Design Leader

Designing Verification Frameworks for Trusted AI

 

Defining Verification Frameworks for AI Responses

I led the design of the system that defines how AI responses are verified and explained across Salesforce’s Lightning Platform.

This system standardizes how AI outputs are presented, including citations and source attribution, making it foundational across product lines and enabling teams to adopt it at scale.

It establishes a consistent interaction model for AI, ensuring users can understand where answers come from and how reliable they are.

Role

  • Defined the interaction model for agentic experiences across platform and product surfaces

  • Established system-level patterns in a zero-PRD, highly ambiguous space

  • Drove cross-org decisions across Design, Eng, and PM to align on a unified approach

  • Shaped feasibility and architecture in partnership with engineering (not just design)

  • Translated strategy into reusable, production-aligned patterns adopted across teams

  • Set quality bar, governance, and rollout strategy for scalable adoption

  • Managed stakeholders and expectations at platform + product levels

 
 

 

Challenges

Existing approaches to AI citations broke down at both system and experience levels.

Scaling

  • Built for a single platform (Lightning), citations did not scale to multi-cloud AI

  • No shared interaction contract across AI products

  • Teams created fragmented, product-specific solutions

  • Legal requirements for source transparency lacked a system-level solution

Experience

  • Inline citations disrupted the natural reading flow

  • High citation density reduced comprehension

  • Expanding source lists pushed primary content off-screen

  • Low engagement with full source views revealed a breakdown in usability

 

 

Strategies

1. Moved citations out of inline text

Inline citations broke reading flow at scale.
→ Shifted to structured surfaces (popover/panel)
→ Tradeoff: less immediate visibility, better readability

2. Designed patterns for multiple surfaces

AI appears in panels, full pages, and embedded views.
→ Created patterns that adapt across contexts
→ Result: reusable system across products

 

3. Introduced citation markers

Connected specific content to its source
→ Result: clearer traceability of AI responses

4. Enabled flexible implementation

Teams had different technical constraints
→ Introduced optional panel-based source views
→ Result: easier adoption across teams

 

User Flow: progressive trust verification system

 
 

 

Impact

  • Scaled across 14 product lines, replacing fragmented, one-off implementations

  • Established a unified approach to how AI responses are handled across products

  • Made source transparency a core part of AI interaction design

  • Accelerated delivery of AI experiences through reusable interaction patterns

 
 
 

Design

Interaction

 

Contextual View (Popover) Triggered by clicking an in-text citation.

Global View (Local Panel) Triggered by the footer/main action button.

Global View (Agentforce Panel) Triggered by the footer/main action button.

 
 

AI Search Experience

Agentforce Panel Experience

Before and after (AI Card)

The new design improves readability by removing inline citation disruption and shifting source access to a popover. This also eliminates layout shifts—keeping the AI card height stable and preventing content from being pushed down on interaction.

Before

After