Dear Reader,
The misdiagnosis
McKinsey’s 2025 Superagency in the Workplace study asked C-suite leaders of large organisations how many employees were using generative AI for at least 30% of their daily work. The executives estimated 4%. The employee survey put the actual figure at around 13%. Employees are roughly three times further into AI adoption than their leaders believe.
This is not the gap most enterprise AI programmes are trying to close. Most are trying to close a knowledge gap by buying licences, running training, and measuring course completion. Salesforce found in 2024 that employees achieved above 90% compliance scores in AI training and then applied roughly a third of the protocols in real work. They knew the material but did not change the behaviour.
This pattern is not new. ERP, CRM and BPM rollouts hit the same wall over the past three decades: trained users, unchanged behaviour, shadow workflows running alongside the system. Change management exists as a discipline because tool-literacy and behaviour change are different problems.
What is different about AI is the depth of the change. ERP changed where you entered the data. AI asks you to change how you think: when to trust the output, when to override your own judgment, what work is yours versus the model’s. That is a cognitive habit change, not just a workflow change. Cognitive habits move slower than workflow steps, which is why “we trained them” produces even less adoption here than it did for ERP.
The evidence on training fundamentals is unflattering. CEB/Gartner’s Metrics that Matter programme reports a scrap learning rate of around 45% (content delivered but never applied on the job). The peer-reviewed transfer-of-training literature has been consistent for forty years: Baldwin and Ford (1988), Blume et al. (2010, meta-analysis of 89 studies), Hughes et al. (2020), and Salas et al. (2012) all converge on one finding. Work environment factors (manager support, peer support, opportunity to apply, accountability) dominate training-content factors in predicting whether learning transfers to behaviour. Most enterprise AI programmes invest in the content factors and ignore the environment factors.
Training answers the question, what do they not know? Effective adoption needs to answer a slightly different question: why don’t they behave differently, even when they know?
Why AI is cognitively different
Previous IT systems changed procedures; the cognitive work (judging risk, sensing when something is off) stayed with the human. AI moves into that layer. The AI-native buyer reasons about supplier risk and margin trade-offs with an assistant, then chooses which outputs to trust. The job becomes exercising judgment over an intelligent system whose reasoning cannot be fully audited. Three problems follow that training cannot solve.
The first is automation bias. Users delegate cognitive work to AI even when its outputs are wrong (Issue #33 covered the design dimension in detail). The 2025 cognitive offloading research from the Swiss Business School (DOI 10.3390/soc15010006) frames it as the difference between AI as a cognitive partner (the human retains metacognitive direction) and AI as a cognitive substitute (the human delegates reasoning); the two produce opposite learning outcomes. Heavy offloading suppresses analytical processing. Training cannot fix a design pattern that pushes users toward blind acceptance.
The second is expertise devaluation. Senior employees derive status and identity from accumulated expertise. AI commoditises some of that knowledge: a junior employee with strong prompting can match a twenty-year veteran in some domains. The veteran resists AI because it threatens something real. Liu and colleagues, in a 2025 two-wave survey of 311 workers, found that AI awareness predicts knowledge-hiding behaviour through psychological resource depletion. Workers who feel threatened protect existing expertise rather than apply new skills.
The third is identity threat, what unlearning researchers call the harder half of organisational change. Argyris named the mechanism in 1977: most training delivers single-loop learning (changing behaviour within existing assumptions), while AI adoption typically requires double-loop learning (revising the governing assumptions). Becker’s 2019 synthesis found that the shift triggers identity threat, disrupts power relationships, and requires sustained environmental pressure. A two-hour workshop does not produce it. Lived experience, peer modelling, and organisational support do, and they are what the transfer-of-training literature has named for forty years.
The manager bottleneck
The most counterintuitive finding in the recent research: the binding constraint on AI adoption is not the frontline but the manager layer.
BCG’s AI at Work 2025 study (n=10,635 employees across 11 countries) found that only 25% of frontline employees report strong leadership support for AI. Where that support is present, the share of employees who feel positive about generative AI rises from 15% to 55%. Leadership modelling moves adoption more than any technology factor in the data.
Gallup’s State of the Global Workplace 2025 adds the structural context: only 44% of managers worldwide have received any management training at all. Asking under-trained managers to model an entirely new cognitive practice, when most of them are not using AI themselves, is a predictable failure pattern. BCG’s dose-response data shows how thin the training-only lever is: employees with more than five hours of AI training are 79% regular users; with under five hours, 67%. And 18% of regular AI users received no training at all. Manager support beats training time as a predictor of adoption.
What reskilling actually is
Reskilling is not training with more hours. The distinction comes from the WEF, McKinsey and ATD consensus.
Training transfers knowledge content, evaluated at Kirkpatrick Level 1-2. Upskilling deepens existing role capabilities. Reskilling combines behaviour change, unlearning of old habits, and organisational support infrastructure to produce a substantially redesigned role. It is measured at Kirkpatrick Level 3, sustained behaviour change on the job observed at 30/60/90-day checkpoints, not at Level 1.
The cost arbitrage is real: WEF puts the saving from reskilling existing employees at 70-92% versus external hiring. That holds only if the reskilling is actually reskilling, not a course with a certificate.
What it requires in practice:
Champions infrastructure. Citi runs a two-tier model: around 25-30 nominated AI Champions support roughly 4,000 voluntary AI Accelerators across 182,000 employees in 84 countries. Peer-driven, badge-based, no compensation linkage. The result is over 70% of employees using firm-approved AI tools. Citi added a baseline prompt-training mandate in late 2025; the voluntary network remains the deep-engagement tier.
Skill taxonomy and role architecture before content. AT&T’s $1B Future Ready programme (Donovan and Benko, HBR, October 2016) consolidated approximately 2,000 job titles into broader role categories and built Career Intelligence, an internal platform showing required skills and growth trajectories, before delivering mass training. Most organisations do the reverse. AT&T’s retrained employees ultimately filled half of the new tech management jobs the company created.
Manager enablement as the first investment. Given that manager support is the strongest predictor of transfer, enabling managers should be the first budget line. It is typically the last.
Measurement at Kirkpatrick Level 3. Percentage of shadow AI converted to sanctioned tools, specific behaviour changes per role, productivity delta in redesigned processes. Without Level 3 measurement, the programme is invisible to leadership and unaccountable to results.
Unlearning infrastructure. Deliberate intervention for the identity shift senior employees face, including coaching for those whose expertise is being partially commoditised.
The Briefing
BCG’s Henderson Institute published AI Will Reshape More Jobs Than It Replaces in April 2026, an analysis of approximately 165M US jobs across 1,500 roles. The finding: 50-55% of US jobs reshaped within two to three years, with 10-15% eliminated within five. The dominant problem is reskilling at scale, and the timeline is shorter than most boards have planned for. The market has no clear direction on this: some companies are announcing cuts in the name of AI, others — already past their own cuts — are reversing them after over-shooting. These are two stages of the same process. PayPal announced in May 2026 a roughly 20% workforce reduction explicitly framed as an AI pivot. Klarna’s CEO, having made the same move earlier, admitted in early 2026 that the company had been too aggressive in replacing people with algorithms and had to resume hiring under a hybrid model.
The regulatory clock is also closing. EU AI Act Article 4 (AI literacy) has been in force since February 2025; the Digital Omnibus on 7 May 2026 softened the deployer obligation toward “supporting improvement” while keeping mandatory human oversight training for high-risk systems. Member states begin strict enforcement on 2/3 August 2026, with Act-framework penalties up to €15M or 3% of prior-year global turnover.
Questions for leadership
Does anyone in your organisation own reskilling? If your AI thinking is only about training and you are shopping for generic training programmes for the workforce, do not expect behaviour to change.
What percentage of your AI investment goes to manager enablement versus end-user licences and training? The data puts the binding constraint one level higher in the organisational hierarchy than where most corporate budgets concentrate.
How do you measure AI competence and deployment? If the answer is course completion or licence utilisation, you are measuring Level 1. What are your Level 3 (behaviour) metrics at 30, 60 and 90 days?
When did your senior leadership team, not the AI Lab or the CoE, last visibly use generative AI for their own work? The 15-to-55% adoption swing tied to visible leadership use shows that such a signal can deliver more than the most expensive vendor-led training programmes.
Summary
The dominant market approach to AI deployment is to buy a tool, run a course, and report completion rates. Data collected across thirty years of IT deployments, four meta-analyses of training transfer, and the 2024-2026 BCG, McKinsey and Wharton survey wave is unambiguous: this approach limits the real level of technology utilisation. BCG’s 2026 AI Transformation Is a Workforce Transformation shows the cost of this mistake: 60% of enterprises generate no material value from AI despite continued investment, and only 5% create substantial value at scale.
Process redesign, role architecture, manager enablement, Champions infrastructure, Level 3 measurement, and unlearning support are much harder, slower, and generate significantly larger costs. But this is where the lever for return on investment sits: MIT-CISR data on the Stage 2 to Stage 3 maturity transition shows gross margin moving from -1.4pp below market to +0.8pp above, with revenue growth jumping 4.7pp.
Stay balanced, Krzysztof Goworek
Krzysztof Goworek is founder of Quintant — supporting AI deployments that deliver measurable business outcomes.