Most enterprise AI projects fail in production, and the reason is consistent and counterintuitive. It is almost never the model.
In 2025, MIT's NANDA initiative studied 300 enterprise AI deployments, surveyed 350 employees, and interviewed 150 leaders. The headline finding: 95% delivered no measurable return. Only about 5% reached rapid revenue acceleration. The report is called "The GenAI Divide," and the divide is not between companies with good models and bad ones.
The number: 95%
Ninety-five percent of enterprise generative AI pilots produced no measurable impact on the P&L. Not a rounding error. The overwhelming majority.
MIT is not the only group seeing this. Gartner, looking at the same window, projected that through 2026 roughly 60% of AI projects would be abandoned because the underlying data was not AI-ready. Two research groups, two methods, one conclusion from opposite ends: these deployments fail on context, not on the model.
The instinct is to blame the obvious things. Model quality. Regulation. Data privacy. MIT checked, and that is not where the failures came from.
It's not the model. It's the learning gap.
MIT's own diagnosis was blunt: the core issue is not the quality of the AI models. It is what they called a learning gap. Generic tools like ChatGPT are flexible enough to help an individual, but they stall inside an enterprise because they do not learn from or adapt to how the business actually works.
Read that again, because it is the whole story. The tools don't know the business. They are brilliant at language and useless about your company, because nobody ever gave them your company in a form they could use.
Picture the pilot that dazzled in the demo. A support assistant answered every test question perfectly, so it shipped. In production it told an enterprise customer their plan included a feature that only runs on the cloud tier, because no one had told it which features ship on which plan. The model was state of the art. It had never been given the one fact the answer turned on.
What "doesn't learn the business" looks like day to day
A tool that hasn't learned your business shows up in four ways, and you have seen all of them.
- Ask three of your AI tools what "active customer" means and you get three answers.
- The rules that actually run the company, why an account gets a discount, how a renewal gets approved, were never written down anywhere a tool can read.
- Every new tool gets taught the business from scratch, and the explanations drift apart.
- You can't tell what any answer was based on, so you can't put it in front of a customer.
These are not four problems. They are one problem: the company was never written down for the machine. (We go deep on this in the pillar piece, "Your AI Doesn't Know Your Company.")
Why a better model won't fix it
The reason this keeps catching teams out is that it looks like a model problem, so they treat it like one. When the AI gives a wrong answer, the instinct is to try a stronger model, add more data, or extend the context window. None of that touches the actual gap.
A model can only reason on what it can read. The definition of "active customer" your company actually uses, the rule that a specific account renews on net-60, the reason a customer gets a discount, none of it is a faint signal a better model would surface. It was never recorded anywhere in a form a model can read, so there is nothing to learn from. The gap is not the model's intelligence. It is that the company's meaning lives in people's heads and scattered tools, and the smartest model in the world cannot infer a decision that left no trace. This is why the 95% who keep tuning models stay stuck: they are sharpening a tool that was never missing sharpness.
The build-versus-buy data says the same thing
MIT found that AI bought from specialized vendors and wired into real workflows succeeded about 67% of the time. Internal builds succeeded roughly a third as often.
The lesson is not "buy, don't build." Plenty of internal builds are excellent engineering. The lesson is that the projects that worked were the ones connected to how the business actually operates. The ones that failed were impressive demos that never learned the company they were dropped into.
People want it. That is not the gap.
Here is the number that kills the "adoption problem" excuse. MIT found only 40% of companies had official LLM subscriptions, but 90% of employees were already using personal AI tools for work.
Demand is not the problem. Your team is reaching for these tools on their own. They stall the moment the work requires the tool to know something specific about your business that it has no way to know.
What the 5% did differently
The projects that crossed the divide did one unglamorous thing. They gave the AI the context it needed: the definitions, the rules, the meaning that make the business specific, in a form the tool could read at the moment it answered.
That is the entire difference between a demo that dazzles and a system you can ship. The demo knows the script. Production needs it to know the business.
What closing the gap actually takes
The 5% did not find a better model. They gave the AI the company, in a form it could read while it worked. In practice that comes down to a few things. The definitions and rules get written down once, with an owner and a date, instead of living in someone's head. The unwritten knowledge, the renewal logic, the account sensitivities, gets captured deliberately rather than left in Slack. All of it gets served to every tool at the moment each one answers, not pasted separately into each tool's prompt. And every answer traces back to the context it used, so a wrong answer points to the rule that needs fixing instead of to a shrug.
That is the difference between the projects that shipped and the ones that stayed demos. The architecture term for writing the company down once and serving it to every tool is a context layer, and building it is the work the successful minority actually did.
The takeaway
If your AI keeps failing in production, the most useful thing you can do is not swap models. It is to write your company down, what your data means and how it actually works, somewhere every tool can read it. The 95% kept tuning prompts and models. The 5% fixed the context.
Your AI doesn't know your company. Fixing that is the work, and it is what we are building Sento to do.
Frequently asked questions
What percentage of enterprise AI projects fail?
MIT's 2025 NANDA report studied 300 enterprise deployments and found 95% delivered no measurable return; only about 5% reached rapid revenue acceleration.
Why do enterprise AI projects fail?
MIT's diagnosis was a "learning gap," not model quality: generic AI tools do not learn from or adapt to how a specific business works, so they stall when the work requires company-specific knowledge.
Is the problem the AI model?
Usually not. The report explicitly found the failures were not driven by model quality. The gap is context, the definitions, rules, and meaning the tool was never given.
What do the successful 5% do differently?
They connect the AI to how the business actually works, giving it the company's definitions and rules in a form it can read at inference time, rather than relying on a generic model to guess.