Why 95% of AI Initiatives Fail (And What the 5% Do Differently)
Created on 2026-02-06 08:36
Published on 2026-02-08 09:00
The research that explains everything you are seeing in your organization
In early 2025, a team at MIT set out to answer a question that had been haunting boardrooms across the world:
Why are so many AI initiatives failing?
Not struggling. Not underperforming. Failing completely.
The researchers studied over 300 AI implementations across 52 organizations. They interviewed executives, analyzed outcomes, and tracked what happened after the pilots ended and the consultants left.
What they found should concern every business leader in Asia Pacific.
95% of organizations are getting zero measurable return from their AI investments.
Not low return. Not disappointing return.
Zero.
The Finding Nobody Wants to Hear
Let me be precise about what MIT’s Project NANDA research revealed.
Across 300+ implementations, the vast majority produced no measurable impact on the profit and loss statement. The pilots worked. The demos impressed. The announcements generated applause.
And then nothing happened.
The technology sat unused. The adoption stalled. The promised transformation never materialized.
Meanwhile, a small minority—roughly 5%—were extracting millions in value. They were implementing in weeks, not years. They were scaling beyond pilots. They were building competitive advantages that would compound over time.
Same technology. Same market conditions. Same regulatory environment.
Radically different outcomes.
The researchers expected to find technology explanations. Perhaps the 5% had access to better models. Perhaps they were in less regulated industries. Perhaps they had larger budgets or better vendors.
They found none of that.
Instead, they found a single sentence that explains everything:
“This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.”
The 5% who succeed are not using better AI.
They are using a different approach.
What the 95% Get Wrong
I have spent twenty-five years watching transformation initiatives fail. At HSBC, at Entomo, at organizations across APAC that hired me to help them change.
The pattern MIT identified is one I have seen countless times.
The 95% start with technology.
They license platforms. They hire “Heads of AI.” They build innovation labs. They run pilots. They announce initiatives with great fanfare.
And then they wait for transformation to happen.
It does not happen.
Here is why:
They deploy AI into organizations that are not ready. The leadership is not aligned. The data is siloed. The processes are broken. The culture is fearful. The governance is unclear.
AI does not fix these problems. AI accelerates them.
They treat AI as a technology project. They assign it to IT. They measure it with technology metrics. They evaluate success based on whether the system works, not whether the organization changes.
The system works. The organization does not change.
They copy visible rituals without building invisible foundations. They see successful companies announce AI initiatives, so they announce AI initiatives. They see successful companies hire AI leaders, so they hire AI leaders. They mimic the visible without understanding the invisible.
I call this the Cargo Cult Trap.
During World War II, Pacific islanders watched American military planes land and deliver valuable cargo. After the war, some islanders built replica runways and control towers, hoping to attract more planes.
The planes never came.
They had copied the visible rituals without understanding what actually made the planes land.
The 95% do the same with AI. They copy the announcements, the hires, the pilots. They do not build the foundations that make AI actually work.
What the 5% Do Differently
MIT’s research identified specific patterns among organizations that succeed with AI.
These patterns have nothing to do with technology selection.
They customize deeply.
The 5% do not deploy generic AI tools and hope for results. They invest significant effort in adapting AI to their specific context, workflows, and data.
Generic tools produce generic results. Deep customization produces competitive advantage.
One organization MIT studied spent three months training their AI on internal documents, customer interactions, and domain-specific terminology before any broad deployment. By the time they launched, the AI understood their business in ways that off-the-shelf tools never could.
They empower line managers.
The 5% do not centralize AI in innovation labs or IT departments. They push decision-making to the people closest to the work.
MIT found that organizations with centralized AI functions consistently underperformed. The AI was technically excellent but organizationally irrelevant. It solved problems that nobody on the front lines actually had.
The 5% let line managers decide how AI integrates into their domains. The result is AI that solves real problems, adopted by people who actually want to use it.
They build partnerships.
This finding surprised the researchers.
Organizations using strategic AI partnerships succeed at 67%. Organizations building AI internally succeed at only 33%.
The conventional wisdom said that building in-house would create competitive advantage. The data says the opposite.
The 5% know when to buy versus build. They recognize that AI technology is commoditizing rapidly. They focus their internal resources on what cannot be bought: understanding their own context, building their own capabilities, designing their own processes.
They partner for the technology. They own the Human Layer.
They accept the J-Curve.
Economist Erik Brynjolfsson’s research established something that most AI vendors will never tell you:
Productivity declines before it improves.
In organizations that are not prepared, productivity can decline by up to 60 percentage points before any payoff materializes. This is the J-Curve.
Most organizations panic at month three. The metrics look terrible. The board asks questions. The initiative gets killed.
The payoff would have come at month nine. But they never got there.
The 5% plan for the J-Curve. They set expectations with their boards. They protect initiatives through the valley. They measure leading indicators, not just lagging outcomes.
They have the patience to reach the other side.
They start with the Human Layer.
Before they deploy technology, the 5% ensure that their organization is ready to receive it.
They align leadership. They prepare their data. They develop capability. They redesign processes. They establish governance. They address cultural barriers.
This work is not exciting. It does not generate LinkedIn posts or conference invitations. It is invisible.
But it determines everything.
A Story of Approach
A Southeast Asian agricultural company wanted to implement a simple computer vision system. AI that could identify crop diseases from photographs.
The technology worked perfectly in testing. The pilot was successful. Accuracy was excellent.
Deployment should have taken two weeks.
It took nine months.
Not because the technology failed. The algorithm was fine.
It took nine months because the organization was not ready.
Data was hoarded. Different departments controlled different datasets. Information flowed up hierarchies but rarely across. Sharing data felt like giving away power.
Political resistance emerged. Middle managers saw AI as threat, not tool. They delayed. They raised concerns. They protected territory.
Trust was missing. Years of internal competition had created silos that no technology could bridge.
Nine months later, they finally deployed. By then, the budget was blown. The executive sponsor had moved on. The momentum was gone.
AI does not fail in the algorithm. It fails in the organization.
This is what MIT’s research confirmed at scale. The 95% who get zero return are not using worse technology. They are deploying into organizations that are not ready.
The Learning Gap
There is another finding from MIT’s research that deserves attention.
Most AI tools do not learn.
Organizations expect their AI systems to improve over time. They expect the technology to adapt to their context, remember their preferences, incorporate their feedback.
This rarely happens.
MIT found that the vast majority of enterprise AI deployments are essentially static. The AI performs the same on day 300 as it did on day 3. It does not remember previous interactions. It does not adapt to evolving needs. It does not get smarter.
This creates a fundamental barrier to scaling.
For AI to become transformative, it must learn from every interaction. It must build what I call the Context Graph: the accumulated record of why decisions are made, not just what happened.
The 5% build systems that learn. The 95% deploy systems that do not.
The Shadow AI Economy
Here is something else MIT discovered that should concern every executive:
Your people are already using AI. They are just not using yours.
MIT found that over 90% of workers are already using personal AI tools for work tasks. ChatGPT on their phones. Claude in their browsers. AI assistants that their employers do not know about.
Meanwhile, only 40% of companies have official AI subscriptions.
There is a massive gap between what organizations provide and what employees actually use.
This is the Shadow AI Economy.
Your most capable people have already adopted AI. They are using it to write emails, analyze data, prepare presentations, solve problems. They are getting more productive every day.
They are just doing it outside your view, outside your governance, outside your security frameworks.
The Shadow AI Economy tells you something important: the adoption problem is not that people resist AI. The problem is that your official AI initiatives are less useful than what people can get for free.
The 5% understand this. They find the power users who are already succeeding with personal AI tools. They learn from them. They build official systems that are actually worth using.
The 95% pretend the Shadow AI Economy does not exist. They mandate adoption of inferior tools. They wonder why engagement is low.
BCG Confirms the Pattern
MIT’s findings are not isolated.
BCG’s 2024 research found that 74% of organizations fail to achieve or scale value from AI. And 70% of the challenges they face are related to people and process, not technology.
70% people and process. 30% technology.
Yet most organizations spend 70% of their AI budget on technology and 30% on people and process.
They have the investment ratio exactly backwards.
The research is consistent. MIT, BCG, and every serious study of AI transformation reaches the same conclusion:
AI transformation is not a technology problem. It is a human problem.
The organizations that succeed recognize this. They invest accordingly.
The organizations that fail keep buying more technology and hoping the human problems will resolve themselves.
They will not.
The Approach That Works
Based on MIT’s research and my own experience across hundreds of implementations, here is what actually works:
Start with the Human Layer.
Before you deploy any AI technology, assess your organization across six dimensions:
-
Leadership and Vision: Do your leaders understand AI enough to lead, not just approve?
-
Data Readiness: Is your data accessible, clean, and governed?
-
Skills and Capability: Can your people judge AI outputs, not just use tools?
-
Process Maturity: Have workflows been redesigned, or is AI layered on broken processes?
-
Governance and Ethics: Are policies clear and accountability defined?
-
Culture and Change Capacity: Does your culture support experimentation?
Fix the gaps before you accelerate.
Choose your partners carefully.
The 67% success rate for partnerships versus 33% for internal builds is not an accident. Building AI infrastructure is hard, expensive, and increasingly unnecessary as the technology commoditizes.
Focus your internal resources on what cannot be outsourced: understanding your context, building your capabilities, owning your data.
Empower the front lines.
Centralized AI functions fail because they are disconnected from real problems. Push decision-making to line managers. Let the people doing the work decide how AI helps them do it better.
Plan for the J-Curve.
Set expectations with your board. Productivity will decline before it improves. Measure leading indicators—adoption, engagement, capability development—not just lagging outcomes.
Protect the initiative through the valley. Do not kill at month three what would have succeeded at month nine.
Find your Sparks.
The Shadow AI Economy tells you who your early adopters are. Find the people who are already using AI successfully. Learn from them. Equip them. Let them pull their peers along.
This is how we achieved 93% awareness and 72% participation in a transformation reaching 33,000 employees at HSBC. Not by mandating adoption. By making it worth adopting.
The Choice
MIT’s research presents every organization with a choice.
You can continue doing what the 95% do. License more technology. Announce more pilots. Perform transformation while getting zero return.
Or you can do what the 5% do. Start with approach, not technology. Build the Human Layer. Invest in people and process. Accept the J-Curve. Partner strategically. Empower the front lines.
The technology is the same for everyone.
The approach is what separates zero return from millions in value.
The divide is not about what AI you use. It is about how ready your organization is to use it.
MIT proved this at scale. The question is whether you will learn from their research or repeat the mistakes of the 95%.
Where does your organization fall on this divide? What would need to change for you to join the 5%?
If you want to assess your AI readiness honestly, comment “SCORECARD” below and I will send you the assessment I built for mid-market APAC leaders. It maps your organization across all six dimensions and shows you exactly where the gaps are.