AI Negotiation Assistant
aPriori helps manufacturing companies simulate production costs. However, sourcing buyers couldn't turn that data into negotiation leverage; I designed the company's first AI assistant to bridge the gap.
My Role
Sole Designer
Timeline
2025 - 2026
Platform
Web app
Team
Designer, PM (2), Engineer (4)
The Problem
Most buyers lacked the expertise to translate aPriori's simulation data into negotiation leverage. They relied heavily on aPriori's internal expert team to bridge that gap.
The Opportunity
Encode that expert knowledge into an AI assistant that delivers actionable negotiation briefs directly to buyers, reducing reliance on specialist support.
My Approach
Microsoft Copilot UX Guideline
This was aPriori's first AI feature. Before designing anything, I used ChatGPT Deep Research to survey emerging AI interaction patterns across the industry. Then pressure-tested findings against primary sources like the Microsoft Copilot UX Guidelines and direct competitor analysis.
Key Insight:
Conversational AI fits open-ended, exploratory tasks where users need a sounding board. Negotiation strategy is inherently ambiguous. That ruled out embedded insight cards or a static dashboard pattern.
The harder problem turned out to be real estate, not conceptual. The product was already visually dense, and adding a chat panel broke the existing UI.
V1: broken UI
My first solution was to fix the panel width to 350px for the smallest supported viewport. This quickly fell apart when stakeholder feedback revealed buyers primarily work on external monitors, not laptops, which made narrow panel insignificant on bigger screens.
350px
V2: fixed-width AI panel
1/3
2/3
V3: responsive panel UI
I redesigned the panel as a responsive flex element: AI takes 1/3 of the screen, main content takes 2/3, with an 800px maximum consistent with industry-standard AI panel sizing. Existing side panels auto-collapse when AI opens, keeping the default state clean.
I built an interactive prototype in Figma Make to communicate the responsive behavior and collapse states to engineers directly, without writing a spec for every edge case.
Status and What's Next
The feature is in beta. Early signals from the expert services team suggest the briefs are directionally accurate, though precision varies by part complexity.
Final UI
The constraint revealed a deeper issue: the product shell wasn't built to accommodate a persistent AI surface. I've drafted a recommendation to restructure secondary panels as header-triggered overlays, freeing up real estate for AI to live more naturally in the product long-term.
Possible Future Enhancement