Skip to content
All case studies
EY Han Young Assistant Director, IT PM March 2026–Present

AI-native Tax Platform

Designing a multi-agent tax assistant, agentic workflows, a verifiable retrieval layer, and an Azure serverless runtime for professional tax work.

Primary outcome

Established the product, knowledge, validation, and runtime foundations for an AI-native tax work platform.

The problem

Turn fragmented tax knowledge and tacit expert judgment into a reusable AI work platform without losing traceability or expert control.

Constraints

  • Tax conclusions require source traceability and professional review.
  • Sensitive work context limits what can be exposed outside controlled workflows.
  • Agent behavior, data preparation, validation, and operations must evolve as one lifecycle.

Approach and key decisions

  1. 01

    Decomposed professional work into collaborating agent roles for retrieval, judgment, review, and response.

  2. 02

    Designed a hybrid retrieval layer with normalized sources, metadata, vector search, source lineage, and citation validation.

  3. 03

    Defined risk-tiered autonomy, human sign-off, failure handling, and evaluation criteria as part of the product workflow.

  4. 04

    Separated authentication, APIs, functions, storage, orchestration, observability, and deployment boundaries on an Azure serverless runtime.

Outcomes

  • Built a multi-agent tax assistant structure for work-context search, judgment support, review, and response.
  • Shaped investment tax review, VAT reconciliation, and contract review into agentic workflows with explicit approval and exception paths.
  • Established a verifiable knowledge architecture connecting source lineage, evidence tracking, citation validation, and expert sign-off.

Technology and methods

Azure Serverless Multi-agent Systems Agent Orchestration Hybrid RAG Vector Search Source Lineage Human-in-the-loop

Evidence and disclosure scope

This case describes the portfolio owner’s role and public-safe architecture patterns. Client-confidential data, prompts, evaluation results, and implementation details are intentionally omitted.

See the full resume context