Most people who write about AI do so from a desk: from universities, think tanks, newsrooms or policy circles. These are important perspectives. But they sometimes lack something crucial: contact with the environments where the technology has to work under time pressure, quality requirements and economic accountability.
I spend much of my time on improvement projects close to production: food, manufacturing, flow problems, quality deviations, downtime, material variation. That is the environment where I see what AI is already doing for real — and what it still cannot handle.
What I see is not that humans disappear overnight. What I see is that an experienced person with AI can do significantly more groundwork, cover more angles, and produce better material faster than before. That alone is enough to start changing organisational logic, staffing and competitiveness.[1]
A representative improvement case
In a typical process problem, the work begins with questions like these: why does the weight outcome vary, why is waste increasing, why do temperature profiles drift, why does the line behave differently between shifts, which experiments are worth running first? Previously, this often meant several tools, manual data extracts, lists of hypotheses, visualisations and writing up reports over several days.
Today a first useful analytical basis can often be produced the same day. AI helps to structure data, suggest analytical paths, draft code, summarise observations and turn technical notes into material that can be discussed with the operation. This does not mean AI does the job for me. It means more of the work can be spent on judgement, verification and getting people on board, instead of on mechanical groundwork.[2]
The multiplier sits in the experience
The most important thing I see from the factory floor is therefore not what AI does on its own, but what the combination of human plus AI does compared to a more traditional way of working. A person who knows what they are looking for can go broader, deeper and faster than before.
That is also why the effect is uneven. An inexperienced user can get a lot of surface and little substance. An experienced user, by contrast, can use the same tools as a multiplier for domain knowledge. The question is therefore not just "how good the model is", but how well it is connected to problem framing, data, process understanding and quality assurance.[3]
Where AI still falls short
It is important to be clear: models hallucinate, miss edge cases and sometimes give answers that sound confident without being robust. They do not always understand the physical constraints of a process. They do not automatically see the informal reality of a factory: shift culture, local workarounds, maintenance debt, measurement errors, organisational inertia.
That is why the best results emerge when AI is combined with strong domain knowledge. An experienced operator or improvement leader often sees things the model does not: when a reasonable suggestion will fail because of material variation, safety requirements or simply how the work is actually done on site.
What industry sees earlier
Industry is useful as a lens because it is concrete. Flow, quality, energy, lead time, scrap and downtime can be measured. There is less room for fluff. Either the process improves or it does not.
That is also why industry-near environments often see the AI shift earlier than the public debate. When the technology saves time in a real workflow, it shows up quickly in execution, staffing and decision-making rhythm. You do not need to agree on every future scenario to see that something is already changing.
What I conclude
I am not trying to write as an evangelist or a doomsayer. I am trying to write as an engineer: observe, compare, test and draw conclusions. The conclusion I draw right now is simple: AI already works well enough in parts of industry-near knowledge work to change how value is created.
That is not the same as saying everything is being automated now. But it is more than enough that management teams, education systems and politicians should talk less about experiments and more about transformation.
Source notes
The essay combines the author's own field observations with broader sources on adoption, productivity and tasks.
- For how AI redistributes tasks rather than simply replacing whole jobs, see Anthropic Economic Index, January 2026 and ILO Working Paper 140.
- On how AI affects productivity and value creation, see PwC Global AI Jobs Barometer 2025 and Stanford AI Index 2025.
- For overviews of how generative AI is used in real tasks, see the Anthropic Economic Index; for broader labour market context, see WEF and IMF.