Performance usually falls before it rises in an AI rollout, and almost no organisation says so in advance. Mike Raven on why the dip happens, why unlearning has to come before re-skilling, and how to measure whether your people can absorb the change.
Performance usually falls before it rises in an AI rollout. People let go of ways of working that used to work before the new ones are reliable, so for a while they are competent at neither, and the numbers show it. Almost no organisation says this out loud in advance, which means a normal phase of the transition gets read internally as a failing programme.
I spoke about this at AI Experience Live in London in June 2026, to a room of HR and transformation leaders. I asked who had openly admitted inside their organisation that things would get worse before they got better. Two or three hands went up. Then I asked who believed it was true. Almost every hand in the room. That gap, between what people know and what they are willing to say, is costing AI programmes more than any technology decision.
The dip is the stretch of a transformation where performance sits below where it started. It is not evidence the programme is wrong. It is what happens when people are asked to drop a method that still works and pick up one they are not yet good at.
Economists have a name for the measured version of this. Brynjolfsson, Rock and Syverson described the productivity J-curve in their 2020 NBER working paper The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, arguing that when a general purpose technology arrives, organisations have to build complementary intangible assets, new processes, new business models and new human capability, and that work is largely invisible in the numbers. The result is that measured productivity growth is understated early and overstated later. Put plainly, the hardest and most valuable work in an AI transition is the work your dashboard cannot see.
Every serious change has a version of this. We just rarely plan for it. The business case goes to the board with a line that rises from day one, reality arrives with a sag in the middle, and everyone starts looking for someone to blame. If the dip had been in the plan, those same weeks would have been read as progress.
Because nobody volunteered for this. When I ask a room who actually asked to own the AI rollout, I almost never get a hand. It lands on people, usually in HR, people operations or transformation, who are already carrying the wellbeing load of the organisation and are now holding a technology programme they did not design and cannot fully control. Admitting it will get worse before it gets better feels like handing critics a weapon.
It is the opposite. Naming the dip is what buys you the room to get through it. Once a leadership team has agreed in advance that a drop is coming, the drop stops being evidence against the programme. Without that agreement, every normal wobble becomes a referendum on whether this should have been started at all.
The sequence that works is unlearn, then redesign, then re-skill. Most programmes go straight to the third step, because re-skilling is the one with a budget line and a vendor attached.
Unlearning is the part almost nobody measures. When I ask who is actively addressing it, as opposed to re-skilling, it is usually one or two people in a room of a hundred. That matters, because the habits being dropped are often the ones that earned people their reputation. Asking a high performer to abandon the method that made them a high performer is not a training problem. It is an identity problem, and it needs handling as one.
The data says this is getting harder. AQai analysed 19,077 first AQ assessments taken between 2020 and 1 August 2026 and found that Unlearning, the ability to deliberately let go of what you know, has fallen by about 1 point a year since 2022, while Mental Flexibility stayed flat. The decline holds when job level, age group and gender are held constant. The clearest way to feel the size of it: the typical score today would have ranked in the bottom 40% of the population 4 years ago. We set out the full analysis in What Is Your Organisation Still Doing Because It Worked?
The useful part for anyone running a rollout is what the decline is not. It is not workload. Work stress averaged 50.7 for the low-Unlearning group against 47.0 for the high group, a relationship close to zero. Uncertainty about change was a different story: 38.5 for the low group against 23.6 for the high group. People who are unsure where the change is going find it much harder to let go of what they have. Which brings the argument back round, because naming the dip is one of the few things a leadership team can do that directly reduces that uncertainty.
That pattern shows up in rollouts too. The deployment completes, the training gets delivered, and people carry on working the way they always have, which is the problem we wrote about in why AI adoption is failing.
Here is where it gets uncomfortable. A lot of what AI removes first is busy work, and busy work has quietly been doing a second job for years. Being busy is how many people know they are useful. Take that away and you have not only saved time, you have opened a question about purpose that the organisation now has to answer.
So the dip is not only a productivity story. It arrives in organisations where a serious share of people are already running on a fried nervous system, and it adds uncertainty about role and worth on top. If a work redesign has no wellbeing conversation inside it, it will stall, and nothing on the dashboard will explain why.
The relationship has to change. Managing a team through this the way you would direct a new starter does not work, for the same reason parenting a teenager the way you parented a 5-year-old does not work. With a young child you tell them what to do and they rely on you for it. Try that with a teenager and you get resistance, and you have earned it. What works is standing beside them and asking what they think.
That shift, from manager to coach, is most of the leadership work in an AI transformation. It comes with a warning: teenagers detect inconsistency instantly, and so do teams. If you are asking people to experiment, drop old habits and sit in discomfort while you personally do none of those things, they will notice before you have finished the town hall.
There is a measurable reason leaders misjudge this. Across 16,491 AQ assessments gathered between 2020 and 2025, executives, VPs and senior managers averaged 74.7 on AQ Mindset while entry-level and early-career employees averaged 64.1, a gap of 10.6 points on a 0 to 100 scale. It is a cross-sectional association rather than proof that seniority raises Mindset, and part of it reflects age, as our research note on the seniority gap sets out. The practical consequence is a perception gap: leadership reads a readiness for change that the frontline does not feel. Treating that caution as resistance is the fastest way to deepen the dip you are trying to climb out of.
Start small and slightly silly, because silly gets remembered. Ask your team to brush their teeth with the other hand for a week. It is clumsy, it is worse than the way they have always done it, and that is the point. They get to feel the dip somewhere nothing is at stake, which makes the real one easier to name later.
Then do the serious version: an everyday audit. Have people list what they do every single day and identify what can be removed. In the C-suite groups I work with, once there is enough safety in the room, leaders will admit that a large share of their work is not especially useful and a small share carries nearly all the value. Acting on that admission moves unlearning faster than any course.
One caution from our own data: unlearning does not respond to effort. It travels with resilience, hope and comfort with not knowing, far more than with grit or hours worked. You cannot make someone unlearn by asking them to try harder. What you can do is make it safe to be wrong in front of colleagues, which is the condition most of this depends on.
Measure it. For most of the last decade, "are our people adaptable enough for this?" was answered with a gut feel. It is now a question you can answer with data, which is what AQ, or Adaptability Quotient, is for.
AQ is measured through the A.C.E. model, which has four parts. Ability covers the adaptability skills that can be developed, including unlearning, mental flexibility, grit and resilience. Character covers why a person adapts, or why they hold still. Environment covers the conditions around them, the team support, the culture and the safety that decide whether adapting is worth the risk. Advanced metrics carry readiness signals on top, including change uncertainty and re-skilling capacity. We explain how it works in The AQ Model & Science.
The gap between agreeing this matters and acting on it is now documented. Deloitte's 2026 Global Human Capital Trends, run with Oxford Economics across more than 9,000 business and HR leaders in 89 countries, found that 85% say building the organisation's and workforce's ability to adapt is critical, 74% say they are making some kind of progress, and only 7% say they are leading in helping their workforce continuously grow and adapt. Almost everyone knows. Almost nobody is ahead.
What measuring gives you is a map instead of a guess. In one engagement run by an AQai certified partner, a leadership team going through an AI-driven restructure assessed 17 people and found 47% scoring high on Mindset while 47% scored low on Change Readiness, which is a very specific problem with a very specific answer. You can read it in our SaaS leadership case study. Without the assessment, that team would have concluded it had a willing group, pressed ahead, and met a dip nobody had planned for.
Commonly, yes, for a period. The fall comes from people letting go of established ways of working before the replacements are reliable, so capability briefly drops in both. Brynjolfsson, Rock and Syverson's productivity J-curve (NBER, 2020) describes the measured version: the complementary work of building new processes and capability is largely invisible in productivity figures, so early gains are understated.
Unlearning is letting go of a way of working that used to succeed. Re-skilling is building the capability that replaces it. Unlearning has to come first, because people cannot adopt a new method while the old one is still what they reach for under pressure. Most transformation budgets fund only the second step.
Because adoption is a human capability question. Across 19,077 first AQ assessments from 2020 to August 2026, AQai found Unlearning declining by around 1 point a year since 2022, and that the decline tracks with uncertainty about change rather than workload. If people are unsure where the change is heading, they hold on to what they know, and the new tooling goes unused.
Through an adaptability assessment. AQai measures AQ using the A.C.E. model, covering Ability, Character and Environment, plus advanced metrics for signals including change uncertainty and re-skilling capacity. That produces a map of where adaptive capacity sits across individuals, teams and the wider system, rather than an assumption that everyone is equally ready.
In practice it usually lands on HR, people operations or transformation, often without anyone volunteering. Whoever owns it needs a leadership team that has agreed in advance that a temporary dip is expected, otherwise the owner carries the blame for a normal phase of the transition.
If you are about to run a work redesign and you do not know how much change your people can absorb, you are guessing, and most organisations currently are. Map adaptive capacity across your leaders and teams before the dip arrives rather than after it.
Start with an AQ certified partner who can run this with your leadership team.
Copyright adaptai Ltd. All rights reserved. AQ® and AQai® are registered trademarks of adaptai Ltd.
.png){kind=small}
.png)
