The only AI filter that matters

The only AI filter that matters - Growth or Costs?

Nine out of ten executives say AI is their number one concern. One in ten feel equipped to do anything about it. That gap isn’t a knowledge problem. It’s a filter problem. Most boards are funding AI initiatives they can’t measure. Most CDOs are defending investments that don’t hit P&L. Most executives are hoping technology solves the problem while their CFO is asking the only question that matters: Does this increase revenue or decrease costs? Everything else is noise.

I sat down this week with John Foong, recent Chief Commercial Officer at Domain. Twelve years at Google, Two years at Uber in global leadership roles. John isn’t an AI specialist - He’s a P&L leader who’s built, sold, and governed AI products at scale. He’s launched AI products that generated millions in revenue and watched others, despite 95% accuracy rates, go nowhere. Three continents. Millions in revenue. He’s seen what works, what doesn’t , and why some die on first contact with reality.

---

The Filter Nobody’s Using

John shared: Whether it’s AI or other technology, the only question that matters is this, does it increase revenue or does it decrease costs? That’s it. Not strategic. Not innovative. Not cool. Revenue or cost. Everything else you’ll hear about AI, efficiency gains, productivity improvements, competitive advantage, is just context. Either it makes money or it doesn’t. Everything else is noise. At Google in the early 2010s, John watched AI models achieve 95% accuracy in radiology scans. Detecting cancer with near-specialist-level precision. It was technically brilliant. It went nowhere. Why? It didn’t displace enough cost. Specialists still had to review every scan. Meanwhile, YouTube’s AI categorisation, initially dismissed as a meme (”the machine can tell if this is a cat video”) eventually displaced millions in manual content moderation costs. One passed the filter. One didn’t.

The Three Assumptions That Kill AI Business Cases

Most pilots die: not in execution, but in the assumptions. Before approving any AI investment, pressure-test three variables. If you get any of them wrong by 30%, your business case collapses.

Price per user - How much will an average user actually pay? Not in theory. - In practice. Most teams overestimate this by 40-50%. They ask “would you pay for this?” Customers say yes. Teams assume that’s the price. It isn’t. Adoption happens at a fraction of the theoretical number. Adoption rate - Realistically, how many users will actually use this? This is where most AI initiatives fail. You build something 95% accurate. You assume 80% of your user base will adopt it. They won’t. Adoption friction kills most AI products. If your AI requires users to change behavior, add a step, or learn a new interface, assume 70% won’t adopt. Time and cost to market - How much will it actually cost to get from beta to enterprise-grade? Teams consistently underestimate this by 2-3x. John’s insight: “It is human nature to be optimistic on each of those three things. We typically overestimate how much money we can make and how quickly we can get adoption, and we typically underestimate just how hard it is to get something to market.” Put your CFO hat on. Act with an owner’s mentality. Would you back this with your own capital, or would you put that money somewhere else? If the answer isn’t immediate, the business case isn’t ready.

---

What Happens When You Get Adoption Wrong

Domain launched LeadScope, an AI product that predicts which homeowners are most likely to sell in the next 12 months. Government data. Browsing behaviour. Property attributes. It increased prediction accuracy from 5% (random chance) to 33%. 5-6x improvement. Real estate agents loved it. Ray White integrated it into their workflows. It generated millions in revenue. And then adoption stalled. Hundreds of agents adopted it out of 50,000 nationally. Less than 10% market share. The technology was brilliant. The product worked. The business case was sound.

“Even if you make great technology, it still doesn’t change the fact that you need to have a commercial, people, go-to-market model that makes it as easy for your people to sell and for your customers to use. Otherwise, it won’t get the adoption that it probably could.” The friction was simple: agents had to log into a separate system, export data, and manually act on it. Ray White’s integration removed one step. Predictions showed up directly in agents’ daily call lists. Adoption followed. The lesson: change will always underperform AI that fits into existing workflows. If your AI product demands a new habit, a new login, or a new process, you’ve added a 70% adoption tax. Most teams don’t account for it. They launch the technology and wonder why adoption is 1/10th of projections.

---

Why There’s No Universal AI Low-Hanging Fruit (Yet)

I asked John: “Where’s the low-hanging fruit in AI?” The one thing every company should do first. His answer: “I don’t think there are that many low-hanging fruits that apply to all industries that AI can give. I think there are some, but they’re quite specific to various functions and various industries.” In other words: There’s no universal playbook. Your AI opportunity is industry-specific. Your adoption model is function-specific. Your business case needs context.

AI will change less in the next year than we think. More in the next 5-10 years than we think. We’re entering the trough of disillusionment. Translation: If you’re pitching “AI will cut your costs by 40% in 6 months,” you’re selling snake oil. If you’re pitching “AI will make your best people 15% more effective and we’ll see bigger wins in 3-5 years,” you’re grounded in reality. The market rewards the second pitch.

---

My Take

John’s filter confirmed what I’ve been testing for two years: ledger before launch. Define your revenue or cost outcome first. Then build backwards. Australian boards don’t know what to look for in AI, they want governance when they actually need clarity on what’s changing and what stays the same: business outcomes are all that matters. The problem is cascade. If the person closest to the problem doesn’t understand why they’re adopting this AI, it doesn’t stick. LeadScope proves it: brilliant technology, 10% market share. Most Australian boards aren’t worried about AI hallucinating yet. They’re worried they don’t understand it well enough to govern it. That’s a knowledge problem dressed as risk. Business outcome first, technology second. Everything else is detail.

** **💼 What Australia’s Top Execs Are Listening To

Miss the full conversation? You’ll miss the filter most boards aren’t using. John Foong (ex-Google, Uber, Domain) goes deeper in this episode of Applied AI Australia, on what kills most AI business cases, and the one question every board should ask before greenlighting another pilot.

“AI that doesn’t touch revenue or cost is just noise.” Listen Now : Spotify 🎙️ Listen now: Apple

🎥 Watch the 60-sec clip:

📬 Subscribe to the newsletter

---

Ramon.