Here's a number that should make every CTO lose sleep - 75%!

75% of enterprise AI initiatives are not delivering expected ROI. The models aren't bad, the data isn't always dirty (though it usually is). It's because most companies did the exact same thing they always do with new technology: they bolted it onto the old way of working and called it transformation.

The data this month is brutal. Multiple industry surveys converge on the same story. Somewhere between 72% and 90% of enterprises are now using AI in at least one business function. That sounds impressive until you learn that fewer than 20% have successfully scaled AI enterprise-wide, and about a quarter are seeing the returns they projected.

Andrew Ng made an interesting observation at AI Dev 26 last week. He argued that development teams should aim for 100% AI-generated code, with human review shifting to high-level design and what he calls "cognitive architecture." That got some eye-rolls on LinkedIn, but the core insight is sound: if you're still using AI the same way you used to write code manually (just faster), you're missing the point.

Harvard Business School researchers Karim Lakhani, Jared Spataro, and Jen Stave, PhD published a piece earlier this year calling this the "last mile" problem. Their argument is that the real bottleneck is not the technology. It's where technical capabilities meet existing organizational design. They advocate for "clean-sheet redesigns" where you imagine your workflows as if you were building the company around AI today, not retrofitting AI into the company you built in 2015. 

This matches what I've seen. And what I got wrong at first.

When we took on the AI-native modernization at Unity, the team's first instinct was to do exactly what everyone else does. Take existing production infrastructure, add an AI layer, and ship it. The pitch deck looked great. The architecture diagrams were beautiful.

It didn't work.

The problem was that our entire production infrastructure had been designed for traditional workloads running at massive scale (not to mention duplicated across a dozen plus acquisitions). Putting AI on top of that was like strapping a jet engine to a horse-drawn carriage. It technically moves faster, but the whole thing shakes apart.

The team figured out that we had to rethink infrastructure from the ground up. Not "add AI capabilities to the existing system" but "design the system as if AI is a core service, not a feature." That meant rebuilding how compute was allocated, how services talked to each other, how data flowed through the organization. It was a complete architectural rethink.

The result was a 20x reduction in operational spend while maintaining 99.99% availability. But the hardest part wasn't the engineering. It was the organizational transformation. We had to move the entire org into a self-service, AI-first culture. That meant changing how teams requested resources, how they thought about capacity planning, how they measured success. Every single workflow changed.

I spent the first three months convinced we could do this incrementally. The team was politely telling me I was wrong the entire time. They were right.

The Three-Layer AI Redesign Test

Before you declare your AI strategy a success (or a failure), run this diagnostic:

Layer 1: Did you redesign the workflow, or just accelerate it? If your humans are doing the same steps in the same order but with AI helping at step 3, you have an automation strategy, not an AI strategy. The question isn't "where can AI help?" It's "if we were building this process from scratch today, knowing AI exists, what would it look like?"

Layer 2: Did you change the org chart, or just the tech stack? AI-native operations require different team structures. At Unity, we couldn't run the new infrastructure with the old org design. Self-service AI means decentralized ownership, which means different reporting structures, different approval chains, different incentives. If your org chart looks the same before and after your AI transformation, your AI transformation hasn't happened yet.

Layer 3: Did you change how you measure success? If you're still measuring the same KPIs you used before AI, you're optimizing for the old system. A 20x reduction in operational spend doesn't show up on a dashboard that measures "tickets closed per week." You need metrics that capture the new value AI creates, not just the old work it accelerates.

The truth is that most enterprise AI strategies are failing not because of technology limitations but because of organizational inertia. The companies that are seeing real returns didn't bolt AI onto their existing operations. They used AI as the reason to rebuild those operations from first principles.

The question worth sitting with this week: if you were building your team's core workflow today, knowing what AI can do, would it look anything like what you're currently running?

#EnterpriseAI #AIStrategy #OperationalTransformation #Leadership #AureviumAdvisors