Ninety-five percent of enterprise AI pilots have no measurable impact on profit or loss. That figure comes from MIT’s Project NANDA, which spent the first half of 2025 reviewing more than 300 public AI initiatives and interviewing the leaders running them. It usually gets quoted as a warning about the technology. MIT’s own read focused on how companies integrated the tools rather than on the quality of the models.
That matches what we see. Almost every failure traces back to the same place, and it happens months before anyone turns on a model.
As an illustration, a large company buys Copilot licenses for everyone, mandates their use, and tracks who logs in. People who don’t use it enough get a talking-to. Nobody ever explains what good use looks like, and the company was never set up to do the interesting things in the first place. The data lives in dozens of places that don’t talk to each other. So the tool answers email a little faster, and that’s the whole return.
That company bought a fine product. It skipped the two steps that make any of this work.
AI is only as good as its context
An AI that knows your business does remarkable work. An AI that doesn’t know your business makes you re-explain the business every single time you open a chat window. That gap is a foundational problem, and no model upgrade can close it.
Walk into most mid-sized companies, and you’ll find the same environment we’ve been fixing since 1993. A spreadsheet on somebody’s desktop runs a critical process. A Google Sheet runs another one. An application someone bought a decade ago holds the customer records. Finance has its own thing. None of them talk to each other, and the person who understands how it all fits together is one retirement away from taking it with them.
Companies in that state can’t optimize much of anything, with or without AI. They’ve been told the optimizations are blocked on budget or headcount. They’re actually blocked because the data can’t move.
Since 1993, our answer has been to reinvent that environment. Connect the systems, normalize the data, get them talking, then put a central administrative application on top that manages it all. AI is becoming that central hub, and that’s the real change: you can now get there without commissioning a large piece of custom software. The prerequisite didn’t change. If your data can’t reach a place the AI can see, and if nobody keeps it accurate, nothing you build on top will hold.
Two steps have to happen first
Step one: connect the data. Every system you want the AI to answer questions about, reason over, or work inside must be reachable, and changes must be trackable. That doesn’t mean migrating everything into one monolith. Connected is enough.
Step two: set the policy that keeps it current. This is the step everyone skips, and it’s a management decision rather than a technical one: where work is allowed to live, and who owns keeping the record straight. Without it, the systems you just connected start drifting out of date the moment people go back to working the way they always have. Get it wrong and whatever you clean today starts going stale inside two quarters.
One step is plumbing. The other is management. Neither one is an AI project, and neither one requires custom software. This is just what a healthy company looks like, and most companies would benefit from doing it even if they never touched a model. AI can help you hold the line once the policy is in place. It can’t write the policy for you, and it can’t make people follow it.
Do both, and the door opens because the AI now has consistent, real-time, accurate ground truth to work from.
The scoping engagement comes first
There are products that handle pieces of this. There is no product that does it all, because every company’s mess is different. Drop an off-the-shelf AI layer on top of that mess, it finds nothing it can connect to, and you get the 95 percent outcome.
So we start with a consulting engagement, and we’re upfront about it being the first of two phases. We spend a month or two interviewing stakeholders, getting real logins to real systems, and cataloging what people actually use rather than what the official system list says they use. What comes out is an audit of everything you have and a plan for where it goes.
Two things about that plan are worth naming. First, a good chunk of it usually has nothing to do with technology. It’s process work you need to do to be ready, and we’ll tell you so. Second, if we get in there and find you’re already in decent shape, we’ll say that too, and the engagement ends. Selling the end state without the front door is how these projects become disasters.
Only after that do we talk about what gets built, and the answer is typically a combination of commercial products, custom software, and AI. The proportions depend entirely on what the audit found.
Two questions always come up at this point: whether a system like this runs unattended, and who is allowed to see what’s inside it. Nothing goes out, and nothing gets decided without a person, and the access question is a decision you want to make early.
The window is open now
Companies that connect their data and fix their processes this year get compounding returns from every AI tool that ships afterward. Companies that buy the tool and skip the plumbing will spend real money, get very little in return, and conclude that AI was overhyped.
The second group will be right about their own results and wrong about the reason.
Source for the 95 percent figure: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025). Based on a review of more than 300 publicly disclosed AI initiatives, 52 structured interviews, and 153 survey responses from senior leaders, conducted from January through June 2025.
Drew Linn is CEO of Counterpart, a custom software firm in Indianapolis. We connect the systems that run your business and become the long-term owner of the complex ones. If you want to know where your data actually stands, that conversation is worth having before you buy anything.