Forbes Argentina ran a piece this week on a number that shouldn't surprise anyone who works with AI every day, but is still worth taking seriously: 95% of enterprise AI pilot projects generate no measurable financial return. We went to the original source -an MIT Media Lab study (Project NANDA) called "The GenAI Divide: State of AI in Business 2025"- to understand what's behind the headline, and the detail is more useful than the number itself.
The study is based on 150 interviews with business leaders, a survey of 350 employees, and an analysis of 300 public AI deployments. Its central finding: companies have already invested between $30 and $40 billion in generative AI, and 95% of those pilots moved nothing on the P&L. This isn't an adoption problem -the money is already spent-, it's a problem of what was done with it.
The real divide isn't "using AI or not"
The study calls it the "GenAI Divide": 5% of companies are extracting millions of dollars in real value from their AI projects, while the other 95% remain stuck with no business impact. The detail that stood out most to us is where the budget goes versus where the return actually is: most of the money goes into sales and marketing pilots -chatbots, content generators-, which is exactly where the measured ROI came out lowest. Back-office automation -internal processes, operations, repetitive team tasks- is what actually cuts real costs and shows up on the balance sheet, and it's the one getting the least budget.
That's not a technical footnote, it's the whole explanation for the failure rate: the money is going where there's more marketing noise, not where the real bottleneck is.
Buying a partner vs. building alone
There's another finding from the same study that, honestly, hits close to home for us as a software factory: companies that bought a specialized tool or partnered with an external provider had a success rate close to 67%. Companies that tried to build the solution entirely in-house, with their own team and no prior experience with the problem, barely reached a third of that rate.
We don't read that as "just hire a vendor." We read it for what it is: building with AI isn't only writing code on top of a model, it's having already seen what breaks when a real process gets automated badly -and that's learned by getting it wrong at other companies first, not at the first one that tries.
Our take, after doing this for real
We said this before on this same blog: the business is never "implementing AI," it's solving a bottleneck. This MIT study confirms that exact diagnosis with hard numbers. The projects that fail almost always share the same pattern: they start from a generic, flashy pilot instead of a specific, measurable problem, they treat AI as a feature you install once and forget, and they don't have a team with enough field experience to notice when the underlying process -not the model- is what's actually broken.
The AI projects that do work in our own case studies share the opposite pattern: narrow, measurable, and back-office. The Previnca WhatsApp agent doesn't try to solve "all of customer service," it solves one specific flow. The risk assessment automation we built for banks with RiskRator isn't a sales chatbot, it's a concrete operational task a person used to do by hand. Neither one was a demo built to show off "we use AI": both are measured in hours saved and errors that no longer happen.
What we'd change, concretely
If we had to summarize in three changes what separates the 5% from the 95%: pick one specific problem you can measure in money before writing a single line of code, instead of an ambitious pilot that sounds good in a slide deck; measure the result on the balance sheet from month one, not in how many people "used" the tool; and treat the implementation as a relationship that gets adjusted over time, not a project that gets delivered and abandoned the moment it works in the demo.
None of this is new, and none of it is exclusive to AI -it's the same thing that separates any well-executed software project from one that isn't. What's new is that there's now an MIT study with 300 real cases confirming that the shortcut of "let's install AI and see what happens" is almost always expensive.
If your company is evaluating an AI project and you don't want to end up in that 95%, let's talk. We always start in the same place: understanding the real bottleneck before deciding which tool solves it.
