Elanco Knowledge Systems · Applied AI · 2023 Interactive · Query lifecycle

ElancoGPT. One question, four accountable blocks.

A decision surface, not a chatbot; designed at the infancy of LLM tooling. OpenAI GPT‑3.5 and Google PaLM 2 behind a federated retrieval layer; every response renders as four typed, independently-citable blocks — fact, inference, risk, action — with a human accountable at every beat. Scroll to follow a single Galliprant query from plain-language ask to customer-ready artifact.

  1. 01 Frame
  2. 02 Ask
  3. 03 Render
  4. 04 Verify
  5. 05 Act
Role
Lead Product Designer · AI surface design
Ship
10 days · 3 design / 7 dev · Apr–May 2023
Stack
GPT‑3.5 · PaLM 2 · custom retrieval · EDS
Adoption
71% weekly active · 18 mo · web / Salesforce / Teams
Scroll
ElancoGPT
RL

What's the Galliprant dosing for senior dogs, and what should I confirm before recommending it to a clinic?

FACTGalliprant (grapiprant) is approved for OA pain in dogs ≥ 9 months at 2 mg/kg PO once daily.
FDA label · NADA 141-455
INFERFor a senior-heavy clinic mix, dosing simplicity is the likely conversion lever vs. competitor NSAIDs.
conf 0.84
RISKNot recommended in severe hepatic impairment. Confirm baseline LFTs before substitution.
Clinical guidance · 2023-08
ACTIONSend the clinic-specific dosing card + LFT-baseline reminder template.
→ create email draft
Send a message…➤
01 / Lifecycle · across actorsAsk → retrieve → render → verify → act

The same lifecycle, mapped across actors.

A query is the start of a workflow, not the end of one. The map below shows what happens between a salesperson asking and a customer receiving a structured answer — each lane holds a specific commitment.

Actor
Ask
Retrieve
Render
Verify
Act
Salesperson
"What's the dosing for senior dogs?"
—
Receives fact / infer / risk / action.
Reads provenance citations.
Copies action block into email.
Retrieval layer
—
Federates OpenAI + PaLM + corpus.
Returns ranked sources.
Surfaces ambiguity for human review.
—
Vet affairs
—
—
—
Audits flagged outputs weekly.
Updates corpus when gaps appear.
Customer (clinic)
—
—
—
Receives sourced answer in email.
Forwards or escalates.
Query lifecycle01 / 05
The infancy of LLMs: no patterns yet, so this was built from Uplook / EDS components.
01 · Persona

No UI patterns existed yet, so I built on a system.

This was the infancy of LLM products, with no established interaction patterns to lean on. I set the surface's design language by composing ElancoGPT from Uplook components so it read as native enterprise software from day one. Two carried most of the surface: the Alert, run in a persistent variant to hold the accuracy warning above the thread, and the Text Field with paired Button, restyled into the composer. The corpus was scoped deliberately: retrieval was tuned to the live needs of multiple key functional teams across Elanco (commercial, vet affairs, IT) rather than a generic knowledge base. At the infancy of LLMs, custom prompts were built around roles and personas, each one steering the model to render the information relevant to that user.

Uplook / EDS · parts reused2 of 56 families
EB.01 Alert · persistent
bg-amber-50border-l-amber-600text-sm

Pinned above the thread instead of dismissible, so the accuracy contract never scrolls away.

EB.02 Text Field + Button
bg-whiterounded-mdbg-elanco-600

The stock pairing, restyled into a persistent composer; no new input component to maintain.

Uplook / EDS · Alert (persistent) · Text Field + Button (composer) · corpus tuned to key functional teams

02 · Ask

The query starts in plain language.

A field rep working a senior-heavy clinic types: "What's the Galliprant dosing for senior dogs, and what should I confirm before recommending it?" No prompt engineering, no keyword formatting — the interface meets them where they already think.

Input: natural language · streamed response

03 · Render

The answer returns as four typed blocks.

Fact, inference, risk, action: each block independently citable, copyable, and attachable to a downstream artifact. Rendered in EDS components, so the AI reads like the rest of the enterprise.

Output shape: fact · inference · risk · action

04 · Verify

Accountability is designed into the chrome.

Every block ships with provenance: the FDA label on the fact, a confidence score on the inference, dated clinical guidance on the risk. A persistent warning names the model; every user is accountable for verifying each response before acting on it — ElancoGPT is an internal tool, so the human in the loop is an Elanco employee, not a customer. The Experimental surface names PaLM‑2 in the same chrome — federation stays invisible.

Provenance: 100% · GPT-3.5 / PaLM-2 · banner: persistent

05 · Act

A block becomes a customer-ready artifact.

The action block copies straight into an email draft, with the dosing card and LFT reminder ready to send. Copy messages exports the transcript; Customise wraps a prompt into a shareable persona deeplink, so teams share links, not prompt lore.

Action → email draft · Copy messages → artifact · persona → deeplink

02 / Outcomes18 mo post-launch

Adopted globally. Time-to-answer collapsed.

Weekly active users71%Eligible commercial + vet org
Time-to-answer−68%vs pre-tool baseline (same query class)
Source citation rate100%Every block ships with provenance
Artifacts / week240+Emails · regulatory notes · briefs
03 / Output shapeLive example

What a Galliprant query returns.

FACTGalliprant (grapiprant) is approved for osteoarthritis pain in dogs ≥ 9 months old at 2 mg/kg PO once daily. FDA label · NADA 141-455
INFERFor this customer's clinic mix (mostly senior dogs), dosing simplicity is the most likely conversion lever vs. competitor NSAIDs. conf 0.84
RISKNot recommended in dogs with severe hepatic impairment. Confirm baseline LFTs before recommending substitution. Clinical guidance · 2023-08
ACTIONSend the clinic-specific dosing card + LFT-baseline reminder template attached. → create email draft

Every block has provenance. Every block is independently exportable. The user controls which blocks travel to the customer.

04 / ArtifactsScreens & demos

The product, in the wild.

One high-fidelity mockup, two live model demos, the shipped surfaces, and the responsive state boards; every screen built on EDS components.

04.1 · The mockup

The UX target the build was measured against: Elanco-branded, on-system, accountable by default.

ElancoGPT product screenshots: mobile and desktop chat surfaces. A new IT employee asks about company culture; ElancoGPT streams a structured answer with three numbered tips
The UX High Fidelity Mockup. Elanco-branded, on EDS components: the streaming response, the structured list, the persistent composer.
04.2 · Live demos · two models

The same surface, federated across OpenAI GPT‑3.5 and Google PaLM 2; the model swap is invisible to the user.

OpenAI 3.5 demo. A live query on the phone surface: streaming, structure, the persistent warning.
PaLM 2 demo. The same surface answering through the Google model; federation is invisible to the user.
04.3 · Shipped surfaces

The desktop screens a session actually opens on: default, experimental, and the pages that let a user shape their own. Each shown at full width, on the same EDS chrome.

ElancoGPT default web surface: personas sidebar, persistent accuracy warning, composer
The default surface. Personas rail, persistent accuracy warning, composer, at chat.elanco.ai. The screen every session opens on.
ElancoGPT Experimental surface running PaLM-2 with a labelled warning banner
Experimental, labelled. The same surface running PaLM‑2, the model and its risk named directly in the warning chrome, so federation never hides behind the answer.
Customise page: title and prompt fields generate a custom-persona deeplink
Customise. Title and prompt fields turn any instruction into a named persona with a shareable deeplink, so teams pass links rather than prompt lore.
About page: introduction, value proposition, accountability, customise
About, in product. Value proposition and the accountability contract, every response verified before acting, stated where users actually work.
04.4 · Responsive states

One system across every breakpoint: the full state walk from beta alert to streamed answer.

Mobile state board: beta alert, default, warning, entered prompt, streamed answer, plus light and dark drawer menus
Mobile states. Beta alert → warning → entered prompt → streamed answer, plus both drawer menus.
Tablet state board: navigation with history rail, secondary menu, and prompt across three layouts
Tablet states. Navigation, history rail, and secondary menu across three layouts.
ElancoGPT Experimental on tablet: warning banner, personas, greeting
Tablet, in frame. The Experimental surface at tablet width.
Mobile flow: six states from beta alert to generated answer
Mobile flow. The six-state walk from beta alert to generated answer.
05 / ReflectionDefend / redo
What I'd defend

Structured output, every time.

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.

What I'd do differently

Ship the audit queue with v1.

Vet-affairs audit ran on a spreadsheet for the first six months. A purpose-built queue could have surfaced corpus gaps faster.

06 / Read nextSiblings + parent
Child · 01

Continuous Discovery

The role taxonomy ElancoGPT uses as its output shape.

Read sibling → Child · 02

Elanco Design System

ElancoGPT renders into EDS components. No new visual grammar.

Read sibling →