If Your AI Rollout Is Working, You Should Be Worried

Enterprises are retraining people to work with AI. Far fewer are checking whether those people can still think without it, or where their reasoning behind their decisions now lives. 

In May 2025, Sebastian Siemiatkowski, chief executive of Klarna, told CNBC that AI had helped shrink the workforce by around 40%. The company had held a hiring freeze for more than a year while its AI assistant handled what Klarna claimed was the workload of 700 customer service agents. A few months later, the company began hiring human customer service personnel, and Siemiatkowski told Bloomberg. (1) “Really, investing in the quality of human support is the way of the future for us.”  

Most coverage filed this under “the technology wasn’t ready.” I think that’s the wrong lesson, and wrong lessons are expensive. 

Klarna’s real problem wasn’t underperformance. It was that nobody inside the company could see the underperformance until customers pointed it out. The dashboards looked good, and the resolution times were acceptable. What had slipped was the quality of judgement on the hard cases, and nothing in the reporting measured judgement. 

That blind spot is the talent story of 2026, formed of two questions: How do we retrain our people for AI, and how do we stop losing what they already know?  

The upskilling gap: it’s not all about the numbers 

Many people can quote The World Economic Forum’s Future of Jobs Report 2025 change figure from memory: employers expect 39% of workers’ core skills to change by 2030.  

Of every 100 workers, 41 will need no significant training, 29 will be upskilled within their current roles, 19 will be reskilled and redeployed, and 11 will need training that won’t be accessible to them. (2)  

Those figures describe volume and reach i.e how many people need training on what. Nearly every enterprise AI capability programme I see is measured the same way: seats filled, modules completed, tools adopted, hours saved. It’s useful operational information but it tells you nothing about whether the judgement sitting on top of the tools is holding up. 

When work moves into a model, three liabilities quietly accumulate, and each of them is celebrated as an improvement before it is recognised as a problem. I have borrowed the use of ‘debt’ from Wolfgang Rohde, whose recent preprint separates “capability masking” from “capability erosion”. (3) I add a timescale as it is this which highlights the failure of most measurement reporting. 

Judgement debt: the individual clock 

Timeframe: a few weeks  

Michael Gerlich (2025) surveyed 666 people across ages and educational backgrounds and found a strong negative correlation between frequent AI tool use and critical thinking, with cognitive offloading doing most of the mediating work. (4)  

Kosmyna, N. et al. (2025) put 54 participants through repeated essay-writing sessions across four months, split between those using an LLM, those using a search engine, and those using nothing. The LLM group showed the weakest, least distributed brain connectivity and reported the lowest sense of ownership over their own work. Many struggled to quote accurately from essays they had themselves produced. In a smaller fourth session, 18 of them swapped conditions, and the former LLM users carried their under-engagement with them into unassisted writing. (5)  

Ujué Agudo and colleagues, publishing in Cognitive Research: Principles and Implications in 2024, had participants assess defendants’ guilt with algorithmic support that was deliberately wrong in a third of cases. Accuracy fell and fell hardest when the AI’s recommendation arrived before the person had formed their own view. (6) 

Human-in-the-loop is only a control if the human in the loop still forms an independent judgement and can detect a confident, fluent, plausible mistake. That’s a trainable skill, and just like all skills, if you do not use it, you can lose it. I suggest to all senior leaders that ‘judgement’ be added to their risk registers. 

Memory debt: the organisational clock (knowledge retention) 

Timeframe: a couple of years 

When our organisations become engrossed in prompt templates and model outputs, something subtle happens to what the organisation knows. The output continues; however, the reasoning behind it can disappear. 

Ozioma, G. et al. (2026) propose a taxonomy of AI’s less visible costs and give this its own category: organisational memory loss. (7) Institutional knowledge has historically lived in mentoring, shadowing and shared routines. Prompt libraries preserve fragments of experience while stripping out the situational context that made it expertise in the first place. Knowledge stops being distributed across people and becomes contingent on somebody else’s architecture. 

Teresa Tung and Philippe Roussiere at Accenture, writing in California Management Review, March 2026, claim: “The real differentiator is not the data or even the models, but the ‘tacit knowledge’ embedded in the judgement of their people.” (8)  

Every part of most AI stacks can be bought; the difference between our competitors and us is tacit knowledge. 

Kate Niederhoffer and colleagues, working with BetterUp Labs and Stanford’s Social Media Lab, documented what they named ‘workslop’: content that “appears polished but lacks real substance, offloading cognitive labour onto coworkers.” Each instance cost the recipient close to two hours of rework. (9) Frustrating and time-consuming without a doubt, but leaders need to ask: how much workslop stays within the company systems, having not been reworked? 

Pipeline debt: the generational clock 

Timeframe: +/- a decade 

Enrique Ide at IESE Business School has modelled what happens when AI absorbs early-career tasks, using an overlapping-generations framework. His conclusion is that these technologies “may reduce entry-level opportunities and threaten the implicit contract in which younger workers exchange their labour for training and the prospect of future promotion.” (10). According to Ide’s calibration, automating between 5% and 30% of entry-level tasks negatively impacts US per capita growth. His model has output sitting nearly 20% below baseline after a century. 

Labour-augmenting technology can help a new starter be more productive and strengthen knowledge transfer. Labour-substituting technology takes the new starter out of the picture and cuts off the transfer. Junior people learn by doing junior work next to experts. Automate the work, and you close the channel the learning travelled through, so the next generation of experts is smaller and less capable, and a less capable workforce produces less.  

Each debt pays a dividend first 

Judgement debt means faster decisions. Memory debt equals consistency and standardisation. Pipeline debt presents a leaner cost base.  

“The result is a system that appears more efficient in the short term while becoming more fragile over time.” Wolfgang Rohde 

Things worth measuring 

Measure dependence. For a given piece of work, investigate what proportion of the reasoning steps the AI is doing, and what proportion the person is doing. Ozioma et al. propose something along these lines as a cognitive dependence measure. It might create an interesting discussion in leadership team meetings. 

AI-off experiments. Pick your three highest-consequence recurring decisions. Once a quarter, have a team work one of them unassisted and score the output against your normal standard. If quality collapses, you’ve found a dependency you didn’t know you had. If it holds, you have evidence your oversight function functions.  

Capture reasoning, not just output. Most knowledge management archives what was decided, but the why it was decided is often lost. Encourage voicing the opposite case out loud before committing. Name what would have to be true for the plan to work, then test the shakiest assumption first; and run a premortem, writing the story of the failure a year out while there’s still time to act. This is not a new practice; what is new is that AI has removed the friction that used to force people to do it. 

Deliberately train people on flawed AI. Korosec-Serfaty, M. et al (2026) ran three experiments with practising HR professionals working alongside AI. (11) They recommend that people need exposure to AI that is sometimes unreliable if they’re going to test it rather than accept it. Nearly every corporate AI programme I’ve reviewed does the reverse: it demonstrates the tool working beautifully and calls that training. 

Protect the tasks juniors learn from. Absolutely automate the formatting on the blank page but let the new starter do the reasoning, make the judgement call, have the conversation about options. They need to own their decisions just as the company leaders do. 

What competence looks like 

The UK’s Department for Work and Pensions, published a set of AI upskilling case studies in June 2026 which told us: 

• Airbus built structured learning pathways, competency frameworks and communities of practice; in a trial of roughly 2,000 employees, the average user saved up to four hours a week when using an AI assistant effectively. 

• Roche made its training mandatory and reached more than 100,000 people quickly.  

• KPMG put over 130 employees, senior leaders included, through an AI-focused apprenticeship rather than a course catalogue. (12) 

The outcomes featured in the above examples are, as mentioned at the start of this article, speed metrics such as hours saved, time reduced, and efficiency gained. The case studies do not mention measuring whether judgement was held, whether reasoning was retained, or whether early careers recruits learnt anything. Some of the best-documented upskilling programmes in the country measure only whether people can use the tools.  

A challenge 

Your AI adoption rate tells you how much of your organisation’s thinking you have moved into a model, which is definitely worth knowing.  

A question for your next board or SLT meeting.  

Does your AI adoption report tell you whether you could take it back? 

Sources 

1. Klarna is hiring customer service agents after AI couldn’t cut it on calls, according to the company’s CEO. Entrepreneur, 9 May 2025. https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396
2. World Economic Forum (2025) The Future of Jobs Report 2025, Chapter 3: Skills outlook. https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/3-skills-outlook/
3. Rohde, W. (2026) Short-Term Gain, Long-Term Fragility: AI Labor Substitution and the Erosion of Sustainable Capability. arXiv:2605.27399 [preprint]. https://arxiv.org/abs/2605.27399
4. Gerlich, M. (2025) AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006
5. Kosmyna, N. et al. (2025) Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, arXiv:2506.08872 [preprint]. https://arxiv.org/abs/2506.08872
6. Agudo, U., Liberal, K.G., Arrese, M. and Matute, H. (2024) The impact of AI errors in a human-in-the-loop process. Cognitive Research: Principles and Implications, 9(1). https://doi.org/10.1186/s41235-023-00529-3
7. Ozioma, G. et al. (2026) AI as Cognitive Ecology: Revealing the Invisible Cognitive, Cultural, and Epistemic Costs of Generative Models. Journal of Artificial Intelligence and AI Ethics, 1(1), pp. 1–16. https://doi.org/10.64978/jaiae.2026.0116002
8. Tung, T. and Roussiere, P. (2026) Tacit Knowledge Is Your Next Competitive Moat. California Management Review, 16 March 2026. https://cmr.berkeley.edu/2026/03/tacit-knowledge-is-your-next-competitive-moat/
9. Niederhoffer, K. et al. (2025) AI-Generated “Workslop” Is Destroying Productivity. Harvard Business Review, 22 September 2025. Research by BetterUp Labs with Stanford Social Media Lab. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
10. Ide, E. (2026) Automation, AI, and the Intergenerational Transmission of Knowledge. IESE Business School, arXiv:2507.16078 [preprint]. https://arxiv.org/abs/2507.16078
11. Korosec-Serfaty, M., Léger, P.-M., Parent-Rocheleau, X. and Sénécal, S. (2026) Critical Thinking in AI-Assisted Deontologically-Governed Professional Decision-Making. Information Systems Frontiers, 23 February 2026. https://doi.org/10.1007/s10796-025-10691-2
12. Department for Work and Pensions and Skills England (2026) Supporting case studies: What works for AI upskilling in the UK. GOV.UK, 10 June 2026. https://www.gov.uk/government/publications/skills-for-ai-what-works-for-ai-upskilling-in-the-uk/supporting-case-studies-what-works-for-ai-upskilling-in-the-uk
13. EU Artificial Intelligence Act, Article 4 (AI literacy) and Article 14 (Human oversight). https://artificialintelligenceact.eu/article/14/