AI does tasks, not jobs

The calculation sounds clean. An AI works around the clock, never falls ill, asks for no bonus and, in many clearly defined tasks, makes fewer mistakes than a human on a bad day. Anyone doing that maths quickly concludes that headcount can be cut. The thought is understandable. It is still misleading.

The flawed assumption

The error is an equation. Technical replaceability is not the same as organisational capability. An AI can take over a task without taking over a job. A job is more than the sum of its tasks. It contains judgement in unclear situations, context knowledge documented nowhere, the ability to escalate when something is wrong, and responsibility someone actually carries.

As a rough rule of thumb: a large share of work is routine and pattern recognition, and there AI is strong. The smaller remainder is judgement, exception and context. That is exactly where the value sits. And that is exactly where it breaks when people are cut first.

The pattern is observable

This is no longer theory.

Ford expanded its automated quality systems substantially and then brought back some 350 experienced engineers over three years. Not because the AI was worthless, but because without that experience it did not find the decisive faults. An AI is only as good as the knowledge it is trained on, and that knowledge had left first. Only in the interplay of people and system did quality return: the group has since reported falling warranty cases and recall costs, according to its leadership in the hundreds of millions.

Klarna went the same way earlier and more visibly. The company had largely automated customer service and, by its own account, pushed the fully automated line too far. Today it runs a model in which AI handles volume and people handle the cases that require judgement and relationship.

IBM automated large parts of its HR processes. Most of it worked. What remained were the difficult cases, the exceptions, the human element. The workforce did not shrink as a result, it grew. The freed-up resources went into roles that require judgement and dealing with people.

These cases are not an argument against AI. They show a pattern. Whoever automates execution while cutting the people who exercise judgement saves in the short term and pays twice in the medium term.

The effect that only becomes visible later

Entry-level and routine work is not just a cost item, it is the training ground. Experts emerge through practice, mistakes and growing responsibility. Automate that stage away entirely and you dry up the source from which the seniors of the day after tomorrow come. The well does not run dry immediately. But it runs dry.

The right question

The real question is therefore not whether an AI completes a task faster. It often does. The question is how a company automates without losing the judgement and the knowledge that keep it steerable.

Value does not come from replacing capacity with machines, but from keeping knowledge inside the company, making decisions traceable and connecting them to execution. That is exactly where I come in. Sensible automation does not mean maximum automation, but automation under human leadership, where people still understand goals, context and consequences and can intervene when it matters.

Before you cut a role

A sober calculation pays off. What does it cost if you have to bring that experience back in twelve months, and will you even find it then. What does the standstill cost when nobody is left who notices that the system is giving the wrong answer. These numbers rarely appear in the first business case. In the end they decide the outcome.

This is the first of three articles. The second looks at why AI recommends surprisingly similar things on strategy questions and how to use it well. The third at what digital sovereignty concretely means for an SME.