You can’t turn a submarine crew into astronauts by bolting the submarine to a rocket. Both jobs involve a sealed metal tube, recycled air and long weeks without daylight, which is why the comparison is tempting. Everything that matters is different: the training, the emergencies, the instinct for when to stay silent and when to call home. Launch the sub and you get brave, highly skilled people in the wrong vehicle, drilled for the wrong disasters.
That, in my view, is why digital transformations fail so often. The crew, the incentives and the tools stay exactly as they were, and only the label changes. In the 2010s the label was “digital.” In 2026 it is “AI,” and the forms are familiar: a transformation task force staffed by the people who ran the last one, an acquired startup whose people walk out faster than anyone else, and an innovation lab whose reports nobody upstairs reads.
None of this is inevitable. But the fixes are not the ones most companies reach for, and they are less comfortable. Here is what the evidence says about why transformations stall, and what I think works instead.
Do 70% of transformations really fail?
You have heard the number. Every keynote about change says that 70% of change programs fail. In 2011 Mark Hughes went looking for the evidence behind it in the Journal of Change Management (Hughes, 2011). He traced the claim through its five most prominent sources, including Hammer and Champy, John Kotter, and articles from McKinsey and Bain. In none of them did he find valid and reliable empirical evidence for the 70%.
That doesn’t mean transformations go well. It means the famous figure is a slogan, and anyone who quotes it deserves the usual critical thinking questions: how do you know, and who benefits? The better numbers are narrower. In 2020 BCG surveyed 825 senior executives and combined the answers with its own work at 70 companies. Its conclusion was that 70% of digital transformations fall short of their objectives. Falling short is not the same as failing, and a consultancy that sells transformation help has a stake in how hard transformation looks. Still, the direction is consistent.
The AI round looks similar so far. In BCG’s October 2024 survey of 1,000 executives in 59 countries, 74% of companies had yet to show tangible value from AI. The detail I find more telling is where the trouble came from. About 70% of the obstacles were people and processes, 20% were technology and only 10% were the algorithms. A year later, in a survey of 1,250 executives, BCG put 60% of companies in the group reporting minimal revenue and cost gains from AI, and only 5% in the group getting substantial value.
So the vehicle is rarely what breaks. What breaks is the crew and the orders it is given, which brings me to the three standard plays.
Why the AI task force changes nothing
Here is the standard play. The board asks what the company is doing about AI. The CEO sets up a task force and fills it with the heads of every existing department, because leaving anyone out would be political. The task force commissions pilots. Each pilot is run by the department whose work it might replace. Nobody’s bonus depends on whether a single pilot ever reaches customers.
I have nothing against committees. They are good at protecting what already exists, which is what they were built for. The trouble starts when you ask a body designed to defend the old business to take it apart. The management scholar Steven Kerr gave this pattern its best name in 1975, in the title of a paper in the Academy of Management Journal: On the Folly of Rewarding A, While Hoping for B. Organizations hope people will take long-term risks and then pay them for this quarter’s numbers. People are not stupid. They do what they are paid to do.
So the task force reports activity instead of change: pilots launched, staff trained, prompts written, a center of excellence opened. That is participation, not engagement, and it tells you nothing about whether the business works differently. A crew rewarded for staying hidden will stay hidden. If your managers are paid on the margin of the product the new thing would cannibalize, the new thing gets a polite pilot, a steering group and a slot in the next quarterly review. Then it gets nothing.
Why acquired startups lose their people
The second play is to buy the future. If your own people can’t build it, buy a startup that already has, and hope its culture rubs off.
The evidence on what happens next is unkind. J. Daniel Kim used US Census Bureau data on about 4,000 high-tech startup acquisitions between 1990 and 2011, covering roughly 350,000 employees. In the first year after a deal, 33% of acquired workers left, against 12% of regular hires with similar skills and experience. Acquired workers were also more likely to start companies of their own, many of them competing with the firm that had just bought them. Kim’s explanation is simple. People who chose to work at a startup didn’t choose the acquirer, and nobody asked them.
The published version of the study, in Strategy Science in 2024, adds the part I find most useful. The turnover gap was smaller when the founders stayed with the acquired team, and when the acquired company was kept structurally separate instead of being folded into the parent. Integration, the thing every acquirer’s playbook calls for, is part of what drives the people out.
Founders leave too. Facebook bought Instagram in 2012. In September 2018 its founders, Kevin Systrom and Mike Krieger, left after clashes with Mark Zuckerberg, Bloomberg reported, as he became more involved in how the app was run.
The more instructive case is Walmart. It paid $3.3 billion for Jet.com in 2016 and then did something unusual: it put Jet’s founder, Marc Lore, in charge of its US e-commerce business. By the time he retired in January 2021, that business brought in more than $25 billion in sales. Jet.com itself had been shut down in 2020. I read that as a deal that worked because Walmart treated it as a hire with real authority and was willing to retire the startup it had paid for. Most acquirers do the opposite. They keep the logo and lose the people.
The innovation lab nobody listens to
The third play is the lab: a separate team, often with better coffee, whose job is to imagine the future. Sometimes it does. That is when the trouble starts.
In 1975 a 24-year-old Kodak engineer named Steven Sasson built the first digital camera. It took 23 seconds to write a 100 by 100 pixel black-and-white image to a cassette tape. He demonstrated it to executives from marketing, the technical side and the business side, and, as he told The New York Times in 2015, they were convinced nobody would ever want to look at their pictures on a television set. Asked how long it would take digital to compete with film, he estimated 15 to 20 years. In 1989 he and a colleague built a digital SLR, and he was told Kodak could sell it but wouldn’t, because it would eat into film sales.
Sasson’s own reading of those early meetings is the most honest remark about corporate incentives I know. Executives being told about something two decades away don’t expect to be at the company when it arrives, he said, so they don’t get excited. Kodak kept the patent, which earned it a fortune in licensing until it expired in 2007. The company filed for Chapter 11 bankruptcy protection in January 2012.
The lab did its job. The organization around it did what it was paid to do, which was sell film. A lab that reports to the people it threatens is a museum of the future, open by appointment.
What actually works in a transformation
The organizations that pull it off tend to do four unglamorous things. The best documented example I know is not a company but a government.
Hire people who have done it before
In 2011 the British government built the Government Digital Service and brought in Mike Bracken to run it. He came from Guardian News & Media, where he had been responsible for the consumer technology of the Guardian, the Observer and their website. That is not a career civil servant sent on an agile course. It is someone who had already done the job somewhere else.
Give them authority, not a seat at the table
GDS sat in the Cabinet Office, and through spend controls its advisers assessed departments’ digital and IT business cases before the money went out. A team that can stop a budget gets listened to. A task force that can only recommend gets thanked.
Retire things
On October 17, 2012, GOV.UK replaced Directgov and Business Link, the two main government websites. A year later the government said the switch was saving up to £50 million a year. Transformations that only add things (a lab, a chatbot, another dashboard) leave the old organization intact and slightly more expensive. Switching something off is the most convincing proof that anything changed.
Change what people are paid for
Kerr’s point is the cheapest to act on and the hardest to accept. If a division head is measured on this year’s margin, she will defend this year’s margin, and by her own lights she will be right. In BCG’s 2025 survey, the small group of companies getting real value from AI was led from the top, ranked its AI projects by expected value and tracked the results closely. That is a description of incentives and accountability, and none of it can be bought from a vendor.
My own trade, market research, has run every play in this post, usually with the same people selling the same studies to the same buyers under a new name. I describe where that leads in You killed market research.
My take
I don’t think the crew is the problem in the sense of being bad at its job. Submariners are superb at being submariners. The problem is the admiral who believes a new mission statement turns them into something else, and the board that applauds the slide.
If you are running an AI transformation in 2026, here is the test I would use. Name the person in charge and ask what she has built before. Ask what she can stop without asking permission. Ask which product, report or team will be switched off if the work succeeds. Ask how the bonuses of the ten most senior people will change. If the answers are “nobody in particular,” “nothing,” “none” and “they won’t,” you have a submarine on a launch pad.
And when you hire people who have done it, check that they really have: in March 2024 Pew Research Center reported that 12% of opt-in poll respondents under 30 claimed to be licensed to operate a nuclear submarine, though the real share of Americans with that license rounds to zero (more in Movin’ on up).
Questions people ask
Why do digital transformations fail?
Usually because the people, incentives and tools stay the same while the label changes. In BCG’s 2024 survey of 1,000 executives, about 70% of the obstacles to getting value from AI came from people and processes, and only 10% from the algorithms.
Is it true that 70% of transformations fail?
Not as an established fact. A 2011 review in the Journal of Change Management found no valid empirical evidence behind the popular 70% figure. BCG’s 2020 survey found that 70% of digital transformations fell short of their objectives, which is a softer claim.
Do employees leave after their startup is acquired?
More often than ordinary hires do. A study of about 4,000 US tech startup acquisitions found that 33% of acquired workers left within the first year, against 12% of comparable regular hires.