The report is ready, the applications are ranked, and the reply to the customer is one click away. Everything seems to be moving faster. Then someone interrupts the demonstration: “That decision is not one we should automate.” In a company eager to see a return on its artificial intelligence tools, this objection can look like a lack of ambition. Yet it could become a mark of professionalism: knowing how to distinguish what a machine can produce from what an organization must remain accountable for.
Looking ahead to September 2026, this question goes beyond proficiency with software. It involves judgment, negotiation and managerial courage. This analysis draws on facts documented before that date; the developments envisaged for 2026 are forward-looking scenarios, not a retrospective assessment of that September.
The trap: confusing a task with a responsibility
Automation rarely arrives as an announced wholesale replacement. It slips in through a small convenience: summarizing a file, suggesting a score, preparing a recommendation. The shift comes afterward. Because the recommendation is well presented and immediately available, it becomes the default decision. The human no longer truly decides; they confirm.
In recruitment, extracting the skills explicitly mentioned in a résumé is not the same as deciding who deserves an interview. In insurance, identifying a missing document is not the same as interpreting a complex family situation. In management, summarizing objectives does not justify concluding that an employee lacks commitment. A single workflow contains both operations that can be standardized and judgment calls that affect people.
The first skill, then, is knowing where to draw the line within a workflow. Rejecting all digital assistance rarely makes sense. Nor does allowing a tool to handle an entire sequence simply because it performs the easiest steps successfully. Sound judgment begins with the question: at what point do we need to interpret, justify or hear a conflicting account?
What real incidents have already shown
In 2024, a Canadian tribunal held Air Canada responsible for incorrect information provided by its chatbot about a bereavement fare. The company could not simply treat the assistant as an independent entity. The lesson extends beyond customer service: automating a statement does not remove the responsibility of whoever puts it into circulation.
Another warning came in a different context: in 2023, US lawyers were sanctioned after submitting court filings citing cases invented by ChatGPT. The problem was not just that errors had been generated. It was that they had made their way into a professional document without sufficient verification. A plausible answer had acquired the authority of legal research.
These cases do not prove that AI is unusable. They show that a fluent output can conceal underlying weaknesses. They also highlight an interpersonal challenge: flagging a problem sometimes means slowing down a process already in use, contradicting an enthusiastic superior or questioning a recent investment.
Identifying situations where people need to stay in control
The right criterion is not simply how difficult a task appears to be. A very simple message can communicate a decision with far-reaching consequences. Conversely, a complex technical operation can be automated if its result can be verified and any failure easily corrected. Four questions help move the discussion beyond the abstract:
- What is the cost of an error? Awkward wording and the denial of a benefit do not carry the same risks.
- Is the decision reversible? Can its effects be corrected quickly, without lasting harm?
- Does the tool have access to the context? An exceptional situation or information missing from the file can change the interpretation.
- Who will be able to explain and defend the choice? Someone must be able to offer an answer beyond “the system recommended it.”
Human judgment becomes particularly important when several criteria pull in opposing directions. Making an exception for a customer, for example, means weighing a rule, a relationship and a precedent. The tool can provide useful information. It must not become a quiet way of avoiding that judgment call.
Beware of human involvement that is purely cosmetic
Requiring human approval is not enough. If a manager has to review hundreds of proposals within an hour, without access to sources or any real ability to change the output, their presence is largely for show. Credible oversight requires time, expertise and the authority to suspend the process.
It is also important to acknowledge that humans make mistakes, become tired and reproduce biases. The meaningful choice is therefore not between an imperfect machine and an infallible professional. It is a comparison of practical organizational arrangements: which is better at detecting errors, allowing decisions to be challenged and assigning clear responsibility?
Turning a refusal into a proposal for management
Saying “I don’t feel comfortable with it” leaves the field open to promises of productivity. To be heard, it is better to specify the scope of the objection, the risk and the alternative. An HR manager might propose automating the formatting of application files while rejecting the automatic screening out of candidates. They are not blocking the project: they are distinguishing assistance from the delegation of decision-making authority.
The strongest approach is to bring forward a few representative cases, including difficult exceptions, and then compare the existing process with the proposed one. This means examining errors, but also verification time, rework and opportunities for appeal. Time saved in producing an output is not necessarily time saved for the company. An instant answer can trigger several days of remedial work.
A limited trial can then establish conditions: which data to use, which results to verify, who decides when disagreements arise and when to stop. The refusal becomes specific: not this level of autonomy, not in these situations, not while these safeguards are missing. This position is more defensible than an enthusiastic yes or a blanket no.
A collective skill, not a solitary act of heroism
The European Union’s AI Act, which entered into force in 2024 with phased implementation, reinforces this approach to risk management. Among other provisions, it sets human oversight requirements for certain high-risk systems. This does not mean that every office tool falls under the same rules; it underscores that oversight is a function that needs to be organized.
Internal culture will matter as much as procedures. If every objection penalizes the person who raises it, employees will learn to stay silent. Managers can instead explicitly ask what should not be automated, document disagreements and recognize incidents prevented. Defending a boundary then becomes a normal part of the job, rather than an individual act of bravery.
What next? If assistants become more autonomous in 2026, the distinguishing skill may be less about knowing how to ask more of them than knowing where to stop them. The strongest companies will not necessarily be those that automate the most, but those that can explain their limits. Saying no, in this context, is not a rejection of progress: it is a way of retaining the ability to choose its direction.
