Early in my career, at MITRE and Lockheed, I integrated the first generation of language tools, the ones that read documents and pulled out the people, places, and organizations inside them, into systems analysts could actually use. We called it natural language processing and named entity recognition, not AI. The point is that the technology for finding meaning in text has worked for decades. So when an AI project fails today, the technology is rarely the reason.
The numbers bear this out. MIT’s Project NANDA found in 2025 that despite tens of billions in enterprise spending, roughly 95 percent of generative AI pilots delivered no measurable return, and the cause was not model quality but flawed organizational integration. The technology works. The deployments do not.
And they fail for reasons that have almost nothing to do with engineering. Leadership decides the company needs AI, usually because a competitor announced something, and mandates it “everywhere.” But everywhere is not a strategy, it is the absence of one, with no specific problem to solve and no honest way to call it a success or a failure. Often the advice making it worse comes from a consultant who sells a single engine and wedges that one tool into every project, fitting the problem to the product instead of the product to the problem. And just as often the leader running it treats AI as a technology decision, because that is the part they feel equipped to own, and misses that the deciding factors were organizational all along.
Underneath all of it is the truth I spend most of my time on with clients. AI does not fix the gap between your business and your technology. It exposes it, and it amplifies it. If your data is scattered, if no one agrees on what a “customer” or a “project” even means inside your own records, if your processes only hold together because a few experienced people quietly carry them, AI runs straight into all of it, faster and more expensively than your old systems ever did.
The fix is simple to say and genuinely hard to do. Pick one painful, well-scoped problem, the kind where you can describe exactly what good looks like before you start. Close the data and process gaps around that one problem first. Give it a real owner and an honest way to measure it. Prove value at small scale, then expand from something that actually worked rather than from a slide that promised everything.
The technology has been quietly doing its job for decades. What never improves on its own is the discipline to aim it at one real problem inside an organization that is ready for it. So before you approve the next AI initiative, ask a deceptively simple question: what specific problem are we solving, and what does good look like in six months? If the room cannot answer it, you have not found a technology gap. You have found where the real work begins, and you have already separated yourself from the ninety-five percent.
