So, the pilot has escaped the lab
Large language models have been with us long enough that describing AI as an experiment is beginning to feel slightly outdated. In Finland, 38% of enterprises were using AI technologies in spring 2025, up by 14 percentage points in just one year; among enterprises employing at least 100 people the figure was 68%, and in information and communications it had already reached 80%. The employee perspective tells much the same story: Finland’s 2025 Working Life Barometer found that 45% of wage-earners said AI was being used at their workplace, rising to 70% in organisations with at least 200 employees. Nearly half of those working in AI-using workplaces felt it had improved their efficiency, while a third felt it had improved the quality of their work. We are therefore past the stage where AI is a curious little pilot sitting safely inside an innovation team’s PowerPoint deck. Particularly in knowledge-intensive work, it has entered the actual workflow, which makes the more interesting question no longer whether we will work with AI, but what happens to the way we think once we do.
And the research is beginning to show something more nuanced than either “AI makes us cleverer” or “AI makes us stop thinking”. A 2025 CHI study by Lee and colleagues surveyed 319 knowledge workers across 936 real examples of GenAI use and found that greater confidence in AI was associated with less critical-thinking effort, while confidence in one’s own abilities was associated with more. More interestingly, the researchers found that critical thinking did not simply disappear; it shifted towards verifying information, integrating responses and supervising the task. Another study by Doshi and Hauser found a similarly interesting trade-off: people given AI-generated ideas produced stories judged as more creative, better written and more enjoyable, particularly among initially less-creative writers, but those stories also became more similar to one another. So AI can improve individual output while quietly narrowing the collective range of thought. Perhaps the psychological change is not that we suddenly cease thinking, but that competent answers arrive so easily that it becomes tempting to stop interrogating them quite as much.
Ah, so the clever bit moved somewhere else
This is where my own experience with AI in software engineering has made me slightly suspicious of the phrase prompt engineering. Efficient use of AI, to me, depends much more heavily on the expertise of the person using it. By efficient, I do not simply mean producing something in twenty minutes that previously took four hours. I mean producing something correct, appropriate to its context and maintainable after everyone has stopped being impressed by how quickly it appeared. Domain knowledge matters before generation because it helps us define and constrain the problem; experience matters afterwards because it helps us recognise whether the generated solution should actually be allowed anywhere near production.
Take an inventory-management application. Someone relatively new to the domain might understandably type, “Build me an inventory warehouse application using Java and React.” The model will oblige, probably rather enthusiastically. An experienced engineer is likely to give it something closer to: “We are building a warehouse inventory system using Java 21, Spring Boot and PostgreSQL, with React on the front end. We need warehouses, stock items, inventory movements and purchase orders. Keep the controllers thin and limited to HTTP concerns, put business rules and transaction handling in the service layer, keep persistence models separate from API DTOs, and design the data layer so stock history is auditable rather than simply overwriting the current quantity. Before generating the implementation, let’s model the relationships: can an item exist across multiple warehouses, should inventory movement belong to both a warehouse and stock item, and would modelling this as many-to-many create problems for auditability or future reporting? Give me the alternatives and trade-offs first.” The difference is not that the second person has discovered a secret way of speaking to an LLM. They already possess a mental model of the system. The prompt is simply that expertise becoming visible.
The same expertise becomes valuable when the answer comes back. Anyone who has spent enough time across development, QA, DevOps and the other moving parts of an Agile environment knows that technically valid and sensible are not synonyms. Something can compile perfectly while placing business logic in the wrong layer, creating an awkward dependency, misunderstanding an entity relationship or introducing a decision that another team will have to live with six months later. An experienced engineer has seen deployments fail, schemas become painful, integrations behave unexpectedly and innocent-looking shortcuts grow up into rather expensive adults. That history influences which questions are asked before generation, but also creates the judgement required to decide which AI suggestions deserve to survive afterwards.
Perhaps experience was doing more than we thought
Much of that judgement has a useful name: tacit knowledge. Ryan and O’Connor describe software development as knowledge-intensive work in which much expert knowledge is tacit – knowledge that is difficult to transfer completely through documentation or verbal explanation. Their discussion of how expertise develops is especially relevant here: tacit knowledge grows through perception, intuition, experience and deliberate practice. In software teams, it is also not entirely individual. Their study found that the quality of social interaction was related to the amount of tacit knowledge within teams, and that team tacit knowledge contributed to effectiveness. In other words, some of what an experienced engineer “knows” has been accumulated from years of code reviews, conversations, incidents, planning sessions and watching how other specialists solve problems – not merely from reading another framework’s documentation.
The paper makes one observation that captures this particularly well: software teams require not only technical expertise but domain knowledge sufficient to understand the significance of the task and its requirements, and that domain knowledge is gained through experience and shared through social interaction. Longer-tenured teams working in a domain therefore develop richer tacit knowledge about that work. And although the study is about software engineering, the principle is much broader. Experienced doctors, designers, researchers, managers and engineers all accumulate things that are difficult to turn neatly into instructions: pattern recognition, organisational context, memories of previous failures and that occasionally irritating feeling that something looks perfectly reasonable but is somehow wrong.
This is why I suspect AI may make experience more valuable rather than less. For a long time, expertise was partly visible in production: the experienced professional could create something the inexperienced professional simply could not. AI is rapidly narrowing that visible gap. But producing an answer and taking responsibility for it remain rather different abilities. As generation becomes cheaper, more of the human contribution moves towards defining the right problem, supplying meaningful context, detecting hidden assumptions and deciding which machine-generated answer is trustworthy. Perhaps prompt engineering was never really the new expertise. Quite often, a sophisticated prompt is simply existing expertise made explicit. And the part we cannot quite write into the prompt – the accumulated judgement that tells us what to question, what to reject and where something is likely to go wrong – is the tacit advantage.
References / Sources
Statistics Finland, Use of information technology in enterprises 2025. Statistics Finland publication
Ministry of Economic Affairs and Employment of Finland, Working Life Barometer: Wage-earners have cautiously positive view of AI use at work, 2026. Finnish Government publication
Lee, H-P. et al. (2025), The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers, CHI 2025. Microsoft Research
Doshi, A. R. & Hauser, O. P. (2024), Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, 10(28), eadn5290. Science Advances paper
Ryan, S. & O’Connor, R. V. (2013), Acquiring and sharing tacit knowledge in software development teams: An empirical study, Information and Software Technology, 55, 1614–1624.
About the author
