrieval layer. The interesting design problem wasn't the model — it was the output shape. Every response renders as four typed blocks: fact, inference, risk, action. Each block is independently citable, copyable, and attachable to downstream artifacts.
The journey shows what happens between a salesperson asking a question and a customer receiving a structured answer. Each lane has a specific commitment.
A field salesperson types: "What's the dosing protocol for Galliprant in senior dogs, and what's the LFT risk?" No prompt engineering. No keyword formatting. The interface meets them where they already think.
Ranked sources surface from FDA labels, internal clinical guidance, and approved customer-facing materials. Provenance is mandatory. A query without source-grounding never reaches the rendering layer.
Fact / inference / risk / action. Each block has a confidence indicator, each is separately citable, each is copyable to a downstream artifact. The AI never returns "an answer." It returns a structured argument.
Low-confidence outputs surface a warning chip and route to a vet-affairs audit queue. 100% of responses ship with provenance by design. Salespeople can defend any claim they pass to a customer.
The Action block becomes the body of a customer email. The Fact block becomes a regulatory citation. The Risk block goes into the clinical-decision log. Structure travels with the content all the way to the customer.
Every block has provenance. Every block is independently exportable. The user controls which blocks travel to the customer.
A chatbot that returns paragraphs is the lowest-value AI surface in an enterprise context. Fact / inference / risk / action made the AI a decision surface, not a reading surface.
Vet-affairs audit ran on a spreadsheet for the first six months. A purpose-built queue could have surfaced corpus gaps faster.