How Do I Get AI Adoption to Work After Previous Transformations Failed?
By refusing to make it look like the last one. Change fatigue is not an attitude problem in your workforce; it is manufactured by prior change history, and the research says each newly announced major change increases it. Your people's scepticism is also rational: fewer than 30 per cent of transformations succeed on McKinsey's own survey data. So do not announce a transformation. Start with one leader who owns a real, painful problem, solve it visibly with the machine as a thinking partner, and let the result recruit the next team. Results do not trigger the fatigue response. Announcements do. The deeper obstacle is that after several failed programmes nobody will admit what they do not know, and that is the specific thing the Havruta Methodology's Flip is built to dissolve: it makes the machine ask, so the gap surfaces without an audience.
The room remembers every programme you ran
Every AI initiative in an established organisation launches into a room with a memory. The people in it have been through the operating-model redesign, the digital programme, the agile rollout, and each one taught the same lesson: a big change is announced with conviction, consumes energy for a year, and the work ends up roughly where it was. The reflex that greets your AI announcement, wait quietly and this too shall pass, is not cynicism. It is trained behaviour, and the training was thorough.
The research on this is unambiguous about where fatigue comes from. A study inside a European financial institution found significant relationships between the number of reorganisations employees had experienced and their level of change fatigue, with prior reorganisations, uncertainty and workload together explaining 43 per cent of the variance, and it states the operational consequence plainly: each time a new major organisational change is planned, change fatigue among personnel increases (de Vries and de Vries, 2021). Fatigue is not a trait some workforces have. It is a stock that every announced programme adds to, including the one you are about to announce.
Which means the standard launch playbook, the town hall, the branded programme name, the cascade of champions, is not neutral packaging. It is the exact stimulus the fatigue is conditioned on. Run it again and you collect the conditioned response, before a single tool is opened.
The scepticism is rational. Respect it
It helps to concede, out loud, that the sceptics have the numbers. McKinsey's Global Survey research on transformations found that fewer than 30 per cent succeed, that only 16 per cent both improved performance and equipped the organisation to sustain the change, and that in traditional industries the sustained success rate falls to between 4 and 11 per cent (McKinsey, 2018). A side note worth knowing: the endlessly repeated claim that exactly 70 per cent of change initiatives fail has no reliable empirical basis, so do not build your case on it. You do not need it. The real base rate is bad enough.
Seen against that base rate, an employee who declines to spend enthusiasm on the fourth announced programme is making a sound bet. They are applying the base rate you gave them. And this reframing matters practically, because leaders who read the quiet room as resistance reach for more communication, more sponsorship, more launch, which is more of the stimulus that produced the quiet. Leaders who read it as rational pattern-matching reach for the only thing that updates a rational prior: evidence.
Your first job, then, is not to persuade anyone. It is to produce one piece of evidence the room cannot argue with, from inside its own walls.
What fatigue does to AI in particular
AI adds a twist the earlier programmes did not have: it arrives wrapped in a fluency assumption. Everyone is presumed to be using it already, competently, so admitting confusion feels like falling behind a peer group that is, in truth, mostly bluffing too. The social mechanics here were mapped long before AI. Edmondson's foundational research on teams showed that admitting an error or asking for help risks making a person appear incompetent, so people withhold exactly the disclosures the team needs to learn, and the withholding is strongest where the stakes of image are highest (Edmondson, 1999). In a senior room, "I don't know how to use this properly" is a sentence almost nobody will say.
The 2026 data shows how far underground this drives things. A survey of 1,250 office professionals found that 66 per cent of those who have used AI at work did so while believing it was not permitted, 39 per cent would rather use AI without telling anyone, and a third would hide it specifically to avoid scrutiny from managers (PagerDuty, 2026). Read that carefully: the learning is happening, at scale, in secret. People are experimenting alone precisely because experimenting in view feels unsafe, in organisations where the last transformation taught them that visible struggle gets noticed unkindly.
So a fatigued organisation does not merely under-adopt. It hides its adoption, hides its gaps, and starves itself of the shared learning that would compound. Any approach that works here has to give people a way to close their gaps without an audience.
The Unjudged Room
Here is the asymmetry that changes the game, and we have given it a name inside the Havruta Methodology: the Unjudged Room. In any human room, "I don't know" carries a social cost, and the cost rises with seniority, which is why the most senior people admit the least. A machine cannot judge. It holds no opinion of you, reports to no one, and remembers no embarrassment. By rights it should be the safest place a leader ever admits a gap, and yet, used as a vending machine, it is where people admit the least, because a vending machine never asks. You type the request, it produces, and the moment for the admission never arrives.
What activates the room's safety is the Flip: instructing the machine to question you before it answers. The moment the machine asks "what is the approval process here?" or "which figure is this based on?", something quietly useful happens. "I don't know" stops being a confession and becomes an instruction, a line that names the missing input and points at where it lives. No colleague heard it. No image was spent. The gap got named, and named gaps get closed.
In the room, "I don't know" costs image. With the machine, it costs nothing, and buys the missing input.
To be precise about the claim, because precision matters here: the machine is not a confidant, a therapist, or a substitute for colleagues, and we never present it as one. The claim is narrower and stronger. It is the one working surface where admitting a gap has no social price, and for a leadership cohort trained by failed programmes to never show a gap, that surface is where the honest learning restarts.
What to do differently this time
Everything above converges on a short list of inversions of the standard playbook. Each one exists to avoid triggering the fatigue the previous programmes installed.
Deliver before you announce. The failed programmes promised first and delivered later or never; that sequence is what fatigue is made of. Invert it. Find one leader whose problem is painful enough to matter, solve it with them using the machine as a thinking partner, and say nothing organisation-wide. The story of the lightened workload travels by itself, and a story that arrives as gossip is trusted in a way no town hall has been for years.
Pick owners, not cohorts. Do not ask which teams should be trained next. Ask which leaders own a problem painful enough to build a habit on. Teams do not adopt methods; they adopt results, and a result needs an owner who bleeds for the problem. One committed owner with a real bottleneck outperforms a cohort of twenty conscripts, and the conscripts join later on better terms, as volunteers.
Make the practice the day job. Transformations fail partly because they ask for new behaviour on top of the existing workload, so the change competes with the work and loses. The Havruta approach works on the day job itself: the sessions run on the team's live decisions and live documents, so the hours spent practising are the same hours the work needed anyway. Nothing extra to sustain means nothing extra to abandon.
Protect the private ramp. Since the evidence says people learn AI in hiding, give them a legitimate hidden path: the machine itself, with the Flip installed, as the place to close gaps without an audience. What you never do is turn early stumbling into content for a dashboard. Measure the programme by results and voluntary return, per the honest adoption readings, not by surveilling who asked the machine what.
None of this is slower than the launch-and-cascade playbook, which is the objection I hear most. The cascade only looks fast because announcements are instant; measured by when the work actually changes, the quiet sequence wins, and it wins without spending the credibility you have left. After three failed transformations, that credibility is the scarcest resource in the building. Spend it on results, one leader and one room at a time.
Frequently asked questions
How do I get AI adoption to work after previous transformations failed?
By refusing to make it look like the previous transformations. Change fatigue is manufactured by prior change history: research inside a European financial institution found the number of reorganisations employees had lived through significantly predicted their fatigue, and that each newly announced change increases it (de Vries and de Vries, 2021). So do not announce a programme. Start with one leader who owns a real problem, solve it visibly with the Havruta discipline, and let the result recruit the next team. Results do not trigger the fatigue response. Announcements do.
My team has change fatigue, how do I introduce AI?
Quietly, and through work rather than messaging. A fatigued team has learned from experience that initiatives cost energy and rarely pay it back, so any launch framing spends credibility you no longer have. Pick the team's most painful piece of owned work, sit with them and solve it using the machine as a thinking partner, and stop there. No branding, no cascade, no certification. When the neighbouring team asks how the work got lighter, adoption has started, and it started without an announcement for the fatigue to attach to.
Why do people stay silent about what they don't know about AI?
Because admitting a gap has a social price, and the price rises with seniority. Edmondson's foundational research showed that by admitting an error or asking for help, a person risks appearing incompetent, so people withhold exactly the disclosures that would help the team learn (Edmondson, 1999). AI gaps are today's sharpest case: everyone is assumed to be fluent already, so nobody asks the basic question. The evidence people even hide their AI use, with 66 per cent having used it while believing it was not permitted (PagerDuty, 2026), shows how much learning is being driven underground.
What is the Unjudged Room?
The Unjudged Room is a concept inside the Havruta Methodology. In any human room, saying "I don't know" carries a social cost, and the cost is highest in the most senior rooms, so leaders suppress the admission. A machine cannot judge, so it should be the safest place a leader ever admits a gap, yet used as a vending machine it never asks, so the moment never arrives. The Flip, instructing the machine to question you first, is what activates the judgement-free property: the admission becomes an instruction that names the missing input. The claim is the absence of social cost, not the machine as a confidant.
Why do most transformations fail?
The honest figure is that fewer than 30 per cent of transformations succeed, and only 16 per cent both improve performance and sustain it, per McKinsey's Global Survey research; in traditional industries the sustained success rate falls as low as 4 to 11 per cent (McKinsey, 2018). The oft-quoted claim that exactly 70 per cent fail has no reliable empirical basis and is best avoided. What the real numbers say is enough: the workforce's scepticism about the next announced programme is not resistance, it is pattern recognition.
How do I make an AI transformation feel different from the last one?
Make it produce before it promises. The previous transformations announced first and delivered later or never, which is the exact sequence that manufactures fatigue. Invert it: deliver a visible result on one team's real work first, at small scale, and let the story travel on its own. Also change what the effort asks of people: a transformation demands new behaviour on top of the day job, while the Havruta approach works on the day job itself, so the practice and the work are the same hours, not competing ones.
References
- de Vries, M. S. E., & de Vries, M. S. "Repetitive reorganizations, uncertainty and change fatigue." Public Money & Management 43(2), 2021.
- McKinsey & Company. "Unlocking success in digital transformations." McKinsey Global Survey, October 2018.
- Edmondson, A. "Psychological Safety and Learning Behavior in Work Teams." Administrative Science Quarterly 44(2), 1999.
- PagerDuty / Wakefield Research. "Shadow AI Is Happening Within Your Organization." June 2026.