An AI expert said this out loud.
“Don’t build your own AI. Every time a new foundation model drops, it outperforms anything we’ve spent months fine-tuning.”
He’s spent years applying AI to healthcare — training models, pushing accuracy, publishing results. What made his work obsolete wasn’t a competitor. It was a routine model update from Google or OpenAI. He laughed and said it anyway: “Just grab the latest model. That’s the winning move.”
That sounds like defeat. The data says it’s strategy.
95% of enterprise AI projects fail to move the needle
MIT’s NANDA Initiative published 「The GenAI Divide: State of AI in Business 2025」.
The finding: 95% of companies deploying generative AI saw no meaningful impact on P&L.
But look closer. The problem wasn’t the AI. It was how companies integrated it. Firms that bought from specialized vendors and built partnerships succeeded 67% of the time. Those that relied on building in-house? One third of that rate.
The builders lost. The buyers won.
I’ve seen this question before
At Hyundai Motor, I spent years asking: what does enterprise AI capability actually mean? Training programs, tools, safe environments, incentives — those were obvious. But concrete organizational initiatives? That’s where I kept hitting a wall.
Earlier in my career, I worked on equipment strategy for a semiconductor manufacturer. Build the critical tools in-house, or source from the world’s best vendors? Two criteria drove the decision: Can we continuously match world-class quality? Does this technology directly determine our core competitive edge?
Apply the same lens to AI.
Generative AI capability comes down to compute and data. It’s a capital game. Almost no company will out-build Google, OpenAI, Anthropic, or Meta — and that’s already been proven. So the real question becomes: does AI model quality directly determine your competitive edge? The AI expert’s experience suggests it does — more than most executives realize.
That last question is yours to answer. Not with data. With judgment.
Real AI capability means being replaceable — by choice
If your answer is yes, here’s what your organization actually needs to build.
When a better model launches, you can’t afford weeks of internal review. You need automated evaluation pipelines that test new models against your specific use cases, score them fast, and let you swap them in without friction. Not the ability to build models — the ability to replace them.
That’s the real internalization play — a structure that keeps you independent from any single vendor, always positioned to use what’s best. An organization that doesn’t break when the technology shifts.
Is your company’s AI architecture ready to swap in next year’s model?