Transformation does not die in a dramatic moment. There's no single meeting where it collapses, no one bad decision you can point to afterwards and say, that's where it went wrong. I've spent enough time in the industry to know, it's not one failure, it's several small ones running quietly in parallel, none of them loud enough on their own to trigger alarm.
A visible failure gets fixed because someone has to explain it. A silent one just accumulates, and by the time you notice, you're not dealing with a problem, you're dealing with a culture. Around 70% of AI transformation failures come down to human and process factors, not technical ones. Leaders who treat AI as a technical project are, almost by definition, underestimating the organisational redesign the work actually requires.
Here are the most common ones.
No Shared Vocabulary
Early on, half the room means simple automation when they say AI. The other half means something closer to an agent making its own decisions. To put it simply, we think we’re talking the same language, but we aren’t. Everyone assumes they're picturing the same thing. Strategy conversations happen, decisions get made, and months later two teams discover they built for completely different ideas of what "the AI thing" was meant to do.
It's sneaky because it doesn't look like a communication problem while it's happening. On the surface it looks like an alignment. Everyone nodded in the meeting.
Pilot Purgatory
A pilot gets built, it works, everyone's pleased, and then it just... stays a pilot. Nobody defined what "done" looks like, or what it would take to move from demo to production, so there's no natural next step, only a slow fade in attention as the next shiny pilot takes its place.
The uncomfortable thing about pilot purgatory is that it can look like progress for a surprisingly long time. You can point to a growing list of pilots as evidence of momentum, right up until someone asks how many of them are actually being used and by who, and the room goes quiet.
Incentive Misalignment
People are still being measured on the old way of working while being asked to adopt the new one. Adoption becomes a personal cost with no corresponding personal benefit, at least not one that shows up on a performance review. So the rational move, if you're the person doing the work, is to quietly keep doing things the way you always have, while nodding along in the meetings about transformation.I think it's just people responding sensibly to what's actually being rewarded, which is often not what leadership thinks is being rewarded.
Ownership Vacuum
Everyone wants to build. Almost nobody wants to be the one who maintains what got built after the initial excitement fades. So things ship, they work for a while, and then they quietly degrade, because the assumption baked into an agent or a workflow six months ago is no longer true, and nobody's job is to notice that.
This is one of the ones I find most avoidable, and also one of the most common, because ownership is unglamorous in a way that building never is.
Fragmentation and Silos
In the absence of a sanctioned path, teams find their own. One group adopts a tool, another group adopts a different one, and neither knows what the other is doing, because nobody asked and nobody's tracking it. This isn't really about shadow adoption though, it's about fragmentation. Silos that were never aligned to begin with just carry that same lack of alignment into how they each approach AI, so the organization ends up with several disconnected, incompatible versions of "adoption" instead of one.
And alongside the fragmentation sits a larger group that isn't adopting anything at all, quietly resistant to changing how they work regardless of what's sanctioned or not. So leadership can point to pockets of activity and call it progress, while the actual picture is a handful of misaligned silos moving in different directions, and a much larger group not moving at all.
Investing Ahead of Learning
Resources get committed before the underlying hypothesis has actually been tested. A big bet gets made on an approach, a tool, a rollout, because it feels decisive, and momentum matters, and waiting for evidence feels slow. But committing ahead of validated learning doesn't just risk the investment, it destroys value even when the eventual answer turns out fine, because the org has now built process, headcount, and expectation around something nobody actually confirmed would work.
Funding, if it's going to mean anything, has to stay aligned on data rather than conviction. That's a harder discipline than it sounds, because conviction is what gets budgets approved in the first place, and asking for evidence before committing can look like hesitation rather than rigor.
Stakeholder Communication Gaps
The people doing the work understand what's actually possible. The people funding and sponsoring the work often have a picture shaped by whatever they last read or saw demoed elsewhere. Nobody is deliberately misleading anyone, but the gap between those two pictures widens quietly, and it's rarely closed until a mismatch in expectations forces an uncomfortable conversation.
In conclusion
None of these, on their own, would kill a transformation. That's the point. They're each survivable in isolation. What makes them dangerous is that they tend to show up together, quietly reinforcing each other, and none of them announce themselves the way a failed project does.
The fix isn't a bigger tech budget or a more ambitious roadmap. It's treating these patterns as the actual risk register, not an afterthought to it. That means naming a vocabulary before strategy conversations start, not after two teams have already built for different definitions. It means setting a success and graduation criteria for every pilot before it starts, so "still a pilot" has to be justified rather than just defaulting to true. It means assigning ownership at the point something ships, not after it's already degraded.
Organisations that get AI transformation right aren't the ones with the most sophisticated models, they're the ones that treat the human and process side with the same rigor they'd apply to the technical side.
Thought pondered by Sarah exploring the intersection of AI, creativity, and human wellbeing