You're Making the Same Mistake

---

I was in a Melbourne warehouse twelve months ago with a retailer spending $100K+ monthly on digital. They wanted to know why investment was not driving a strong ROI. I asked for their customer data to build lookalike audiences. “We’ve got orders. We don’t have data.” A decade of transactions. Zero usable intelligence. I’ve seen this pattern a thousand times. Organisations invest in technology without building the data foundation. Then wonder why nothing works. ADAPT’s 2025 State of Data & AI in Australia report surveyed 450+ CDAOs and CIOs. The finding: Australian enterprises invest an average of $28M annually in AI, yet 72% report failing to achieve measurable ROI. So I called Samir Ghoudrani, AI Director, PWC to confirm my thesis: the AI failure mode is identical to the digital failures I’ve been watching for years. Turns out I was right. And the problem is so obvious, organisations don’t even know its missing.

1. You Have a Context Problem, Not a Technology Problem

AI models are reasoning engines, not magic boxes. That Melbourne retailer? They thought buying better ad tech would fix their problem. It didn’t. Because the technology was fine. The context was broken. Samir calls this the bottleneck. Most organisations are feeding AI raw data - transaction logs, customer IDs, timestamps, and expecting brilliant insights. Think about it, what if I asked you ‘Can you fix my problem, but I will not tell you anything about my business?’ This is what we are doing with AI. The gap between your data and usable AI isn’t technical. It’s something Samir calls “knowledge engineering.” Not data engineering, we’ve had that for 20 years. Knowledge engineering is turning raw data into context AI can actually reason with. When I deployed NotebookLM earlier this year, to replace a 20-year employee, I made this exact mistake. Dumped everything, emails, files, two decades of work, into the system. Week one was chaos. Wrong answers. Week three, I rebuilt it properly. Categories. Structure. Context. Result: the replacement hire, her onboarding six weeks to capability instead of six months.

•   Week 1 vs Week 3: Same AI, different results. The knowledge layer made the difference.*
  

---

            *“Decreased speed to capability from 6 months to 6 weeks, here’s how”*

---

2. You’re Drowning in Data, But Starving for Knowledge

Samir pointed out something I’ve been seeing for a decade: companies have too much data and not enough at the same time. Open any data warehouse. Millions of transactions, billions of logs. Ask your BI team to extract insight? Three weeks and two dashboards later, you’ve got nothing actionable. Now ask a new executive how long it took to understand your business. Three months minimum. What helped? Coffee chats. Shadowing people. Conversations in hallways. Not your data warehouse. Not your documentation. Your most valuable intelligence lives in conversations, not databases. Knowledge engineering is the work of capturing what people know, the context, the relationships, the stories behind the numbers, and making it usable for AI. That Melbourne retailer had every transaction for 10 years. But they couldn’t tell me which customers were most valuable, what attributes made someone likely to buy, or which campaigns actually drove revenue. The data existed. The knowledge didn’t.

Under Australia’s emerging AI governance standards, including the Voluntary AI Safety Standard and the federal Guidance for AI Adoption, enterprises are increasingly expected to document how their AI systems reach conclusions, especially in higher‑risk use cases. You can’t just point at a data warehouse; you need clear knowledge layers that explain how decisions are made.”

---

               *“Most organisations don’t realise they’re missing the knowledge layer...”*

---

3. You’re Using AI for Memory, Not Reasoning

I see this in nine out of ten AI implementations I audit. They treat AI like Google. Ask it to recall facts. Retrieve information. Find documents. That’s using a reasoning engine for memory. Massive waste. Samir’s right: the power isn’t in what AI remembers from training. It’s in what it can reason about with YOUR context. Give it rich knowledge, customer relationships, product nuances, internal processes, decision history, and it stops being a search engine. It becomes something that can actually think about your business. I saw this with NotebookLM. Week one, I was asking it to find files. Week three, I was asking it to solve problems using everything that 20-year employee knew. Different game entirely. Stop asking AI to guess, Start giving it your knowledge and asking it to reason.

4. Your AI Needs Onboarding, Just Like a New Hire

Samir’s analogy nailed it: think of AI as the smartest employee you’ll ever hire. Now imagine that employee shows up day one. No handbook. No training. No context about your business, your customers, your history. Useless, right? That’s what most organisations are doing with AI. Brilliant technology. Zero onboarding. The knowledge layer is your AI’s onboarding. It’s how you teach the system what matters, how things connect, why certain decisions were made. You need a four-person task force for knowledge engineering: someone technical, someone who knows the business, someone AI-literate, and an executive sponsor who’ll protect the work.

5. You’re Asking the Wrong Question About Failure

When AI doesn’t work, executives blame the model. “ChatGPT is bad.” “This tool isn’t right for us.” “We need a different vendor.” Wrong question. Samir’s right: the question isn’t “Why is the AI failing?” The question is “What context have I failed to provide?” That Melbourne retailer blamed their platform . The technology was fine. They’d given it nothing to work with. I’ve watched this pattern for 10 years in digital strategy. Now I’m watching it with AI. Organisations invest in capability without building foundation. Then wonder why ROI never shows up. Stop chasing better models. Start building better knowledge.

What This Means for You

As Samir mentioned, AI models improve every three months. Your window to catch up is shrinking. The competitive advantage won’t go to whoever has the best AI. It’ll go to whoever has the best knowledge layer feeding AI.

48 Hour Action

• Pick one painful workflow. Customer support tickets. Sales qualification. Compliance reviews. Something costing actual money. Look at your P&L.
• Don’t touch AI yet. Assign a team.
• Map what data exists. Customer history. Product catalog. Past decisions. Write it down.
• Ask what’s missing. Context in people’s heads. Relationships between data points. Stories behind numbers.
• Assemble a four-person task force:
• Technical lead (data engineer)
• Business expert (knows the workflows)
• AI specialist (knows the models)
• Executive sponsor (you, to protect the priority)
• Build one knowledge layer for one workflow. Turn data into story. Most organisations will skip this and throw ChatGPT Enterprise at their data warehouse.

---

Before you build one more AI workflow, answer this: Do you have a knowledge layer, or just a data warehouse?

Take the 5-minute AI Readiness Scorecard. It’ll show you exactly where your bottleneck is, Strategy, Data, People, Governance, or ROI. Most organisations discover it’s not the one they expected.

→ AI Readiness Scorecard

---

AI works. It’s whether you’re giving it anything meaningful to work with. Note: The views and opinions expressed in this post and podcast are Samir’s personal views and do not necessarily reflect the official policy or position of PwC.

Ramon

This Week’s Episode Full conversation with Samir Ghoudrani (PWC): Why 72% of Australian AI investments fail at the knowledge layer, the four-person task force approach, and what happens when you treat AI like memory instead of reasoning. Spotify Podcast Apple Podcast