Companies spent billions on generative AI this year. Most still can’t find last quarter’s signed contract. Is that ironic? Depends on your definition of ironic. It is, at minimum, expensive.
A widely cited study from MIT’s Project NANDA confirmed what employees already suspected. Ninety-five percent of enterprise generative AI pilots failed to deliver measurable value. That isn’t a rounding error. It’s closer to a verdict.
The Chatbot Didn’t Organize the Filing Cabinet
Here is what actually happened at most companies. Leadership bought a generative AI subscription. Employees got a chatbot that writes decent email drafts. Nobody’s documents got any easier to find.
That gap is not a technology failure. Generic AI was never built to fix filing chaos. Ask it to find last April’s signed NDA, and watch it draft a lovely email about NDAs instead. Widen the request to every contract signed in five years, and it offers a bulleted list of contract-management tips instead. Helpful. Not remotely what you asked.
So the underlying mess stayed exactly where it was. Contracts still live in someone’s inbox. Invoices still hide inside mislabeled folders. The AI arrived; the chaos never left.
Hiring Instead of Fixing
Faced with this gap, companies did what companies do. They hired people to manage it. Job boards filled with titles like “Head of Generative AI Operations” almost overnight. Somewhere, a new “VP of Prompt Governance” is updating a LinkedIn headline right now.
Ask ten of these hires what their job actually involves. Expect ten different answers. Mostly, they are highly paid people manually doing what better document infrastructure would do automatically.
This is the talent bubble. In fairness, hiring feels like progress. It is, technically, an action. Headcount grew to compensate for a problem nobody fixed at the source. Adding staff to route documents by hand is not a strategy. It is a workaround wearing a strategy’s clothes.
Buying Beats Building, by a Wide Margin
MIT’s researchers found something else worth sitting with. Buying tools from specialized vendors succeeded roughly 67% of the time. Building internal solutions succeeded barely one-third as often.
The lesson is not subtle, though plenty of AI budgets will ignore it anyway. Budgets have never been known for reading research papers. Purpose-built tools already solve document chaos directly, without a new department to run them. A generic chatbot was never the right layer for this problem.
Optix, for instance, classifies and routes incoming documents automatically, no custom model or five-person team required. Document capture and workflow automation were solved problems well before the current AI spending wave began.
What the Successful 5% Actually Did
The report didn’t just measure failure. It also studied the outliers who succeeded. What did they do differently? Not what most executives would guess.
First, they focused on back-office workflows, not flashy sales tools. More than half of GenAI budgets went to sales and marketing. The stronger returns showed up in back-office automation instead.
Second, winners favored vendors who learn. Static software that behaves the same on day one and day 200 doesn’t adapt to a company’s real workflow. The successful 5% wanted systems with memory, ones that improve as they process more of an organization’s actual documents.
Third, they treated vendors like outcome partners, not software licenses. They wanted a result, not a login.
None of this describes a chatbot bolted onto existing chaos. It describes infrastructure that learns a company’s own document patterns over time, the back-office problem MIT flagged as underrated.
The Bottom Line
Buying a chatbot was never going to fix a filing cabinet. Ninety-five percent of generative AI pilots prove the point at scale. A “Head of Generative AI Operations” looks impressive on a business card. It does not locate a single contract. Before hiring another one, ask what’s actually broken.
If the honest answer is “we still can’t find our documents,” the fix isn’t more headcount. It’s infrastructure built for exactly that problem, not a chatbot bolted onto the chaos. Simple? Mostly. Common? Apparently not.
Contact us today to learn more about how Optix by Mindwrap classifies and routes your documents automatically, no custom model or new department required.