There is a reason why this author uses a pen name: sharing anecdotes from colleagues, politics and the industry is considerably easier. Over recent months, I have been informally „interviewing“ old and new colleagues alike — Weggefährten, fellow travellers — about the adoption of artificial intelligence and how it changes their work, the training of juniors and graduates, and how organisations are managed. The pattern is clear: we are living in different parallel universes.
What used to be data management is now AI transformation. All sorts of people flock into the gold rush. They do not understand the technology, any more than they understood data. Processes, policies and pointless PowerPoints rule the dominion. This article is not, however, about rolling out Copilot to a large corporation which is now writing emails with AI chatbots that are consumed by AI chatbots.
If we take a step back, it started with the models becoming impressively good. Then came tool calls and agents with specialised tasks. But lately it has been almost quiet in mainstream media. The real action is happening out of sight.
As Marc Andreessen and other frontier voices have observed, the real disruption is bottom-up: workers and engineers quietly weaving AI into daily practice, often without notifying their managers and far out of sight from the executive‘s 17th floor. In programming the change is gigantic. Andrej Karpathy, one of the architects of modern deep learning, has argued that traditional programming languages are an intermediary layer built for humans, not machines — and that a sufficiently capable AI will not need the intermediary. Why should a potent AI use an abstraction layer created so that humans can communicate with a machine? Frontier thinkers increasingly foresee AI generating machine code — assembler or even optimised binary — directly. Prompt to executable, no compiler required.
The evidence is already in the infrastructure. This author is aware of large consulting projects making sense of codebases written in the 1980s; insurance companies are pouring millions into AI-driven refactoring of legacy risk calculation models. Tools such as IBM’s watsonx Code Assistant for Z and CloudFrame’s agentic AI are already extracting business logic from COBOL monoliths and turning them into cloud-native Java or Python, often in weeks rather than years. Bankdata in Denmark offers a sharper example still: a cooperative serving roughly 90 Danish banks that spent years locked inside a COBOL core it could barely extend, now migrating with AI assistance at a pace that would have been unthinkable three years ago. COBOL still processes an estimated three trillion dollars of daily commercial transactions. The reckoning, long predicted and long deferred, has arrived.
However, the real magic is sometimes happening closer to home. Ask your somewhat tech-savvy friends and you will learn that they have quietly build new apps or smartly organise their everyday life with Claude Code and the likes.
All this while the traditional working world is getting stuck in the morass of AI chatbot conversations: keeping AI chatbots and their humans busy adding no value. Organisations still write mails, misuse JIRA and Confluence to pollute their codified institutional knowledge on tools that were never designed for the job but which cannot do without the revenue from it.
Some industries are adopting faster. Ex-colleagues in investment banking confirm that junior analysis work has been largely replaced with Claude Code or OpenAI‘s codex. And because it is genuinely enjoyable, even partners are chipping in again. The preparation and presentation work rests with the mid-level professional — for now.
Two problems emerge from this.
It is tempting to let the model run without supervision (--dangerously-skip-permissions) and let the model create the work, review it and challenge it entirely unsupervised. One colleague put it plainly: eighty per cent of the time it is right, and twenty per cent of the time people do not notice. The real problem, however, is not that we stopped checking. It is that we stop understanding, stop reviewing, stop learning and stop challenging ourselves.
This translates directly to the second problem. When we no longer require colleagues to perform the analytical work in the first place, how do we train the junior and, in time, the senior staff that organisations so desperately need? Expertise is not transmitted by proximity to a correct answer. The psychologist K. Anders Ericsson documented this over decades: it is built through deliberate practice, the effortful, corrective repetition that is as uncomfortable as it is irreplaceable. Future learning will look different, but it cannot look like passive observation of AI outputs. The most effective path runs through small, cross-functional teams of five to seven people which we have argued for many times and in our experience works best. This is the size J. Richard Hackman identified as ideal for sustained performance, and the size at which the Ringelmann effect warns that social loafing begins to dilute effort. Juniors will pair directly with AI agents on live problems, reviewing outputs in real time, iterating and building the muscle of discernment that no textbook can teach.
So why is structural adoption so slow? PwC’s 2026 AI Performance Study makes the divide brutally clear: three-quarters of AI’s economic value is captured by just twenty per cent of companies — the leaders who treat AI as a growth engine rather than a pilot project. McKinsey’s State of Organisations 2026 finds eighty-eight per cent of firms experimenting with AI yet eighty-one per cent reporting no meaningful bottom-line gains. The spectacle of activity masks the deeper architecture that actually works: lean teams, outcome metrics and a relentless focus on value over headcount. The reason most organisations cannot close that gap is structural rather than cultural. Success is still measured in headcount. A director who manages two hundred people outranks one who manages twenty, regardless of what either produces. Labour unions rationally protect employment volumes. Governments campaign on jobs created, not productivity gained. No AI strategy document changes this until the measurement itself changes.
In ageing societies, the arithmetic becomes urgent. Europe’s median age now stands at 43.1 years, with shrinking working-age cohorts and rising pressure on pensions and healthcare. Success can no longer be measured by the size of the payroll.
The way forward is clear.
- Measure outcomes, not activity. Stop measuring digital transformation by the number of people kept busy and start measuring it by outcomes delivered. Headcount is a legacy metric.
- Build small, empowered teams. Groups of five to seven that work with AI agents daily are where real skill transfer and innovation happen fastest. Scale the model, not the headcount.
- Let the frontier lead. Encourage bottom-up adoption in programming and analysis rather than top-down consultant roll-outs. The quiet revolution already under way in investment banking and legacy-system refactoring shows what is possible.
- Redesign learning around the twenty per cent. Juniors must own the portion the model gets wrong — that is where judgment is forged. Review, challenge and iteration are the curriculum.
- Enforce conditionality on AI investment. Tie budgets to measurable productivity gains, not activity metrics. In an export-heavy, ageing continent like Europe, the prize is not more jobs but more value per job — or we risk watching the twenty per cent pull further ahead while the rest admire their PowerPoints.
The two universes will not stay parallel forever. One is already generating optimised binary from prompts; the other is still writing emails about writing emails. Europe’s choice is whether to close the gap before the frontier leaves the rest of us permanently behind.
Subscribe to The Friday Thinkers for weekly analysis of the forces reshaping Europe’s economic and political landscape. It is free.
The Friday Thinkers