Companies are spending enormous sums on artificial intelligence, and most of them are getting almost nothing back.
In 2025, researchers at MIT looked at what was actually happening inside businesses, rather than what was being claimed on stage or on LinkedIn. Their report, The GenAI Divide: State of AI in Business 2025, drew on interviews with around 150 company leaders, a survey of roughly 350 employees and an analysis of 300 real AI deployments. The headline finding was blunt. About ninety five percent of company AI pilots were delivering no measurable impact on the bottom line. Only about five percent were producing real returns.
It is tempting to read that and decide AI is overhyped, but that would be the wrong lesson. The same tools that produce nothing in one company are transforming another. The difference is rarely the technology. It is almost always the way the organisation around it is led.
This is where the most useful conversation tends to begin. In Helen Sterling’s work as an AI keynote speaker, the leaders who get real value from AI are almost never the ones with the cleverest tools. They are the ones who understand that the hard part of AI is not the technology, it is the people and the way the work is organised around it.
So the real question for any leader is not which tool to buy. It is why the vast majority of AI projects fail, and what the small number that succeed are doing differently.
In Brief:
Around 95% of company AI pilots deliver no measurable return, and the reason is almost never the technology. The barriers are organisational: unclear goals, a culture of fear around AI use, tools bolted onto unchanged processes, and budgets chasing visibility rather than value. This article examines what the research shows about why most rollouts fail, who the “secret cyborgs” are and why they matter, and what the small number of successful organisations do differently.
What 95% of AI pilots have in common, and why it matters
Plenty of organisations can point to an AI pilot. Far fewer can point to a result. That is the gap the MIT researchers called the GenAI Divide: high adoption on one side, almost no real transformation on the other.
Despite tens of billions of dollars going into enterprise AI, the study found that the overwhelming majority of pilots never turned into anything that showed up in the numbers. One manufacturing executive summed up the mood neatly, telling the researchers that while the hype online says everything has changed, inside their operations nothing fundamental had shifted.
That is worth sitting with, because it cuts against the daily noise. The problem in most companies is not that they have failed to start with AI. They have started. They have run the pilot, bought the licences and sent the all-staff email. The problem is that nothing much happened next, and very few leaders have a clear explanation as to why.
Why failing AI rollouts are almost never a technology problem
The most important finding in the MIT report is not the ninety five percent figure. It is the reason behind it. The barriers, the researchers concluded, are mainly organisational rather than technological. They describe a learning gap: companies that cannot fit AI into their actual workflows, structures and culture.
This matters because it points the finger somewhere uncomfortable. If the tools were the problem, the answer would be easy, you would simply wait for a better model or buy a more expensive one. But the same models are quietly delivering results in some organisations while failing completely in others. As Ethan Mollick, the Wharton professor and author of Co-Intelligence, has put it, the thing that decides whether AI delivers is no longer the capability of the technology or even the ability of the individual using it. It is the structure, the policy and the leadership around it.
In other words, buying a better tool will not fix a rollout that is failing for human reasons. And most of them are failing for human reasons.
Why spending more on AI does not guarantee better results
If money alone solved this, the biggest budgets would be winning. They are not.
The MIT study found that companies were putting more than half of their AI budgets into sales and marketing, the visible, exciting use cases, while the strongest returns were sitting quietly in the back office: the unglamorous work of processing documents, handling routine service and streamlining internal operations. A lot of AI spending is chasing visibility rather than value, and it shows up in the results.
There is a similar lesson in how companies get their tools. The research found that AI brought in from specialist outside vendors tended to succeed about twice as often as systems a company tried to build for itself. The usual reason is not that internal teams lack talent. It is that organisations underestimate how much of the work is integration, the slow, detailed business of fitting a tool into how people actually work, and overestimate how much is the model itself.
There is one more reason pilots stall that costs nothing to avoid, and it is the most common of all: vague goals. A great many AI projects begin without a clear definition of the problem they are meant to solve, or any honest description of what success would look like in the numbers. That makes them almost impossible to judge and very easy to quietly drop. Add in messy, scattered or locked-away data, which is the raw material every one of these tools depends on, and you have a project that was never set up to succeed in the first place. None of that is a failure of the technology. It is a failure of planning.
The pattern across these findings is the same. What matters is not how much you spend, but where you point it and how carefully you connect it to the real work.
Why your people are already using AI, just not where you can see it
Here is something that surprises a lot of leaders. While they are still debating their AI strategy, their employees have quietly got on with it.
The MIT researchers found that only around forty percent of companies had bought official AI subscriptions, yet close to ninety percent of employees said they were using personal AI tools to get work done. Mollick has a name for these people. He calls them secret cyborgs: employees who have worked out how to use AI to do their jobs faster and better, but who keep it to themselves.
This is not a small detail. It means the most valuable knowledge in your organisation, the practical understanding of what AI can and cannot do in your specific business, already exists. It is just hidden. The people who have figured out where AI genuinely helps are not telling anyone, so the company carries all of the risk of unsupervised AI use and gains almost none of the shared learning.
Why fear, not capability, is the real blocker to AI adoption
Why would someone hide a tool that makes them better at their job?
Usually because they are afraid of what honesty might cost them. If an employee reveals that AI now does most of a task in minutes, they worry the reward will be a smaller team, not a lighter workload. If they admit they used AI to write something, they worry it will be judged as cheating or laziness. And in many companies the rules themselves push usage underground, long lists of restrictions and approval committees built around everything that could go wrong, with very little said about what could go right. Mollick points out that this kind of fear-first approach is almost a recipe for shadow AI use, because people keep doing the thing, they just stop talking about it.
Underneath this sits a vision gap. Many leaders have told their people that AI is urgent. Far fewer have painted a clear, believable picture of what an AI-supported future actually looks like for the humans in the building. Into that silence, people pour their worst assumptions, and adoption stalls. This is why trust and a healthy culture turn out to be such an advantage. If your employees do not believe you have their interests at heart, they will keep their best AI work to themselves, and you will be the last to know what is possible.
What the organisations succeeding with AI do differently
The organisations on the right side of the divide are not the ones with secret access to better technology. They are the ones who lead the change as a human change, not a software upgrade.
Mollick offers a simple structure for this, built on three parts working together: leadership, lab and crowd. The crowd is everyone in the organisation, encouraged to experiment with AI on their real work rather than warned away from it. The lab is a small group of people who are genuinely good with AI, often not from the IT department, whose job is to take what the crowd discovers, refine it and spread the best of it to everyone else. And leadership is the part that makes the other two possible: setting the incentives, removing the fear, and being clear that the gains from AI will be reinvested in better work and growth rather than used to thin out the team.
The successful companies also do something the failing ones skip. They redesign the work around AI instead of bolting it onto a process that never changes. The MIT research is clear that the projects which delivered were the ones tightly woven into how the business actually runs, not the ones dropped on top as a shiny extra. That is slower and less exciting than buying a tool, and it is the whole game.
Finally, the best leaders use AI themselves. It is very hard to make good decisions about something you have never properly touched, and Mollick’s consistent advice to leaders is the least complicated thing he says: use the tools, a lot, on real work, until you understand what they can and cannot do. Strategy built on hands-on experience beats strategy built on a vendor’s slide deck almost every time.
What leaders should do now to get real value from AI
None of this requires a bigger budget. It requires a different approach. Here are five practical places to start.
1. Start with the problem, not the tool
Before buying anything, get specific about the business problem you are trying to solve and the result you expect to see. The pilots that fail tend to begin with the technology and go looking for a use. The ones that work begin with a clearly defined problem and a clear idea of what success would look like in the numbers.
2. Make it safe to use AI
Your people are already using AI. Your job is to make it safe for them to do so in the open. Be honest that productivity gains will not be used to punish the people who find them, invite employees to share what is working, and reward the ones who do. The knowledge you need is already in the building, and fear is the only thing keeping it hidden.
3. Put the money where the returns are
Resist the pull towards the visible, demo-friendly use cases and look hard at the unglamorous ones. The back-office work that nobody photographs for the annual report is often where the real savings sit. Spend on value, not on visibility.
4. Redesign the work
A tool dropped on top of an unchanged process rarely delivers. The harder, more valuable move is to rethink how the work flows once AI is part of it, and to integrate it properly into the way people actually operate. This is the slow part most rollouts skip, and it is the part that separates the five percent from everyone else.
5. Lead from use
Spend real time using AI on your own work. You cannot set a sensible direction for something you have only read about, and the leaders who experiment personally make far better calls than the ones who outsource the question to consultants or wait for a vendor to tell them how to run their company.
Why the gap between AI leaders and laggards is widening, not closing
There is a reason to treat this with some urgency, because the divide between the companies that make AI work and the ones that do not is not standing still. It is widening. The organisations that have built a culture of safe experimentation are learning what works faster, spreading it faster, and getting comfortable with each new wave of tools before their competitors have finished arguing about the last one. Their secret cyborgs come forward, their small AI-fluent teams turn good ideas into shared practice, and the advantage quietly compounds.
Meanwhile the companies on the wrong side keep running isolated pilots, keep driving their best AI users underground, and keep waiting for a certainty that is not coming. Mollick makes a sharp historical point here. A large share of the competitive edge that the best twentieth-century firms built came not from technology but from a willingness to experiment with how they managed and organised themselves. The companies that treat AI as a chance to rethink how they work, rather than a product to be bought, are likely to pull away in exactly the same way. The real cost of getting this wrong is not a wasted pilot. It is falling steadily behind the organisations that got it right.
Where this leaves you: treating AI as a leadership challenge, not a software purchase
The divide between the companies that get value from AI and the ones that do not is real, but it is not a technology divide. The failing rollouts share a pattern. They treat AI as a purchase, restrict it out of fear, bolt it onto unchanged work, and spend on whatever looks impressive. The successful ones treat it as a leadership challenge. They define the problem, build enough trust that people will tell the truth about how they work, redesign the work itself, and lead from hands-on experience.
That should be encouraging, because it means the deciding factor is something leaders can actually control. The tools will keep improving on their own. What will not improve on its own is the culture, the clarity and the trust that decide whether those tools ever turn into results.
This is what Helen Sterling speaks about as an AI keynote speaker. She helps leadership teams understand why AI delivers in some organisations and stalls in so many others, and what leading the change well actually looks like in practice.
If you are planning a conference, leadership event or company offsite where your people need a clear, honest and practical conversation about getting real value from AI, Helen Sterling can help your leaders see why most rollouts fail, and what the best ones do instead.
About Helen Sterling: AI Keynote Speaker
Helen Sterling is an AI keynote speaker and author of The Irreplaceables, winner of Best AI & Future of Work Speaker of the Year 2026 at the DKS Speaker Awards. Her work centres on why AI succeeds in some organisations and stalls in others, and what leaders need to do differently to move from pilot to real return. She helps leadership teams understand that getting AI right is a culture and strategy challenge as much as a technology one.
If you are looking for a speaker who can help your senior team understand why AI adoption so often falls short, and what to do about it, get in touch here.