After $30–40 billion of enterprise AI spend, MIT's Project NANDA reviewed more than 300 AI initiatives and reached a conclusion that should concern every executive currently signing off on an AI programme.
Ninety-five per cent of organisations are seeing no measurable P&L impact. Not modest returns. Not results that are hard to attribute. None.
Only five per cent of integrated AI pilots are extracting meaningful value - and the gap between that five per cent and the rest is not a gap in technology. It is a gap in how organisations think about what AI transformation actually requires.
"Most GenAI systems do not retain feedback, adapt to context, or improve over time."
MIT Project NANDA - The GenAI Divide: State of AI in Business 2025
That observation cuts to the heart of why most AI deployments plateau. The tools are capable. The workflows around them are not designed to extract that capability, feed it with real-world signal, or compound it over time. The AI does something impressive in isolation. Then nothing changes.
The friction problem
MIT's finding is that the primary barrier to AI value is not infrastructure, regulation, or talent. It is learning - specifically, the absence of mechanisms that allow AI systems to improve through use and allow organisations to improve alongside them.
The reason this barrier persists is straightforward: organisations avoid friction.
Embedding AI meaningfully into a high-value workflow is disruptive. It requires changing how decisions are made, who is accountable for what, what gets measured, and how information flows between systems. That is uncomfortable. It generates resistance. It takes longer than a pilot. So instead, organisations insert AI into the path of least resistance - a summarisation tool here, a drafting assistant there - and measure the results of that insertion against the cost of a platform licence.
The results are predictably thin. Not because the technology is weak, but because isolated intelligence produces isolated value. The AI is not connected to anything that matters enough, deeply enough, to move a number on a P&L.
The MIT research is unambiguous on what success looks like: deep integration into specific, high-value processes; continuous learning through feedback loops that allow the system to adapt; and evaluation based on business outcomes rather than technical benchmarks. Each of those three features requires friction. Each requires the organisation to change, not just the technology stack.
The iceberg most organisations never look at
There is a useful way to think about where AI value actually comes from. Roughly ten per cent of the outcome is determined by the model - which underlying AI you use. Another twenty per cent is determined by data quality and availability. The remaining seventy per cent is determined by transformation: the people, processes, and change management that determine whether the technology produces anything real.
Most organisations spend their attention - and their budget - on the ten per cent. Model selection consumes weeks of evaluation. Vendor comparisons dominate steering committee agendas. The question of which large language model will power the deployment gets treated as the central strategic decision.
It is not. The model is largely undifferentiated at the capability level that matters for most enterprise use cases. What is differentiated - and what determines whether the investment returns anything - is the seventy per cent that sits below the waterline. The process redesign. The integration depth. The feedback mechanisms. The change programme that makes adoption real rather than nominal.
This is not a comfortable conclusion for organisations that have already bought a platform. It means the hard work is still ahead of them, and that the platform was the easy part.
Real automation requires new processes, not improved ones
The instinct when deploying AI is to find an existing workflow and make it faster. That instinct is understandable and almost always insufficient.
Adding AI to an existing process optimises that process. It does not transform it. The process was designed for human execution - with the hand-offs, approval steps, and information structures that human execution requires. AI inserted into that process inherits all of those constraints. You get a faster version of a workflow that was not designed for AI, rather than a workflow that takes advantage of what AI actually makes possible.
Real automation - the kind that produces the step-change results that justify the investment - requires asking a different question. Not "how do we do this faster?" but "if we were designing this process from scratch today, knowing what AI can do, what would it look like?" The answer is almost always structurally different from what exists. Fewer hand-offs. Different decision points. Different accountability structures. Data flowing in ways the old process never contemplated.
That is an operating model question, not a technology question. And it is where the seventy per cent lives.
Why this matters more in the GCC than anywhere else
Across the UAE, Saudi Arabia, and the wider Gulf, AI adoption is being driven at a pace and with an executive mandate that most markets do not have. National AI strategies, sovereign wealth fund commitments, and Vision 2030 programmes are creating real pressure to deploy and demonstrate results - quickly.
That pressure is a double-edged thing. It accelerates investment and cuts through the organisational inertia that slows AI programmes in more cautious markets. But it also creates an incentive to deploy visibly rather than effectively - to launch pilots that can be reported upward rather than transformations that take the time they require.
The MIT findings are a warning for exactly this dynamic. Ninety-five per cent failure at a global level means the odds are already against any given deployment. Deploying at speed, without the process redesign and integration depth that the five per cent got right, does not improve those odds. It reproduces the failure mode at GCC scale.
The opportunity, however, is genuine. Organisations in the region building new functions - new shared services, new digital government capabilities, new financial institutions - do not carry the legacy process debt that makes transformation so costly in established markets. They can design for the seventy per cent from day one rather than fighting their way through decades of accumulated workflow. That advantage is real. The question is whether it will be used.
Orchestrated intelligence, not isolated intelligence
The five per cent that MIT identifies as genuinely successful share a common architecture: AI that is deeply embedded in consequential workflows, connected to the systems that hold the data those workflows depend on, equipped with feedback mechanisms that allow it to learn through use, and governed by outcome metrics that keep the programme honest.
That is not a description of a pilot. It is a description of a programme - one that treats AI as the mechanism of an operating model transformation rather than a tool added to an existing one.
The ROI does not live in isolated intelligence. It lives in orchestrated intelligence at scale: AI that knows the context, retains what it learns, improves through every interaction, and is embedded deeply enough in real work that its output actually changes decisions.
Getting there requires friction. It requires the process work that most organisations avoid. It requires measuring outcomes rather than activity. And it requires treating the seventy per cent - the transformation, the people, the change - as the primary investment, not the afterthought.
The ninety-five per cent did not get that wrong because they lacked ambition. They got it wrong because they confused buying AI with deploying it.