Lexa

Translating Regulation into Executable Systems

Designing an AI-assisted system to transform how complex tax rules are authored, validated, and deployed at scale.

Lexa Hero

IMPACT

RuleGen transformed tax rule authoring from a manual, knowledge-heavy process into a human-in-the-loop, AI-assisted workflow designed for scale and compliance.

~80 hours → <5 hours

Per rule file over time

80%+ Accuracy

Target, exceeding industry LLM baselines

Reduced Dependency

On institutional knowledge which encompasses 90% of Rule authoring

Lower Risk

Operational and compliance risk reduced through structured review

This work established a foundation for onboarding 40% of rule files by end of 2026, enabling future AI-assisted compliance capabilities.

TIMELINE, ROLE & SCOPE

Timeline

Sept–Dec 2025 - Led design and launch of Lexa, an AI-assisted rule authoring experience, delivering MLP under a 3.5-month timeline for compliance-critical workflows.

Role

Lead UX Designer
Experience Strategy · Human-AI Interaction · UX/UI · Research · Branding

Scope & Responsibility

Defined the end-to-end authoring experience, balancing AI automation with human oversight to enable a scalable, trustworthy, human-in-the-loop compliance system.

Defining the Rule Authoring Experience Strategy

  • Led user research to extract mental models of tax experts authoring, testing, and deploying rules across jurisdictions.
  • Designed a purpose-built, agentic AI interaction model beyond existing chat templates to support compliance-critical review.
  • Partnered with PMs and engineers to define safe human-in-the-loop workflows, guardrails, and failure handling.
  • Advocated for live UAT observation, uncovering usability gaps missed by asynchronous testing and driving key experience changes.
  • Led product branding (name, icon, visual direction) to support future expansion beyond the tax organization.

Case Study — Deep Dive

Problem

Amazon operates in hundreds of tax jurisdictions worldwide, each governed by unique and frequently changing legislation. Translating these regulations into executable tax rules is a high-risk, manual process.

Today, tax experts must:

  • Read complex legislation documents
  • Review Tax Requirements Documents (TRDs) and Business Requirements Documents (BRDs)
  • Analyze existing rule files to identify impacted logic
  • Decide whether to create new rules or modify existing ones
  • Ensure dependencies across rule files remain intact

This work is time-consuming, error-prone, and heavily reliant on institutional knowledge. In some regions, a single regulatory update required 100+ hours of combined PM and authoring effort, slowing production deployments and increasing compliance risk.

Why This Matters

As Amazon continues to expand globally, regulatory change is no longer an exception, it is a constant.

Manual rule authoring does not scale with:

  • The volume of legislative updates
  • The growing complexity of interdependent rule systems
  • The need for faster, safer deployments

Delays or errors in tax rules can lead to incorrect tax calculation, compliance exposure, and customer trust impact. Improving this workflow is not just about efficiency, it directly affects business continuity and regulatory confidence.

Opportunity

Advances in large language models created an opportunity to re-imagine rule authoring.

Instead of experts...

  • • Manually interpreting legislation
  • • Searching for impacted rules
  • • Encoding logic from scratch

We shift their role to...

  • • Reviewing AI-generated drafts
  • • Resolving conflicts
  • • Applying judgment where it matters

"The goal was not to remove humans from the loop, but to move effort upstream, from manual creation to informed validation."

The Solution

RuleGen is an AI-assisted rule authoring system that helps tax experts translate legislation into structured, executable rules with confidence.

Using a multi-agent AI architecture, RuleGen:

  • Analyzes uploaded legislation and regulatory inputs
  • Extracts decision logic and rule conditions
  • Identifies relevant existing rules and dependencies
  • Generates draft rules in a standardized internal format
  • Surfaces conflicts and areas requiring human judgment

All generated rules remain editable, reviewable, and never auto-deployed, ensuring governance and control remain with the author.

What We Built

Conversational UX

Tax experts upload legislation and interact via natural language, using a familiar chat paradigm that reduces training time.

Explainability & Trust

Legislative summaries and generated rules are traceable to source logic, building confidence in AI outputs.

Human-in-the-Loop

AI acts as an assistant, ensuring all rules remain editable and require explicit human approval before deployment.

Design Challenge We Solved

Our initial concept placed the chat interface side-by-side with the existing rule authoring UI.

The Pivot

During design reviews and early feedback, we discovered this approach created serious usability issues:

  • Cramped layouts on laptops
  • Reduced readability of generated rules
  • Increased cognitive load during review
  • Higher risk of oversight in compliance-critical tasks

We pivoted to a full-screen chat experience, allowing users to:

  • Focus entirely on analysis and review
  • Toggle back to the rule authoring system only when needed

This change significantly improved clarity and reduced error risk—critical in regulated workflows where accuracy matters more than speed.

Design

Lexa Design Screen 1
Lexa Design Screen 2
Lexa Design Screen 3
Lexa Design Screen 4
Lexa Design Screen 5

Re-imagined UX (Q1 2026)

The next phase of RuleGen focuses on depth, memory, and auditability.

Planned enhancements include:

  • Chat history with timestamps for reference and auditing
  • Smart suggestions for rule placement within files
  • Visual indicators distinguishing AI-generated vs human-authored rules
  • Context memory across conversations to support longer workflows

This re-imagined UX was reviewed with business stakeholders to align on direction and investment priorities.

North Star Vision

RuleGen is part of a broader roadmap toward an AI-powered tax compliance assistant.

The long-term vision includes:

  • Proactive monitoring of legislative changes
  • Automated impact analysis across rule systems
  • Validation of rules against regulatory requirements
  • Auto-generated compliance and audit reports

This roadmap guides product decisions through 2027, evolving RuleGen from a rule authoring assistant into a comprehensive compliance intelligence platform.

Testing & Feedback

We are actively testing RuleGen with expert tax users.

Early Feedback

  • Strong comfort with chat interface
  • Requests for confidence indicators
  • Need for clearer error handling
  • Desire for quick reference to existing rule data

Our Response

  • Confidence indicators planned for early 2026
  • Refining error messaging based on real inputs
  • Exploring lightweight context panel for Q1 2026

What This Work Established

Lexa established a scalable operating model for regulatory rule authoring, one where AI initially operates with human oversight, but is designed to progress toward autonomous execution as confidence, accuracy, and safeguards mature. The work reframed rule creation from a manual translation task into a system capable of automated interpretation, validation, and controlled deployment.

More than an MLP launch, Lexa laid the foundation for AI-assisted and eventually AI-driven, compliance workflows by proving that transparency, traceability, and progressive autonomy are prerequisites for safely removing humans from the loop, not barriers to it.