Nearly every enterprise AI pilot returns nothing. The ones that work didn't buy a bigger model — they built for one problem and nothing else. Here's what that means for campus safety.
Last updated July 21, 2026
Ninety-five percent. That's the share of enterprise AI pilots that delivered no measurable return on the P&L, per MIT's 2025 study of 300 real deployments. Not "modest return." Not "slower than we hoped." Zero. Companies fed somewhere between $30 and $40 billion into these projects, and for 19 out of every 20, the money went in and nothing came out the other side.
The reflex is to blame the models. Too dumb, too hallucinatory, not ready yet. MIT went looking for that and didn't find it. The failure isn't the model. It's the canyon between a model that can do impressive things in a demo and a tool that actually understands the job it got dropped into. Generic AI is flexible, which is precisely why it wins for one person at a keyboard and dies the second you ask it to run a real workflow. Flexibility is a feature for a human and a liability for a process.
This is the whole "AI for AI's sake" problem in one number. Buy the technology, announce the technology, and assume returns show up because the technology is impressive. They don't. Impressive and useful turn out to be different products.
Because the thing being rewarded isn't results. It's the badge. A board asks the leadership team what their AI strategy is — and in higher ed, 72% of leaders say they'll be automating administrative work with AI within two years, so the question is being asked in every boardroom right now — and "we bought an AI thing" is a much faster answer than "we found one problem worth solving and solved it." Vendors know exactly how that meeting goes, so they sprint to staple "AI-powered" onto whatever they were already selling. Gartner has a name for this: agent washing. Of the thousands of companies marketing agentic AI, they estimate maybe 130 are the real thing.
And the market is about to find out. Gartner also expects more than 40% of agentic AI projects to be scrapped by the end of 2027, killed by rising costs and business value nobody can locate. That's not the models losing a step. That's the sticker peeling off in the rain.
Now, the honest counterargument, because it's a good one. General-purpose AI is genuinely great. ChatGPT is the most useful individual tool most of us have touched in a decade, and its flexibility is real value, and you'd be an idiot to rebuild from scratch what a foundation model hands you for free. All true. But flexibility for a person and flexibility for a workflow are opposites. The tool that can do anything is, by construction, built to do nothing in particular — and "nothing in particular" is a rough spec to hang a compliance obligation on.
From the same MIT study, and it's almost boring in its simplicity. They picked one pain point and built for it. The lead author's line stuck with me: the winners "pick one pain point, execute well." And when it came time to get it into production, tools bought from outside vendors reached the finish line about twice as often as the ones companies tried to build in-house.
Specialists who've run the same implementation a hundred times beat generalists doing it once. That's the return. That's where it lives.
A campus safety office does not have an "AI problem." It has a Clery problem and a Title IX documentation problem — but stop at the compliance line and you've named maybe a third of it. It also has a daily-crime-log problem, a dispatch-records problem, a records-request-from-general-counsel-due-tomorrow problem, a shift-change-handoff problem, a the-same-incident-keyed-into-four-systems-that-don't-talk problem, an annual-security-report-still-assembled-by-hand-in-September problem. And underneath every one of them, the officer-spends-three-or-more-hours-of-every-shift-writing-reports problem — that one's from a national survey, and it's more than half of them.
Clery and Title IX just happen to be the pieces with a federal deadline bolted on, so they get the budget and the anxiety. But the deadline was never the burden. The burden is the administrative overhead layered on top of every operational task the team touches, every day, regulator watching or not. Peel the compliance labels off and the shape underneath is identical: the people who should be doing the work are stuck documenting it.
General-purpose AI walks into that building, gets handed a login, and has no idea that a timely warning has a legal clock on it, or that a case file has to survive a federal audit two years from now. It'll write you a beautiful email. It will not keep you compliant — and it won't give the shift back to the people working it.
Purpose-built AI walks in already knowing. It knows what a Clery-reportable incident is because that is the only thing it was ever built to know. It treats the annual security report as a deadline with teeth, not a document. It does the work the general tool only gestured at during the demo.
Purpose-built has a real cost, and I'd be lying to skip it: it's narrow on purpose. A tool built to run Clery won't draft your strategic plan or brainstorm your all-hands. You're buying a scalpel, and a scalpel is a terrible way to butter toast. But nobody's return ever came from a tool that could kind-of do everything. It came from one thing done all the way.
The AI that finally generates the return AI was supposed to generate won't be the one with the biggest model or the slickest demo. It's the one built to solve your actual problem and nothing else. Everyone else is buying the sticker.
And 95% of them are getting exactly what a sticker is worth.
A real conversation about your operations. No sales deck, no demo environment.
Book 30 minutes →