A blue-sky exploration of agentic AI in SAS Viya, shaping how humans collaborate with intelligent systems in enterprise decisioning.
The work defines how agents can be legible, trustworthy, and controllable through end-to-end scenarios and key UX touchpoints—laying the foundation for a cohesive, enterprise-ready agentic AI experience.
SAS® Viya Copilot ↗
Protected Under NDA
The full case study contains confidential product details and is available with a password. Please reach out if you'd like access.
Shaping the foundation of agentic experiences
I was one of four designers on this exploratory project, contributing domain expertise in intelligent decisioning to shape agentic AI in SAS Viya.
We moved from rapid brainstorming and AI-assisted research to defining a focused use case—conducting competitive analysis, mapping user journeys, and developing end-to-end mockups. The final concept was presented to stakeholders, gaining leadership alignment and paving the way for potential implementation.
My contributions included:
- Collaborating cross-functionally in a fast-moving, ambiguous environment.
- Designing a multimodal interaction framework for Viya Copilot.
- Synthesizing explorations into a cohesive set of design foundations.
Working through ambiguity
Every constraint shaped the design process — tight time, unclear scope, and a complex multi-product ecosystem each demanded a distinct response.
Three lenses on how enterprise users relate to AI agency
Workshop & Exploration
Cross-functional workshops with PMs, data scientists, and ML engineers to map the decision-making landscape. Sessions surfaced deep tensions between automation desire and accountability concerns.
AI-Assisted Journey Mapping
Leveraged AI tooling to accelerate synthesis of interview transcripts into annotated journey maps, surfacing recurring friction at handoff moments where human judgment and machine action intersect.
Future-State Experience Models
Speculative experience models depicting how user archetypes interact with agentic AI across short, medium, and long-term horizons — stress-tested in concept evaluation sessions.
From autonomous vision to near-term copilot
Early concepts explored a long-term, autonomous vision — a dashboard-like system where agents proactively monitor and act with minimal user intervention.
Through stakeholder alignment, we shifted toward a more practical MVP: a chat-based interaction within Viya Copilot, enabling step-by-step collaboration while gradually introducing agentic behaviors.
How the interface matured across three iterations
Each version of the Copilot side panel addressed a specific friction point — from raw text output to structured, agent-aware communication that builds user trust incrementally.
The raw text format was fast to prototype and surfaced all the information at once — but user feedback indicated the response felt like a "wall of text." Analysts had trouble quickly extracting the specific section relevant to their current task, and the agent attribution was nearly invisible.
- About 12% of SME borrowers are showing a consistent increase in late payments over the past 90 days.
- The Macro Trigger Monitor detected a slowdown in manufacturing output (–2.3%) and a rise in local unemployment (+1.8%), both linked to higher delinquency risk.
- Confidence score for overlap: 0.87 (high)
- Top contributing features:
- Unusual transfer frequency
- Spike in short-term credit utilization
- Change in repayment cycle behavior
The accordion significantly improved scannability, and the tooltip gave power users access to agent metadata on demand. However, the hyperlink still felt like a UI pattern borrowed from the web rather than a native AI interface component — stakeholders flagged that it didn't clearly communicate "this came from an agent" to less technical users.
The agent token pattern tested significantly better across all user archetypes — including non-technical stakeholders who had previously overlooked the hyperlink entirely. It established a reusable design pattern for communicating AI agent provenance that was adopted into the broader Viya Copilot design system.
What we learned about designing for agentic uncertainty
The project confirmed that agentic AI UX cannot be designed in isolation from the underlying decision logic. The most impactful design decisions were not about the interface surface, but about where to place the boundaries of agent autonomy — and how to make those boundaries visible and adjustable in real time.
Designing for agentic AI is fundamentally a trust design problem. The hardest questions are not about capability — they're about when to show your work, when to ask permission, and when to act.