Case Study — Grant Thornton

AI-Powered
Document Intelligence

How we cut Transfer Pricing document preparation from 2 days to 3–4 hours — without replacing the consultant.

Role

Solution Design & SME Collaboration

Client

Grant Thornton

Domain

Tax & Transfer Pricing

Type

Enterprise AI

AI ProductDocument IntelligenceWorkflow AutomationGenAITax & Consulting
Section 1

The Problem

Every year, Grant Thornton's tax consultants had to rebuild Transfer Pricing Local Files from scratch — for each subsidiary, across multiple entities. The content? Largely the same as last year. The work? Entirely manual.

Consultants were spending up to 2 full days per document — extracting data from prior-year Local Files, Master Files, and Transaction Matrix records, then re-assembling it into a standardized template. Repetitive, time-consuming, and low on strategic value.

They tried Microsoft Copilot to speed things up. It made things slower. Outputs were inconsistent, required heavy correction, and added friction rather than removing it. The team needed something purpose-built — not a general-purpose AI tool.

2 Days

Per Local File Document

100%

Manual Effort

Variable

Output Quality

Section 2

My Role

I was part of a cross-functional delivery team responsible for designing and implementing an AI-driven solution to this problem. My contribution spanned solution design, working closely with tax Subject Matter Experts (SMEs) to translate domain requirements into system logic, and ensuring the workflow aligned with how consultants actually work — not just how the technology works.

Section 3

The Solution

We built an AI-driven workflow that could read and understand prior-year documentation — Local Files, Master Files, Transaction Matrix data, and meeting notes — and automatically pre-populate large sections of the new Local File template.

01
Standardizing Inputs First

Before automation could succeed, we mapped and cleaned the source documents. This alone unlocked consistency that Copilot never had access to.

02
Rules + AI, Not AI Alone

We worked with tax SMEs to define extraction rules and transaction-type logic. The AI handled volume; the rules ensured accuracy.

03
Human-in-the-Loop by Design

Consultants review and approve AI-generated sections before finalizing. This wasn't a workaround — it was a deliberate product decision to build trust and maintain quality control.

Section 4

Results

Time per document

3-4h
vs 2 days

Manual effort automated

70-80%

Output accuracy

~90%

Beyond the numbers: Consultants shifted from document assemblers to reviewers and strategic advisors. That's the real outcome.

Copilot failed.
A purpose-built solution succeeded.

This project is a case study in why general-purpose AI tools often underdeliver in enterprise workflows — and what it actually takes to make GenAI work at scale.

These aren't technical lessons. They're product lessons.

Section 6

What I Took Away

1. Adoption Starts with Trust

If the output isn't reliable, users go back to manual. The 90% accuracy target wasn't arbitrary — it was the threshold where consultants felt confident enough to stop double-checking everything.

2. The Workflow Matters as Much as the Model

The biggest unlock wasn't a better AI — it was standardizing the input documents first.

3. Scope the Automation Honestly

We targeted 70–80% automation intentionally. The remaining 20–30% requires human judgment — and a good product knows where to draw that line.

Section 7

Skills Demonstrated

GenAI Product DesignStakeholder ManagementSME CollaborationWorkflow AutomationDocument IntelligenceHuman-in-the-Loop SystemsEnterprise AIRequirements Translation

Interested in Learning More?

I've delivered multiple enterprise-scale products. Let's discuss how I can drive impact for your organization.

Built with v0