OVERVIEW
I designed the interaction model for an AI assistant that helps legal and financial teams make bulk changes to structured data through conversation, with guardrails that ensure every automated action is reviewable, bounded, and reversible.
​
The challenge was not adding AI, but defining the rules that make automation safe: how the system interprets intent, when it asks for confirmation, and what happens when something goes wrong.
PROBLEM
Legal and financial professionals manage large, structured datasets spanning dozens of interdependent entities.
Bulk updates such as jurisdictions, attributes, or ownership states required repetitive point-and-click actions across many items, often taking hours and increasing error risk.
​
While AI could compress this work into a single instruction (for example, “Move all US entities to the Cayman Islands”), speed without control undermines trust in regulated environments.
PROBLEM
Legal and financial professionals manage large, structured datasets spanning dozens of interdependent entities. Bulk updates required repetitive actions across many items, often taking hours and increasing error risk.
​
AI could compress this work into a single instruction, but in regulated environments, speed without control undermines trust.
MY ROLE
I defined the end-to-end AI interaction and conversation model governing how intent was interpreted, scoped, previewed, confirmed, executed, and reversed.
​
I partnered with Product and Engineering to establish system prompts and conversational rules that constrained AI behavior in high-risk workflows, shaping how ambiguity, confirmation, refusal, and recovery were handled across the product.
IMPACT
The assistant became a key enterprise sales capability, demonstrating how AI could safely automate high-impact workflows in regulated environments.
​
Bulk edits spanning dozens of entities could be completed through a single conversational command.
Preview, confirmation, and undo patterns were adopted as reusable standards across the product, underpinning subsequent AI features.
IMPACT
-
The interaction model became a core enterprise sales asset, demonstrating AI capabilities to prospective clients and generating pipeline interest.
-
Reduced bulk edits spanning dozens of entities into single conversational commands.
-
Preview, confirmation, and undo patterns adopted as reusable standards across the product, underpinning subsequent AI features.
VALIDATION
I designed targeted feedback surveys to evaluate clarity, correctness, tone calibration, and recovery in failure states. Insights informed refinements to confirmation flows, error handling, and conversational constraints.
SYSTEM-LEVEL EXPLORATION
I evaluated where the AI interaction model could and should not extend across additional modules, defining boundaries for safe AI adoption and informing longer-term product direction.
STRATEGIC PROTOTYPE
I built a rapid prototype connecting the assistant to adjacent workflows, aligning leadership on AI direction and supporting enterprise sales conversations.
STRATEGIC TAKEAWAY
In high-stakes enterprise systems, AI earns trust through structure, not speed. Adoption follows when users can understand, verify, and recover from automated actions.