Bought AI Systems Succeed About 67% of the Time. DIY Internal Builds, About 22%.

In plain terms: The same MIT study behind this week's "95% of AI pilots show no payback" number also looked at what the winners did differently. AI systems bought from a specialized vendor or partner reached a successful deployment about 67% of the time. Systems a company built internally from scratch got there only about 22% of the time. Same technology. Roughly a third of the success rate. The pattern is quiet but hard to miss.

All week I have been writing about AI failing: 95% of pilots with no payback, companies quitting at more than double last year's rate, small businesses stuck "exploring," software sitting unused. Today is the one finding that points at what actually works.

Because the MIT researchers did not just count failures. They looked at how the small group of winners got there. And the split is stark.

What the research says

In MIT's GenAI Divide report, the teams that reached a real, working deployment fell into two camps. The ones who bought from a specialized vendor or built a partnership succeeded about 67% of the time. The ones who tried to build it internally, on their own, succeeded about 22% of the time.

I am going to give you the two numbers and let them sit, because that is the honest way to read them. You will see this stat floating around framed as "buying is 2x better" or "3x better," and the multiplier shifts depending on who is rounding. The two raw numbers are the finding: about 67 versus about 22. Draw your own line through them.

The honest note

Same caveat as Monday, because it is the same report. The famous 95% figure went viral and got sloppy, and this number rides in the same car. It leans on a set of interviews and public deployments that the report itself calls "directionally accurate," and "success" means reaching production with measurable results, not some perfect universal scorecard.

So read this as direction, not decimal. The exact gap could be wider or narrower than 67 to 22. What makes it worth your time is that it points the same way as everything else this week, and the same way as plain experience: the projects that die are usually the ones a company tried to invent from nothing, and the projects that ship usually had someone who had built the thing before.

Why buying wins

It is not that internal teams are worse. It is that building an AI system from scratch is a different job than running your business, and doing it well the first time is genuinely hard.

Someone who builds the same kind of system over and over has already hit the walls you are about to hit. They know which integrations break, where the data gets messy, what "done" actually looks like, and how to keep a human on the final call so the thing is safe to turn on. That is not magic. It is reps. When you buy or partner, you are buying the reps, and the reps are most of what separates a pilot that ships from a pilot that quietly dies.

The internal build, by contrast, spends its first months learning lessons that already exist. Sometimes that is worth it. Usually, for one specific workflow, it is the expensive way to arrive at the same place.

What this means for you

You are a small business, so you were never going to staff an internal AI lab anyway. That sounds like a disadvantage. This week's data says it is the opposite. The 22% camp is full of companies that had the budget to build it themselves and chose to, and mostly lost. You get to skip straight to the thing that works: pick one narrow problem and have someone who has built it before install it.

That is the whole idea behind a catalog of proven, fixed-price systems. Each one is a job that has already been built and run, not a science project that starts at zero on your dime. You are buying the reps, which is the exact thing the winning 67% had and the losing 22% did not.

You do not have to be the one who figures it out. In fact, the data says you are far better off if you are not.

FAQ

So I should never build anything in-house? Not never. If something is truly core to how you make money and truly unique to you, building it can be worth the low odds. But for the common workflows, lead follow-up, scheduling, reactivation, invoicing, those are solved problems. Buying the solved version is not settling. It is playing the odds the research just handed you.

Isn't "buy from a vendor" just what a vendor would want me to believe? Fair to ask, so weigh the source: this is MIT looking across 300 deployments, not a vendor's case study. And the mechanism is plain common sense. Someone who has installed the same system fifty times will beat someone installing it once. You would bet that way about a roofer or an accountant. AI is no different.

How do I avoid buying the wrong thing? Buy for one specific problem, not "AI" in general. The 22% failures usually started with a vague mandate to "adopt AI." The 67% started with a narrow job and someone who had done it before. Name the problem first, then buy the fix for exactly that.