---
title: "KPMG's 100-Page-Prompt AI Disaster Shows What Not to Do"
description: "5 source-backed AI claims from KPMG's 100-Page-Prompt AI Disaster Shows What Not to Do, with key statistics, context, and the original AI Adopters Club source."
url: "https://kbanc.com/claims-library/kpmg-100-page-prompt-ai-disaster"
source: "https://aiadopters.club/p/kpmgs-100-page-prompt-ai-disaster"
date: "2025-08-30"
topics: ["implementation", "strategy", "business"]
generated: "2026-08-31"
---

# KPMG's 100-Page-Prompt AI Disaster Shows What Not to Do

By Kamil Banc | August 30, 2025

## Claims

1. **KPMG's 100-Page Tax Prompt** (source summary): KPMG built a tax advice agent that uses a 100-page prompt on its Workbench platform.
2. **Dramatic Drafting Speed Gains** (source summary): KPMG's system drafts 25-page client documents in one day instead of two weeks.
3. **CDO Admits Prompt Excess** (source summary): KPMG's CDO admitted mega-prompts probably won't be necessary once their agent runtime matures.
4. **Jagged Frontier Study Finding** (source summary): A Harvard Business School study with BCG found consultants' accuracy dropped 19 percent on complex tasks.
5. **Tax Chatbot Reliability Failures** (source summary): TurboTax and H&R Block's AI chatbots were wrong or useless up to half the time.

## Evidence

### Quote
> "Nothing says 'we've never actually built anything that works' like thinking that more pages equal better AI." - Kamil Banc

### Key Statistics
- **100-page prompt**: The size of the instruction document KPMG built for its tax advice agent, which the author argues reflects a misunderstanding of LLM design.
- **25 pages in one day vs. two weeks**: KPMG's claimed productivity gain for drafting client tax documents using the AI agent.
- **19% accuracy drop**: Decline in consultants' performance on complex tasks when using GPT-4, per a Harvard Business School and BCG study on the 'jagged frontier.'
- **Up to half the time**: Rate at which TurboTax and H&R Block AI chatbots gave wrong or useless answers in Washington Post testing.

## Context
The article's methodology combines reporting on KPMG's disclosed AI implementation with the author's practitioner experience implementing LLMs across hundreds of systems. Kamil Banc applies software engineering principles—modularity, unit testing, version control, and deterministic validation—to argue that monolithic prompts are unmaintainable and error-prone. His recommendations center on retrieval-augmented generation, separating knowledge retrieval from reasoning, and building testable components with validation at every step. For practitioners, the takeaway is to start with focused prompts, build around LLM strengths like pattern recognition and text transformation, and avoid relying on models for math, citations, or factual accuracy without supporting systems. The broader lesson is that AI failures at major consultancies stem from technical competence gaps rather than corporate bureaucracy.

## Source
- Original: [KPMG's 100-Page-Prompt AI Disaster Shows What Not to Do](https://aiadopters.club/p/kpmgs-100-page-prompt-ai-disaster)
- Cite: kbanc.com/claims-library/kpmg-100-page-prompt-ai-disaster

## Primary Evidence
- [KPMG just admitted](https://www.theregister.com/2025/08/20/kpmg_giant_prompt_tax_agent/) (theregister.com; supports claims 1, 2, 3)
- [TurboTax and H&R Block's AI chatbots](https://www.aiaaic.org/aiaaic-repository/ai-algorithmic-and-automation-incidents/turbotax-hr-block-chatbots-provide-inaccurate-tax-advice) (aiaaic.org; supports claim 5)