MBA Consulting Project · TY2025

An AI-powered tax strategy agent
for high-net-worth clients

Built on n8n, Pinecone, and OpenAI. Retrieves personalized tax strategies from a 111-strategy vector database and delivers advisor-grade recommendations in seconds.

111

Strategies Vectorized

5

HNW Client Personas

2

n8n Workflows

5

Top Strategy Recommendations

System Architecture

Three-Workflow Pipeline

Each workflow handles a distinct stage — from ingesting tax return data to retrieving strategies to generating advisor-grade output.

01 — INGEST

Process Tax Returns

Reads TY2025 tax return JSON for each persona from Google Drive and structures the data for strategy matching.

n8n Google Drive JSON

02 — RETRIEVE

Vector Strategy Search

Embeds the client profile and queries Pinecone to retrieve the 20 most relevant tax strategies from the 110-strategy playbook.

Pinecone OpenAI Embeddings RAG

03 — SYNTHESIZE

AI Strategy Output

OpenAI GPT-4o synthesizes the retrieved strategies into five prioritized recommendations with IRC citations, estimated savings, and implementation guidance.

Claude n8n AI Agent IRC Citations

Client Profiles

Five HNW Personas

Each persona represents a distinct high-net-worth archetype with simulated TY2025 tax return data.

P01

Daniel Mercer

Pre-Exit Tech Founder

AUM$15M

FilingMarried Filing Jointly

ProfileAge 44

P02

Priya Shah

Public Company Executive

AUM$13M

FilingSingle

ProfileAge 51

P03

Miguel Alvarez

Real Estate Operator

AUM$17M

FilingMarried Filing Jointly

ProfileAge 39

P04

Eleanor & Robert Whitman

Multigenerational Wealth Family

AUM$18M

FilingMarried Filing Jointly

ProfileAges 67 & 69

P05

Aisha Nazim

Mobile Consultant & S-Corp Owner

AUM$12M

FilingSingle

ProfileAge 34

How It Works

User Flow

From workflow design to real-time strategy — a walkthrough of the prototype in three steps.

Technical Design

System Architecture

The system is built on n8n workflows connected to Pinecone for vector retrieval and OpenAI GPT-4o for language generation. Persona data and tax returns are fetched directly from Google Drive. A stateful session layer tracks each conversation across multiple turns.

01

Workflow One

Strategy Playbook Vectorization

Reads the 111-strategy Master Tax Strategy Playbook from Google Drive, generates OpenAI embeddings for each strategy, and upserts them into Pinecone with rich metadata — IRC codes, client archetypes, risk scores, and estimated savings ranges. Run once per playbook update.

Key Nodes

Google Drive — download playbook Excel

Extract from File — parse strategy rows

Code — structure pageContent + metadata

OpenAI Embeddings — text-embedding-3-small (1536d)

Pinecone — upsert 111 strategy vectors

02

Workflow Two — Decommissioned

Process Tax Returns

Originally built to index persona tax returns into Pinecone for RAG-based retrieval. Decommissioned after testing confirmed that direct Google Drive lookup by persona ID was more reliable and less prone to hallucination than vector search over structured JSON data.

Why it was removed

Tax returns are structured JSON — not free text

Vector similarity search added noise, not signal

Direct GDrive lookup by persona ID is faster and exact

Workflow 3 now fetches returns directly on demand

03

Workflow Three

Analyze & Output Strategy

The live conversational agent. Receives user messages via webhook, classifies intent with GPT-4o-mini, fetches persona data and tax returns from Google Drive, queries Pinecone for the 20 most relevant strategies, assembles a full client context prompt, and returns five ranked recommendations via OpenAI GPT-4o.

Key Nodes

Webhook — receive sessionId + message

Intent Classifier (GPT-4o-mini) — SELECT / CONFIRM / FOLLOW_UP / UNKNOWN

Pinecone Query — top 20 strategy vectors

OpenAI GPT-4o — synthesize 5 ranked recommendations

Session State — track conversation across turns

Workflow One — Node Map

Strategy Playbook Vectorization

Workflow Three — Node Map

Analyze & Output Strategy

Technology Stack

Built With

n8n

Workflow orchestration & webhook routing

Pinecone

Vector database for strategy retrieval

OpenAI GPT-4o

LLM for strategy synthesis & follow-up

OpenAI Embeddings

Semantic search for strategy retrieval

Claude

Front-end development & n8n workflow implementation support

Client Profiles

Five HNW Client Personas

Each persona represents a distinct high-net-worth archetype. Select a client to explore their profile, life events, and AI-identified tax strategy triggers.

Arrows navigate between clients · Tabs switch between events and strategies

Interactive Prototype

AI Tax Strategy Agent · Live Demo

A live conversation with the n8n-powered tax strategy agent. Connect, select a client from the top panel, then ask any tax strategy question.

Select Client

Connect above, then select a client from the sidebar to begin

Thinking...

Session: not started

Market Context

Market Analysis

The U.S. accounting industry is at a structural inflection point — talent shortages, AI acceleration, and shifting client expectations are converging to create a narrow but significant window for innovation.

Industry Overview

The Accounting Industry Faces a Definitive Inflection Point

75% of CPAs are set to retire within 10 years. PE capital has flooded in — $50B+ since 2019 — validating accounting as an institutional-grade asset class. AI is no longer optional; it is the unlock that separates firms that exit at 6–7× EBITDA from those stuck at compliance margins.

$157B

U.S. CPA Market

Total revenue at today's prices

6–7×

Exit EBITDA Multiple

Potential with AI-lifted margins

75%

CPAs Retiring

Within the next 10–15 years

$50B+

PE Capital Deployed

Into CPA firms since 2019

01 — Market Sizing

TAM / SAM / SOM

TAM

Total Addressable Market

$145.5B

Overall U.S. accounting services — audit & assurance, tax services, client advisory, and bookkeeping & payroll.

SAM

Serviceable Available Market

$51B

Accounting services targeting high-net-worth individuals (HNWIs) and SMEs — the segments with the highest advisory complexity and willingness to pay.

SOM

Serviceable Obtainable Market

$85M

0.1% market share under a Virtual Family Office model delivering AI-powered tax strategy at scale. Represents near-term ARR target.

Market projected to reach $180.3B by 2030, with Client Advisory (CAS) growing fastest at ~12.3% CAGR as firms shift from compliance-driven to relationship-driven revenue models.

Market Share 2025

Total $145.5B

Audit & Assurance

$55.3B38%+1.8%

Tax Services

$40.7B28%+2.0%

Client Advisory

$27.6B19%+12.3%

Bookkeeping

$21.9B15%+3.0%

Market Share 2030

Total $180.3B · projected

Audit & Assurance

$60.5B34%Stable

Tax Services

$45.0B25%Stable

Client Advisory

$49.4B27%↑ 8pts

Bookkeeping

$25.4B14%Shrinking

$10.9B → $68.8B

AI in Accounting Market — growing from $10.9B (2025) to $68.8B (2030) at a 44.6% CAGR. Source: Mordor Intelligence

02 — Market Trends

Tailwinds & Drivers

Talent Cliff

75% of CPAs retiring within 10–15 years

AI Adoption Accelerating

61% view AI as opportunity · 80% of returns AI-assisted

Private Equity Validation

$50B+ PE deployed · 1.0× vs. 1.9× revenue multiples by market

Margin Expansion Potential

30% → 45%+ margins with AI · 6–7× EBITDA exit potential

Shift to Subscription Advisory

70%+ recurring revenue → premium valuation multiples

Process Efficiency Gains

50%+ reduction in complex analysis time

03 — AI Impact

Five Developments Reshaping the Industry

Advanced

Bookkeeping & Automation

End-to-end workflows completed autonomously · anomalies flagged in real time

Near-Full Auto

Tax Preparation & Compliance

Prep time: hours → minutes · human review at final stage only

Slow but Strategic

Audit & Risk Analysis

Regulatory complexity slows pace · professional judgment remains essential

Full Agentic

Workflow & Efficiency AI

Email triage, document routing, scheduling · AI as invisible operational backbone

Next Frontier

Advisory Services

Real-time scenario planning · blends machine precision with human judgment

AI Readiness Checklist

Data hygieneClean, structured, well-tagged records

Systems interopAcross practice management tools

Change managementLeadership, staff buy-in, quick wins

Platform trustAuditability + client data controls

Key Indicators 2024–2025

AI-assisted tax returns80%

Analysis time reduction50%+

Agentic client workflowsAccelerating

PE due diligenceAI readiness now standard

04 — Competitive Landscape

Key Players & Gaps

Automation layer

AI intelligence layer

Wealth aggregation

Hover a pill for details · Secondary stage shown in ( )

01

Data Input

Receipts, invoices, bank feeds, payroll

Vic.ai(+Cat)

BILL(+Cat)

Tipalti(+Com)

Ramp(+Cat)

Expensify(+Cat)

Dext(+Cat)

FreshBooks

Masttro(+Rep)

02

Categorization

Journal entries, reconciliation, ledger

QuickBooks(+Rep, +Plan)

Xero(+Plan)

Zoho Books

Botkeeper

Eton Solutions(+Rep)

03

Review & Reporting

Audit trails, controls, P&L, cash flow

DataSnipper

BlackLine

AppZen(+Com)

MindBridge

NetSuite(+Com)

Addepar(+Plan)

04

Planning & Strategy

Tax strategy, forecasting, cash optimization

Holistiplan(+Com)

Hive Tax AI(+Com)

Zocks

april(+Com)

Hazel AI

05

Compliance

Lease, revenue recognition, tax docs, audit support

Sage Intacct(+Rep)

Sage for Accountants

Trullion(+Rev)

Luminary(+Plan)

Gap

HNW Tax Advisory

The white space no existing tool occupies

No tool delivers agentic, multi-turn tax strategy purpose-built for HNW complexity — multi-entity structures, cross-state income, estate planning, and coordinated IRC strategy — at advisor scale. Stage 4 tools address individual return analysis or general planning; none operate at the intersection of HNW client profiling, vector-retrieved strategy intelligence, and conversational advisory dialogue.

↳ This prototype demonstrates exactly this layer

With Thanks

Acknowledgments

This prototype is a supplement to my MBA elective — Analytics Consulting Lab — for the consulting project titled 'A Business Case for an AI-Powered CPA Practice.' The consulting project produced the market analysis presented here. It was a collaborative effort.
My sincere thanks to the teammates who contributed to the research, analysis, and thinking behind it — and to the mentors whose guidance shaped the approach.

Teammates

Apeksha Gupta

linkedin.com/in/apekshagupta90

Joe Christopherson

linkedin.com/in/joe-christopherson

Anabel Roza

linkedin.com/in/anabel-roza

Mentors

Prof. Russell Walker

University of Washington

linkedin.com/in/russellwalker

David Sokolovsky

Managing Director & Partner, The Learner Group (HighTower)

linkedin.com/in/davidsokolovsky

Project Background

Product Discovery

How I identified the problem, designed the prototype, and validated the approach through iterative testing.

Phase 01 — Problem Identification

Identifying the Gap

Who is this for

Tax strategy advisor in an AI-enabled Virtual Family Office (VFO) — financial advisor or CPA — serving HNW clients. In-house tool; sensitive client data stays within the VFO.

The pain point

Identifying tax saving opportunities requires reviewing prior returns, referencing tax code, and synthesizing across documents — all within fixed billable hours. Scale is structurally capped. Goal: auto-surface opportunities for advisor review before client delivery.

How it's done today — and why it fails

Advisors manually review prior-year returns. Clients disengage between filings, fragmenting tax and financial planning. Peak client engagement — right before the filing deadline — is when advisors are most overloaded. A real-time strategy tool at that moment could sustain year-round engagement.

Phase 02 — Solution Design

Designing the Prototype

What was built

A proof-of-concept conversational web application for tax strategy advisors serving HNW clients ($15M+ AUM). The advisor selects a client from five pre-loaded personas; the system retrieves the client's simulated TY2025 tax return from Google Drive and generates personalized strategy recommendations. Built as a supplement to the Analytics Consulting Lab project "A Business Case for an AI-Powered CPA Practice."

The primary job

Recommend personalized tax-saving strategies curated from the VFO's proprietary strategy playbook — differentiated from generic LLM outputs by grounding every recommendation in firm-specific expertise.

The one feature that matters most

All recommendations must come from the strategy playbook. If a scenario isn't covered, the agent declares so rather than advising independently. Enforced through a self-hosted Retrieval Augmented Generation (RAG) model.

Phase 03 — Validation

Testing & Iteration

Playbook grounding via RAG

Early responses drew from generic LLM knowledge — not the playbook
Fixed by embedding all 111 strategies into Pinecone with structured metadata (category, entity type, IRC code, risk score)
Hard constraint enforced: agent must decline if a scenario isn't covered, rather than advise independently

Session state reliability

n8n static data cleared on every workflow redeploy — discovered mid-testing
Fixed by passing session object through node outputs rather than relying on global static data
Google Sheets identified as production migration path to handle restarts and concurrent users

Intent routing & edge cases

Regex-only persona detection failed for inputs like "show me the first person's return"
Replaced with GPT-4o-mini intent classifier handling digit, word-form, and natural language selection
Five end-to-end tests validated: persona selection, analysis, follow-up, mid-session switch, unknown input

Phase 04 — Current State

What I Built

Live agentic workflow

Two active n8n workflows on self-hosted Hostinger VPS — see Architecture tab for full node detail
Tax returns fetched directly from Google Drive rather than Pinecone — more reliable for structured data
Session state passed through node outputs; Google Sheets identified as production persistence layer
Returns five ranked recommendations with IRC codes and savings estimates per client

Five HNW client personas

Pre-Exit Tech Founder, Public Company Executive, Real Estate Operator, Multigenerational Estate, S-Corp Consultant
Each with simulated TY2025 tax return JSON and structured data maps across up to 12 IRS forms
Life events and strategy triggers curated from the playbook per persona

Infrastructure & design decisions

Self-hosted stack (n8n + Pinecone + OpenAI + Google Drive) for data privacy — client data stays within the VFO
Basic Auth webhook with session-scoped credentials; no persistent user accounts required
Markdown responses rendered client-side via marked.js for structured, readable output

Key Insights

What I Learned

01

Playbook grounding is the core differentiator

Early responses drew from general LLM training, not firm strategy
Hard constraint enforced: decline if scenario isn't in the playbook
Transforms the tool from a chatbot into an auditable advisory system
The constraint is what makes it safe to put in front of a client

02

Structured retrieval beats RAG for known data

Full Pinecone pipeline built and tested for tax return retrieval
Then decommissioned — direct Google Drive lookup proved more reliable
Vector search over structured JSON added noise, not signal
RAG is a tool for unstructured text, not a universal data layer

03

Orchestration over autonomy for high-stakes advice

Every decision is a visible n8n node with inspectable inputs and outputs
Deterministic routing made the system debuggable and auditable
Autonomous agents offer flexibility — but not observability
In compliance-sensitive domains, observability is not optional

An AI-powered tax strategy agentfor high-net-worth clients

Three-Workflow Pipeline

Five HNW Personas

User Flow

System Architecture

Strategy Playbook Vectorization

Analyze & Output Strategy

Built With

Five HNW Client Personas

AI Tax Strategy Agent · Live Demo

Market Analysis

Product Discovery

What I Learned

An AI-powered tax strategy agent
for high-net-worth clients