The Announcement Nobody Talks About
A few years ago, I watched a large enterprise make a decision that felt, at the time, like a bold bet on the future. The company, a global leader in its industry, assembled a team of AI and data engineers, invested millions in salaries, infrastructure, and tooling, and set out to build a proprietary AI platform from scratch.
The ambition was real. The talent was real. The money was very real.
The platform never delivered what it promised.
Eventually, the announcement came quietly: the platform would be sunset. The AI team would be either made redundant or redeployed to other projects. And the company would migrate to a commercially available, off-the-shelf solution as the more pragmatic path forward.
I wasn't surprised. I had seen patterns like this from the inside.
What I've since realized is that this story isn't a cautionary tale about one company's missteps. It's a pattern. And it's one that enterprise leaders keep repeating despite decades of evidence pointing in the opposite direction.
The platform never delivered what it promised. The announcement came quietly. The team was made redundant. The company moved to off-the-shelf.
The Off-the-Shelf Reality Check
There's something important buried in these migration stories. Companies don't move to generic AI tools because they're better. They move because they're cheaper and faster to deploy than maintaining a failing internal build.
I've tested several of these tools myself. They're genuinely useful for a category of work: summarizing emails, drafting documents, pulling data into familiar interfaces. Repetitive administrative tasks where the domain is general and the stakes are low.
But the moment you push toward anything operationally complex, real-time decision support, cross-system anomaly detection, financial quantification of inefficiencies, you hit a wall. Generic AI tools aren't designed for that. They were never meant to be.
The harder question, does this AI actually move our operational outcomes, gets deferred. Leadership, having spent years watching an in-house build collapse, is relieved to have something to show. But relief is not a strategy.
The risk is that "we deployed an AI tool" gets reported as "we have an AI strategy."
We Already Learned This Lesson. With ERP.
Here's what strikes me most about the corporate AI build failure pattern: we've been here before. Exactly here.
In the 1990s and early 2000s, a wave of large enterprises decided to build their own ERP systems. The logic was seductive: we're complex, our processes are unique, off-the-shelf systems won't fit us, and we have the engineering talent to do this ourselves.
Some built their own warehouse management systems. Some built their own transportation management systems. The ones that went furthest down this path almost universally arrived at the same place: years late, hundreds of millions over budget, and saddled with a system that was already outdated by the time it launched, because the market had moved faster than any internal team could match.
The companies that thrived were the ones that bought best-in-class platforms and focused their internal engineering on integration, configuration, and the genuinely proprietary problems only they could solve.
The AI era is replaying this dynamic almost beat for beat. And the enterprises still trying to build proprietary AI infrastructure in-house are making the same bet that failed a generation ago.
There are narrow exceptions. If your competitive advantage is AI, if the model itself is the product, then building may be justified. But if you're a manufacturer, a retailer, a logistics operator, or a professional services firm trying to use AI to run your operations better, you are not in the business of building AI. You are in the business of using it.
The enterprises winning with AI aren't the ones who spent three years building internal platforms. They're the ones who asked: who has already solved this for my domain?
What Good Partnership Actually Looks Like
Partnering with an external AI provider isn't the same as deploying a generic tool and calling it a day. The distinction matters enormously.
A generic AI tool gives you general-purpose capabilities that improve productivity at the individual task level. That's valuable, but it's table stakes. Your competitors have the same access to the same tools.
A domain-specific AI platform gives you something different:
- Intelligence that understands your operational context
- Financial quantification of impact in terms your CFO can act on
- Compounding improvement: it gets smarter with every week of deployment because it's learning from your data
That's the compounding advantage that generic tools can never deliver. The build-vs-buy decision, properly framed, isn't "should we use AI or build AI?" It's: what is our actual source of competitive advantage, and are we investing in that, or in infrastructure that a specialist can build better, faster, and cheaper than we ever could?
The answer, for the vast majority of enterprises, is clear. It just took a few billion dollars in failed internal builds to make it obvious.
