What If AI Fails to Cross?
This morning, I asked students of the SUSS-IAL Master in Boundary-Crossing Learning and Leadership a question that unsettles anyone who works in learning: What if the learner isn’t the problem? I walked out with a sharper one. What if AI isn’t the problem either?
Here is my answer, and I suspect many leaders won’t like it. An organisation that cannot learn across its own boundaries will not be rescued by a technology that learns faster than it does. It will simply hit its walls sooner, and more expensively.
Every organisation I work with has an AI pilot it is proud of. A chatbot that answers HR questions. A tool that drafts marketing copy. An agent that summarises customer complaints. The demo is impressive. The steering committee nods. And then, almost always, nothing happens. The numbers confirm it. McKinsey’s November 2025 State of AI survey found that 88% of organisations use AI in at least one function. Yet only about a third have begun scaling AI across the enterprise, just 7% say it is fully scaled, and only 39% can point to any impact on enterprise profit. MIT’s Project NANDA went further, reporting that 95% of generative AI pilots stall and never reach scaled adoption. That headline figure is contested. The direction is not.
And the next wave is already lining up to repeat it. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Its analysts described most current agentic projects as early experiments driven by hype and often misapplied.
The usual explanations are technical: the data isn’t ready, the model hallucinates, the vendor over-promised. MIT’s own researchers rejected that. They said the failures were less about the models and more about how organisations try to use them.
Here is my argument. AI doesn’t fail to scale. It fails to cross. A pilot lives comfortably inside a boundary: one team, one function, one sponsor. Value lives across boundaries: between levels of the hierarchy, between functions, between the organisation and the world outside. Until your AI can cross those boundaries, it will never scale. It will just multiply.
(I write a fortnightly series on continuing education, capability development, AI, and the messy realities of organisational transformation. I’m also the author of Making S.E.N.S.E. of AI for SMEs. You can find his full archive on Medium.)
Pilot purgatory is a feature, not a bug
Here is the provocation I want to put on the table. Most organisations are not failing to scale AI. They are succeeding at not scaling it.
Think about how a typical pilot is designed. One sponsor. One function. A ring-fenced budget. A small, enthusiastic team. Clean data. Heavy human oversight. Every one of those choices has the same effect: it keeps the AI safely inside a boundary, where it cannot disturb anyone’s authority, budget or way of working.
That is not an accident. Harvard’s Robert Kegan and Lisa Lahey called it immunity to change: people and organisations hold hidden competing commitments that quietly defeat the change they say they want. An organisation can be sincerely committed to AI transformation and, at the same time, deeply committed to keeping its hierarchy, its functional fiefdoms and its decision rights exactly as they are. The pilot satisfies both. It lets you look innovative without having to change.
Pilot purgatory, in other words, is the organisational immune system working exactly as designed.
The dynamo lesson we keep forgetting
We have been here before. In 1990, Stanford economic historian Paul David published a famous paper, The Dynamo and the Computer. He asked why electricity took so long to show up in productivity statistics. Factories began electrifying in the 1880s. The big productivity gains didn’t arrive until the 1920s. The reason was not the technology. Early factories simply swapped a steam engine for an electric motor and kept everything else: the same central drive shaft, the same multi-storey layout, the same flow of work. The gains came only when a new generation of managers redesigned the factory around electricity, with individual motors on each machine, single-storey plants and work organised around the flow of materials rather than the location of the power source. That took about forty years.
Most organisations today are doing exactly what those early factory owners did. They are bolting AI onto the old drive shaft. Economists Erik Brynjolfsson, Daniel Rock and Chad Syverson describe the result as the productivity J-curve: general-purpose technologies first require large, invisible investments in new processes, skills and structures, so measured productivity dips before it rises. Organisations that refuse to make those investments never get past the dip.
BCG has been saying the same thing in consultant language for years. Its 10–20–70 rule says successful AI at scale is 10% algorithms, 20% technology and data, and 70% people and processes. Yet in too many organisations, the budget conversation runs almost exactly in reverse: months debating models and platforms, a few slides on change management at the end.
That 70% is where every boundary lives.
The five boundaries every AI rollout must cross
Chris Ernst and Donna Chrobot-Mason of the Center for Creative Leadership mapped the boundaries that shape organisational life into five types. Read them as a map of where your AI pilots go to die.
Most AI strategies are written as if only the technology needs to travel. In reality, every one of these boundaries is a border crossing with its own customs check. The two that kill the most pilots are the first two: vertical and horizontal.
Vertical boundaries: the silicon ceiling
BCG’s 2025 AI at Work survey of more than 10,600 employees in 11 countries found something every CEO should find alarming. More than three-quarters of leaders and managers use generative AI several times a week. Regular use among frontline employees has stalled at 51%. BCG called it a “silicon ceiling”.
Now look at what sits on either side of that ceiling. When leaders show strong support for AI, the share of employees who feel positive about it rises from 15% to 55%. Yet only about a quarter of frontline employees say they get that support. Only 36% feel properly trained. At the very top, the view is rosier, and wrong. In IBM’s 2026 CEO study, 86% of CEOs believed their employees have the skills to collaborate with AI. The same CEOs estimated that only 25% of their workforce actually uses it regularly. That is not a skills gap. It is a perception gap across a vertical boundary. The top believes the capability is there. The bottom knows it isn’t being used.
Get Michael Low’s stories in your inbox
Join Medium for free to get updates from this writer.
The vertical boundary leaks in the other direction too. MIT found a thriving “shadow AI economy”: employees quietly using consumer tools to get real work done while the official pilots stall. Think about what that means. The most valuable AI experimentation in your organisation is happening below the level where decisions are made, and nobody above can see it. So the vertical boundary blocks AI both ways. Top-down mandates don’t land. Bottom-up innovation never rises. And the layer caught in between, middle management, is exactly where MIT’s researchers said ownership should sit. They called for line managers, not just central AI teams, to lead the integration of AI into daily work.
There is a deeper reason the vertical boundary resists AI. Organisational theorist Jay Galbraith argued in the 1970s that hierarchy exists largely to process information: when rules can’t handle an exception, it travels up the chain until someone with enough context can decide. In that sense, the hierarchy was the original AI. It was how organisations made sense of complexity. Now a model can summarise, analyse and recommend faster than any reporting line. The informational case for many layers is weakening. Their power is not. So the vertical boundary has every incentive to hold AI at arm’s length, because AI quietly questions why it is there.
The frontline is not resisting AI. It is standing on the other side of a glass floor, waiting for someone to come down.
Horizontal boundaries: your AI is wearing your org chart
In 1967, computer scientist Melvin Conway made an observation that became known as Conway’s Law: organisations design systems that mirror their own communication structures. Siloed teams build siloed software. AI is no exception. Look at most AI portfolios and you can reconstruct the org chart from them. Marketing has its content generator. Finance has its forecasting tool. HR has its chatbot. Each is a sensible pilot. Together they change almost nothing, because value in an organisation rarely sits inside a function. It flows across them.
This is where Eliyahu Goldratt’s Theory of Constraints becomes brutally relevant. Goldratt showed that any system is limited by its bottleneck, and that improving anything other than the bottleneck is an illusion of progress. Now imagine marketing uses AI to produce ten times more campaign content. Brilliant. But legal still reviews every piece by hand, and operations still can’t fulfil the demand it creates. The bottleneck didn’t move. It just got a longer queue. Marketing reports a productivity win. The organisation sees no difference in revenue. Multiply that across every function and you have the McKinsey numbers: widespread use, minimal enterprise impact.
BCG put it bluntly. Its global AI leader, Sylvain Duranton, said real returns come when companies redesign how work gets done, not when they simply roll out tools. Redesigning how work gets done means redesigning the handoffs between functions, and nobody owns the handoffs. McKinsey’s data points the same way. Its 2025 survey found that AI high performers, the roughly 6% of companies getting significant bottom-line value, are nearly three times as likely as others to have fundamentally redesigned individual workflows. They are also three times more likely to say their senior leaders visibly own and champion AI. In other words, the winners cross both boundaries at once: they redesign the work horizontally, and they lead it vertically.
Every function wants AI. No function owns the workflow. That is the horizontal boundary in one sentence.
Why scaling AI is so hard: the boundaries are knotted together
If vertical and horizontal boundaries were separate problems, we could fix them one at a time. They are not. They are knotted together, and the knot is where scaling stalls.
Consider what it takes to scale one AI use case across a customer journey that spans sales, operations and finance.
- Every horizontal problem becomes a vertical one. The person who can change the workflow between three functions usually sits two or three levels above all of them. So a frontline fix has to climb the hierarchy, cross over and come back down. Most never make the round trip.
- Every vertical decision lands horizontally. A leadership mandate to “use AI” gets translated by each function into its own local pilot, which is exactly how you end up with Conway’s Law.
- Data belongs to everyone and no one. Customer data sits in sales, cost data in finance, process data in operations. Scaling AI means sharing it across functions, which means someone senior has to settle who owns it, who can see it and who is liable when it is misused.
- Agents cross both boundaries in a single action. A human employee who needed data from finance had to ask, and the asking was a control. An AI agent that reads a contract, checks a budget and triggers a purchase order crosses three functions and two levels of authority in seconds. Your approval structures were never designed for that.
This is not a distant scenario. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028, up from essentially none in 2024. Every one of those decisions will cross a boundary that a human used to guard.
Then the other boundaries pile on. Your workflows increasingly run on models you don’t own, crossing the stakeholder boundary with every API call. And across Southeast Asia, a use case that works in Singapore must cross geographic boundaries into markets with different languages, data rules and expectations of authority. So when someone tells you their AI pilot “just needs to be scaled”, ask them a simple question: how many boundaries does it have to cross, and who has the authority to open each one? If they can’t answer, it isn’t a scaling plan. It’s a hope.
Three shifts that let AI cross
In my recent talks on scaling AI, I’ve argued that organisations need three shifts. Each one is really a decision to cross a boundary deliberately rather than hope AI drifts across on its own.
1. From AI to Work. Stop asking “where can we use AI?” Start asking “how should this work be done now?” That question forces you to look at the whole workflow, across functions, instead of the tool inside one of them. It is how you cross the horizontal boundary.
2. From Task to Organisation. A task-level pilot saves a few hours. An organisation-level redesign changes roles, decision rights, handoffs and incentives. That needs leaders with the authority to rewrite how levels and functions work together. It is how you cross the vertical boundary.
3. From Pilot to Transformation. A pilot proves something can work. A transformation makes it the way work is done, with owners, governance, measures and capability that outlast the demo. It is how you stop multiplying pilots and start compounding value.
Underneath all three is the discipline of my S.E.N.S.E. framework. Spot the real bottleneck, not the most visible pain. Empathise with the people on every side of the boundary, especially the frontline nobody consulted. Narrow to the workflow where crossing creates the most value. Build the Solution across the boundary, not inside it. Evaluate at the enterprise level, not the function level.
The question I’ll leave you with
The truth is this — most organisations don’t have an AI scaling problem. They have a boundary problem that AI has exposed. The technology is ready to cross. The organisation is not. So try this. List every AI pilot you are running. Next to each one, write down how many boundaries it would have to cross to create enterprise value, and the name of the person who owns each crossing.
If most of those names are blank, you don’t need another pilot. You need leaders who are willing to open the borders.
