Why 95% of AI Projects Never Deliver and What Your People Have to Do With It
Everyone's buying AI. Almost nobody's getting value from it.
MIT's Project NANDA put a number on it in their State of AI in Business 2025 report: despite an estimated $30–40 billion in enterprise investment in generative AI, 95% of organisations see zero measurable return. Not "less than expected." Zero. The research drew on 300 publicly disclosed AI initiatives, 150 leadership interviews, and 350 employee surveys.
When you see a number like that, the instinct is to blame the technology. The wrong model. The wrong vendor. Bad timing. But the models work. The gap is between what AI can do and what most organisations actually know how to do with it. That gap lives in your people, and closing it is the reskilling challenge most businesses haven't taken seriously yet.
What's actually going wrong
Picture the typical rollout. A team gets access to a new AI tool, starts using it, and ships whatever comes out. Nobody stops to ask whether the output is grounded in the right data, whether it makes assumptions nobody has verified, whether there's a hallucination buried in paragraph three.
AI systems, large language models in particular, are confident by default. They produce fluent, plausible-sounding output whether or not that output is accurate. A model doesn't flag uncertainty the way a junior analyst might; it just answers. When nobody on the team has the skills to interrogate that answer, errors travel through workflows until they cost something real.
MIT calls this the "learning gap" — a human failure, not a technical one. Teams that can't challenge, verify, and direct AI outputs will fail just as reliably as the tools that let them down.
What AI literacy actually looks like
There's a version of AI literacy that gets talked about a lot, and it's too shallow to be useful. "Understand what AI can and can't do" sounds right. In a working context, though, it means something more specific.
A language model will fabricate a plausible statistic when it lacks the right data. Knowing how to prompt to reduce that risk, and how to verify what comes back, is a learnable skill. So is understanding that AI outputs reflect patterns in training data, not ground truth about your business. A model doesn't know your customers, your risk appetite, or the three things your CFO always pushes back on. Someone on your team does. The skill is combining what the model produces with what your organisation already knows.
Then there's the judgement question: which tasks should actually go to a model, and which ones shouldn't? The people who get real value from AI are the ones who've thought carefully about where that boundary sits, and who stay engaged when the stakes are high. That kind of thinking gets built through doing real work with real data, not through a generic training course.
The distribution problem
MIT's research also points to something structural. In most organisations, AI capability is concentrated in a small group of specialists while the rest of the business watches from the sidelines. When that's the shape of things, adoption stalls across the board. Business-unit teams don't trust outputs they don't understand. Leaders without an AI background struggle to make informed calls about what to automate or where the risks sit, and early-career staff build habits around AI tools without ever developing the critical thinking to use them well.
The 5% of organisations generating real value look different. They have AI-capable people spread across the business: line managers driving adoption in their own teams, domain experts applying AI to workflows they actually own. Spreading that capability is a workforce transformation challenge. Buying better software doesn't solve it.
Why most training falls short
Most upskilling efforts share the same flaw: they separate learning from doing. Skills built in the abstract don't transfer because they were never built in context.
What works is learning embedded in your actual business environment, using your data, your processes, your specific use cases. A finance team that builds a working AI workflow with real KPI data learns things no generic case study can teach: where the model gets it wrong, how to structure outputs their colleagues will actually trust, why verifying before shipping matters. That understanding sticks because it was built in the context where it needs to work.
Completion metrics tell you who sat through the training. They say nothing about who can actually use AI well under pressure.
The real bottleneck
The 95% of organisations getting nothing from AI are stuck because the people around the technology haven't caught up. They don't know enough about how it works, where it breaks, or how to fill the gaps it leaves. The technology isn't the bottleneck.
Building genuine AI capability across an organisation (finance, operations, IT, leadership) is harder and slower than buying better software. It requires real reskilling, delivered in context, measured by what people can actually do. But it's also what separates the businesses that extract lasting value from AI from the ones still waiting for their investment to pay off.
Etiq Reskilling helps organisations build practical AI capability where it matters — through role-based learning pathways, applied projects grounded in your own data, and verified assessments that measure readiness rather than just completion. Learn more about how Etiq works.
