The 100-Page Prompt That Changed Everything: KPMG's TaxBot Revolution
How a Big Four firm turned two weeks of expert work into one day of AI magic - and what it means for enterprise AI
The Scary Beginning
John Munnelly, KPMG's chief digital officer, remembers the exact moment everything changed. Late 2022. ChatGPT drops. His team starts experimenting.
But then came the real shock - a single document on KPMG servers, listing thousands of employees’ credit card numbers - discovered during those early tests. “That absolutely scared the pants off me,” Munnelly admitted.
ChatGPT got blocked instantly. Innovation screeched to a halt.
As a result, KPMG immediately blocked ChatGPT, paused AI experiments, and reassessed their risks. Not long after, a graduate posted a screenshot of the ChatGPT block with a snarky comment about the firm’s “fear of innovation.” Papers ran the story. PR disaster. Munnelly realized the firm could no longer simply hide from AI - they had to master it.
The Platform Strategy That Actually Works
Here's where KPMG got really smart: they didn't just build an agent. They built an ecosystem.
Enter KPMG Workbench, a global, multi-agent platform with retrieval-augmented generation (RAG), agent hosting, and support for models from OpenAI, Microsoft, Google, Anthropic, and Meta.
Why? Because KPMG decided it was wise not to assume that any single vendor would dominate LLMs.
Smart. Very smart.
While everyone else is having religious wars over GPT vs Claude, KPMG built Switzerland. The agent picks the right model for the job. Cost optimization. Quality optimization. Risk mitigation.
This isn't just technical foresight - it's strategic insurance.
By 2023, staff were trained not just in AI usage, but in crafting prompts that actually work.
The 100-Page Prompt That Nobody Expected
Then came the 100-page prompt. Not of code. Not of configuration. Of instructions.
For months, a dedicated team collected partner-written tax advice - often “stored all over the place,” including on personal laptops - and merged it with Australia’s tax code to create the foundation for TaxBot.
Fast-forward to today. KPMG's TaxBot doesn't just work - it transforms businesses. "It does what our team used to do in about two weeks, in a day," Munnelly said at Forrester's APAC summit.
The secret sauce? A 100-page prompt that took months to craft.
Most people write prompts like text messages. KPMG wrote a prompt like a legal brief.
Here's what went into it:
Partner memorandums scattered across laptops (yes, laptops)
Australia's complete tax code
Decades of institutional knowledge
Workflow specifications
Quality checkpoints
Human handoff protocols
The result? A system that takes 4–5 inputs, consults a human for direction, and produces a 25-page draft tax opinion overnight.
Why Size Matters (In Prompt Engineering)
"Why 100 pages?" I hear you asking. "Isn't that overkill?"
Not in enterprise. Here's why:
Enterprise isn't about clever shortcuts - it's about bulletproof processes.
When you're advising on billion-dollar mergers, "close enough" doesn't cut it. That 100-page prompt isn't bloat - it's a specification. Every edge case. Every regulatory nuance. Every quality gate.
Think of it as the difference between a food truck and a Michelin-starred restaurant. Both make food. One has a laminated menu. The other has a 200-page operations manual.
KPMG chose the Michelin approach.
The Human Element Nobody Talks About
And the people? TaxBot didn't replace anyone. Staff surveys suggest employee satisfaction has risen as AI frees them to spend more time working on challenging tasks.
Turns out, tax professionals don't want to spend weeks compiling boilerplate advice. They want to solve complex problems. They want to be strategic advisors, not document assemblers.
"They just don't want to do the boring stuff," Munnelly said. "They want to get out there and help clients with chewy problems."
The 25-page draft? That's not the end product. That's the starting point. The human expert takes that foundation and builds something truly valuable.
The Revenue Plot Twist
Here's the part that'll make your CFO sit up straight: clients asked to buy the agents. "We have additional revenue streams that we didn't expect."
Think about it. KPMG built TaxBot for internal efficiency. Clients saw the speed and quality. Said "we want that."
Suddenly, KPMG isn't just selling advice - they're selling the machine that makes advice.
That's not cost reduction. That's business model transformation.
The Enterprise AI Playbook Emerges
KPMG's journey reveals the blueprint for enterprise AI success:
Phase 1: Containment & Learning
Block the public tools. Build private sandboxes. Learn the risks before you scale.
Phase 2: Infrastructure & Platform
Don't build apps. Build platforms. Don't pick winners. Build bridges to all of them.
Phase 3: Knowledge Archaeology
Find the expertise buried on laptops. Digitize the tribal knowledge. Make the invisible visible.
Phase 4: Process Engineering
Write the 100-page prompt. Define the workflows. Engineer the handoffs.
Phase 5: Human Augmentation
Don't replace experts. Amplify them. Free them from drudgery. Let them do expert work.
The Future Is Smaller (Prompts)
Munnelly thinks 100-page prompts probably won't be necessary in future, because KPMG has since built an agent runtime service that allows agents - writers, editors, and managers - to interact.
He's probably right. As agent orchestration matures, we'll move from monolithic prompts to collaborative workflows. Writer agents. Editor agents. Manager agents. Each with specialized, smaller prompts.
But here's the key insight: The 100-page prompt wasn't wasted. It was research.
Every line of that prompt taught KPMG something about their domain. About their processes. About where humans add value and where they don't.
When they move to multi-agent systems, that knowledge doesn't disappear. It gets distributed. Refined. Operationalized.
What This Means for You
Every enterprise, from startups to global firms, can take away these lessons:
1. Security First, Innovation Second
The graduate who mocked KPMG's ChatGPT block? He missed the point. Moving fast and breaking things works in consumer apps. In enterprise, breaking things breaks clients.
2. Platform > Point Solutions
Don't build 47 different AI apps. Build one platform that hosts 47 different agents.
3. Vendor Agnostic = Future Proof
The LLM wars aren't over. Don't pick sides. Pick all sides.
4. Process Documentation = Prompt Engineering
Your best prompts aren't clever tricks. They're crystallized expertise.
5. Human-AI Collaboration > Human Replacement
The goal isn't eliminating experts. It's eliminating the stuff that prevents experts from being experts.
The Real Revolution
TaxBot isn't revolutionary because it writes tax opinions. It's revolutionary because it democratizes expertise at the speed of business.
That M&A deal can't wait two weeks. The market won't pause for your research cycle. But now it doesn't have to.
KPMG didn't just build a better tool. They built a time machine. One that brings next week's deliverable to today.
And that 100-page prompt? It's not documentation. It's a declaration of intent - that accuracy, compliance, and quality don’t get traded for speed.
You can have both. But only if you do the work.