After being presented as a machine to replace workers, artificial intelligence is starting to encounter a more complex reality. Since the rise of generative AI, many companies have explained their workforce reductions, hiring freezes or reorganizations by the productivity gains promised by the new tools.
The idea was seductive: automate repetitive tasks, reduce costs, produce faster, with less staff.
But a backlash is beginning to appear. Not because AI would be useless, nor because companies renounce automation. On the contrary, it is taking root in organizations. What changes is the understanding of its real role: AI works better as a tool to augment human labor than as a total substitute for employees.
The real trend, therefore, is not pure replacement of humans by machines. It is a reconfiguration of work: some tasks disappear, others are accelerated, but the functions of judgment, control, relationship, responsibility and expertise remain difficult to automate entirely.
What companies thought they could automate
The first wave of enthusiasm around AI mostly touched functions considered repetitive: customer service, internal support, operational marketing, translation, standardized writing, simple coding, human resources, document analysis or administrative tasks.
In this context, several large groups explained they could slow down certain recruitments thanks to AI. IBM, for example, announced as early as 2023 a pause or slowdown on some back-office roles likely to be automated.
Salesforce, for its part, now says that its AI coding tools allow it to keep its engineering headcount at a stable level, without recruiting as much as before in this area.
These announcements do not mean that all employees are replaceable. They show rather that companies are looking to identify the areas where AI can absorb part of the workload.
The danger begins when this logic is pushed too far: replacing experienced people with even imperfect systems can create new costs, less visible but sometimes heavier.
The reality check: automating is not understanding
The most telling case is customer service. AI can answer simple questions, retrieve a procedure, summarize a file, or guide an agent. But it quickly reaches its limits when the situation becomes ambiguous, emotional, urgent, legal or commercial.
That is precisely what Gartner highlighted in early 2026: by 2027, 50% of companies that reduced their customer service staff in the name of AI should rehire personnel for similar roles, often under different titles. This figure is important, because it does not describe the AI’s failure; it describes the failure of an automation thought out too quickly as a removal of human.
Klarna, the Swedish fintech specializing in split payments, had long been touted as one of the most spectacular examples of automation. The company had highlighted its AI chatbot, capable, it claimed, of performing the work of hundreds of advisers. But it later reintroduced more humans into its customer service, recognizing that customers must still be able to speak to a real person when the situation demands it.
This turnaround does not mean AI has failed. It means it was mispositioned. It is effective at absorbing simple requests, speeding up responses and assisting teams. It is far less reliable when entrusted alone with the relationship with an unhappy customer, a particular case, or a decision that affects the company’s reputation.
The mistake some companies made was to confuse speed with quality. An automatic response can be fast, but if it is inaccurate, cold or unable to resolve the real problem, it destroys trust.
And in many sectors, trust is worth more than the savings from removing a few positions.
The IBM example: Hire differently rather than permanently cut
IBM illustrates another important evolution. After talking up the automation possible for certain functions, the group plans to triple its recruitment of entry-level profiles in the United States in 2026. But this is not simply going back to the old model. Junior roles are redesigned to integrate AI from the start.
This decision says something essential: a company that stops hiring beginners risks breaking its own skill chain. Who will become manager, expert, architect or project leader in five years if young profiles are no longer trained today?
AI can help a beginner be more productive, but it does not replace learning the craft, understanding customers, corporate culture, or the ability to manage unforeseen situations.
The question is therefore no longer just: “How many positions can AI replace?” It becomes: “How to train workers capable of working with AI?”
The Salesforce example: the engineer does not disappear, their role changes
Salesforce offers a more nuanced example. Marc Benioff explains that the company did not need to significantly increase its engineering headcount, because AI coding tools allow producing more with a stable team. By contrast, Salesforce continues to recruit in commercial roles, where human relationships, persuasion, understanding the client and negotiation remain central.
The lesson is clear: AI does not mechanically remove jobs. It shifts value within jobs. In software development, producing code becomes faster. But designing an architecture, understanding a need, securing a system, maintaining a complex application, verifying quality or arbitrating between several technical choices remain highly human tasks.
The engineer of tomorrow will not only be the one who writes code. It will be the one who knows how to steer AI agents, verify their results, understand their mistakes, integrate their outputs into a reliable system, and assume the technical responsibility for the final product.
Why AI augments better than it replaces
The most rigorous studies show that AI yields real gains when it assists employees. In customer support, an AI assistant can help structure responses, retrieve the right information and reduce the gap between beginners and seasoned professionals.
That is precisely where the value lies: AI acts as a skills accelerator. It does not necessarily replace the professional; it helps them work faster, better document their decisions, reduce repetitive tasks, and focus on the most important cases.
But this logic requires one condition: humans must stay in the loop. When there is no supervision, risks increase: invented responses, context errors, misinterpretation, privacy issues, unjustified decisions, or inability to handle exceptions.
The MIT NANDA report on AI in business in 2025 also underscores the brutal reality: despite 30 to 40 billion dollars invested in generative AI, 95% of the organizations studied saw no measurable return. The problem is not only technological. It mainly lies in integration: many companies bought tools, ran pilots, delivered demonstrations, but did not truly transform their work processes.
This figure sums up the current paradox: AI impresses in a demonstration, but it only creates lasting value when it embeds itself in a profession, a method, a decision chain, and human accountability.
And if this backlash is only temporary?
However, we must avoid a too comfortable reading. Some experts think that current AI limits are only a transitional stage. Models are progressing rapidly, agents are becoming more autonomous, systems connect to internal databases, and verification tools improve.
In this scenario, the current return of humans would not be a definitive victory over automation, but a phase of adjustment. Companies may have reduced their workforce too quickly, before the tools were reliable enough. The next generations of models could thus fill some of the current gaps, particularly in customer support, coding, document analysis, or administrative tasks.
This counter-argument is serious. It requires nuance in the diagnosis. AI does not yet massively replace all employees, but it is advancing fast. Jobs that rely mainly on standardized tasks remain exposed. Conversely, roles that combine expertise, human interaction, responsibility, creativity, arbitration and field knowledge resist better.
The conclusion is not: “AI will replace no one.” The conclusion is more like: “AI will replace tasks, transform jobs and favor workers who can use it with discernment.”
The real challenge: not erasing the company’s human memory
The risk for companies is not only social. It is also economic. When an organization fires too quickly, it loses part of its memory: special cases, customer habits, known mistakes, informal arbitrations, operational details that procedures never fully describe.
This human memory is difficult to replace with an AI model. Data can be stored, but experience cannot be reduced to a documentary base. It is built in real situations, mistakes, interactions with customers, internal constraints, and decisions made under pressure.
That is why the most prudent companies are not just looking to automate. They aim to reorganize work around a duo: fast machine, responsible human. AI handles the volume; humans handle the exception, control, strategy and trust.
Thus, the current backlash does not mark the failure of artificial intelligence. It marks the failure of a too simplistic promise: that a company could quickly replace a large portion of its employees with automatic tools.
AI is powerful when used as a co-pilot. It becomes fragile when used as a complete substitute. Companies are discovering that productivity does not depend only on speed of execution, but also on quality, trust, responsibility, and the ability to manage unforeseen situations.
The true revolution in work will therefore not be a humanless company. It will be a company capable of combining the efficiency of machines with the judgment of professionals. AI does not render humans unnecessary; it makes those who know how to use it, control it, and leverage it for real value more valuable.
