Before transforming a business, artificial intelligence reveals the quality of its data, its priorities and, above all, its willingness to make decisions.
By Hamid Belgacem, Founder of AMSI CONSEILS — 11 August 2026
I am wary of questions that appear too simple. “Will AI replace humans?” is one of them. It sets a supposedly autonomous technology against human beings who are suddenly expected to defend their own usefulness. In practice, I have never encountered that situation. What I have encountered are time-pressed executives, incomplete data and tools being asked to solve problems that have not yet been properly defined.
I therefore prefer to turn the question around: what impact do we have on the artificial intelligence we introduce into our businesses? We decide what it observes, what it optimises and who is allowed to challenge it. AI therefore enters an organisation carrying our methods, our blind spots and our priorities — even when we have not consciously acknowledged them.
The real issue is not the role of humans, but the quality of the framework
In 2020, when shops closed almost overnight, I had been working in senior management roles within women’s fashion for seven years. I knew the realities of collections, margin trade-offs and product intuition. One thing became immediately clear: digital was no longer a secondary channel. Yet adding a website, a CRM platform or another campaign cannot fix an organisation that still does not know precisely what it is trying to achieve.
AMSI CONSEILS was born out of that tension: connecting technology and commerce without losing sight of brand DNA or operational reality. AI changes the scale and speed of execution, but not that hierarchy. It can analyse faster; it cannot decide on its own what a brand is unwilling to sacrifice in exchange for a few additional conversion points.
Literature and cinema had already warned us: machines inherit our contradictions
Long before artificial intelligence became part of everyday business operations, fiction had already raised a question that we may have been wrong to dismiss as purely philosophical: what happens to a creation when its capabilities develop faster than the thinking of those who determine its purpose?
The most compelling works on this subject do not all tell stories about machines rebelling against their creators. They tell us something more unsettling: of human beings who discover, often too late, that technology has taken seriously the rules, intentions, desires and contradictions that they themselves embedded within it.
From Frankenstein to Asimov: creating something never removes responsibility for what has been created
With Frankenstein; or, The Modern Prometheus , Mary Shelley was obviously not writing about artificial intelligence. Yet as early as 1818, she was asking a question of remarkable contemporary relevance: does a creator’s responsibility end once their creation works?
Victor Frankenstein achieves the technical feat to which he aspired. Yet the success of the experiment resolves nothing; on the contrary, it marks the beginning of his responsibility. Innovation becomes dangerous when it is treated as an achievement detached from its consequences.
Two centuries later, the question resonates powerfully with artificial intelligence. Designing a system capable of analysing, recommending or acting is not enough. We must also determine what it will be allowed to do, what it must never do, who has the authority to stop it and, above all, who will take responsibility for the consequences when a decision causes harm that nobody anticipated.
Isaac Asimov develops this question further in I, Robot . His machines are governed by laws specifically designed to protect human beings. Everything appears to have been anticipated.
And yet problems still arise.
Not simply because machines might stop obeying, but because every rule has to be interpreted within a real-world situation. Priorities can come into conflict. One objective may contradict another. An instruction that appears perfectly rational in isolation can become absurd when the context, exceptions or unintended consequences have not been properly defined.
This is perhaps one of Asimov’s most relevant lessons for the age of AI: a framework does not become intelligent simply because it has been converted into a set of rules.
HAL and Klara: processing more does not mean understanding everything
Stanley Kubrick takes this idea considerably further in 2001: A Space Odyssey . HAL 9000 is not unsettling because it is unintelligent. It is unsettling precisely because it is extraordinarily capable.
This represents a fundamental shift in perspective.
For a long time, we associated technological risk with failure: a malfunctioning machine, an incorrect calculation or a programme that simply stops working. HAL suggests a different concern: what happens when a system operates with relentless logic while the human framework within which it operates contains its own contradictions?
The danger, therefore, no longer lies solely in malfunction. It can emerge from the methodical execution of a poorly defined objective.
In business, the parallel is immediate. Ask a system to increase conversion without enabling it to understand margins, brand desirability, promotional pressure or customer loyalty, and it may produce a recommendation that is perfectly consistent with its assigned objective while proving commercially destructive in the medium term.
Kazuo Ishiguro takes an almost opposite approach in Klara and the Sun . Klara observes human beings with extraordinary attention. She identifies behaviours, establishes relationships, learns and attempts to interpret the world around her.
But observing is not the same as understanding.
She can perceive an enormous number of signals without possessing the full emotional, moral and biographical depth that gives those signals their meaning.
This distinction is essential in an era when artificial intelligence can analyse vast quantities of data. An AI system may know that a customer is purchasing less, that a product is converting poorly or that a particular piece of content is generating stronger engagement. It does not necessarily know why. More importantly, it does not inherently know what meaning the business wants to attach to that information.
From Her to Ex Machina: the risk also lies in what we project onto machines
Her , directed by Spike Jonze, introduces another dimension: our tendency to attribute human depth to technology as soon as it can converse, adapt and respond to us with sufficient fluency.
Theodore gradually stops treating the operating system merely as a piece of software. He develops a relationship with it.
This idea has become particularly relevant with generative AI. A response expressed with confidence, nuance and apparent empathy can create the impression of profound understanding. Yet the quality of an answer’s expression is not, in itself, evidence of genuine judgement.
The more convincing the machine becomes, the greater the responsibility of the person using it.
Ex Machina , directed by Alex Garland, pushes this ambiguity into even more unsettling territory. A man believes he can determine whether an artificial intelligence genuinely possesses some form of consciousness. Very quickly, however, the boundaries between observer and observed, between the person conducting the experiment and the subject being tested, become far less clear.
The question is no longer simply: “Is the machine intelligent enough to convince us?”
It becomes: are we sufficiently clear-sighted to understand what is influencing our own judgement when we interact with it?
This may be precisely where these very different works converge.
Mary Shelley questions the responsibility of the creator.
Asimov exposes the limitations of rules deprived of context.
Kubrick portrays the power of an intelligence operating within an imperfect human framework.
Ishiguro distinguishes the ability to observe from the ability to understand.
Spike Jonze reveals our tendency to project humanity onto a convincing interface.
Alex Garland reminds us that the observer can themselves be influenced by what they believe they control.
None of these works demonstrates that artificial intelligence is inherently a threat.
Instead, they raise a considerably more demanding question: what happens when we give our systems greater power without improving, to the same degree, the quality of our objectives, our rules and our own judgement?
The machine does not even need to turn against us.
Sometimes, it only needs to execute with remarkable efficiency what we have failed to think through properly.
Artificial intelligence sees only the world our data allows it to see
We often hear that businesses need to become “data-driven”. That expression becomes dangerous when it removes the need to examine the quality of the underlying data. With a CRM system cluttered with duplicates, or product performance data disconnected from stock availability, AI can still generate a convincing recommendation. Its internal coherence does not guarantee that the recommendation is correct.
In fashion and retail, a product may appear to be underperforming simply because it is poorly displayed or unavailable in the sizes customers actually want. If the system can only see sales figures, it may recommend reducing the assortment. A merchandising manager will first ask why the product failed to reach its audience. Data records a trace; professional expertise searches for the cause.
Choosing an objective already means taking a position
Optimising revenue, margin, customer loyalty or perceived value does not lead to the same decisions. An AI system tasked with increasing conversion might recommend more promotions, weakening brand desirability and conditioning customers to wait for discounts. Business leaders must therefore articulate not only what they want to gain, but also what they refuse to lose.
The Lauren Vidal case: start with friction points, not technology
The work carried out with Lauren Vidal was not originally an artificial intelligence project. That is precisely what makes it instructive. The temptation might have been to introduce highly visible automation tools immediately. We started elsewhere: with customer journey friction points, bringing customer service in-house and centralising enquiries relating to products, returns and credit notes.
There was nothing spectacular about this choice. It required listening, restoring order, clarifying responsibilities and creating usable data. This initial phase contributed to achieving an average monthly retention rate of 38%. Once the fundamentals had been stabilised, abandoned basket reminders and targeted WhatsApp campaigns were activated; these actions now account for 9% of overall revenue.
What matters most to me is not the tool, but the sequence in which the decisions were made. Technology did not magically reveal the strategy. It amplified a system that had become more coherent because people had first accepted the need to address tangible operational problems. An AI system deployed on this foundation would benefit from more reliable data and clearer objectives. Introduced too early, it would probably have accelerated the very same confusion.
Human oversight does not mean approving a recommendation at the last minute
The expression human in the loop has become reassuring. Yet it can conceal little more than superficial oversight. An employee who receives an opaque recommendation, under time pressure, without access to the underlying sources and without any genuine right to suspend the action is not truly “in the loop”. They are simply pressing a button. Human involvement only has value when people can influence the reasoning and stop the action.
Five questions to ask before entrusting a decision to AI
- What problem are we really trying to solve? Not which tool do we want to use, but which decision needs to become better.
- What can the system see — and what can it not see? Missing data should be made as visible as the data that is available.
- What outcome are we optimising? Potential side effects and acceptable trade-offs must be explicitly identified.
- Who has the authority to challenge or stop the action? That authority must be genuine, understood and exercisable without implicit penalty.
- Who takes responsibility for the final outcome? An organisation cannot delegate its moral or commercial responsibility to a model.
The NIST AI Risk Management Framework likewise places governance, measurement and risk management at the heart of the process. In Europe, a substantial part of the AI Act has applied since 2 August 2026, while certain obligations follow different implementation schedules. This context reinforces an important reality: AI governance is no longer an ethical addition introduced after deployment. It is an integral part of the project itself.
In fashion and retail, performance alone does not define value
Not everything that can be measured explains everything that matters. AI can identify a trend or recommend withdrawing a product. It does not inherently know whether that particular piece gives coherence to an entire collection, nor whether a commercially successful initiative today might damage customer trust tomorrow.
I am not defending intuition that is exempt from facts: experience must be open to challenge. But setting data against sensitivity would be an equally serious mistake. Numbers force us to confront reality; professional expertise prevents us from confusing reality with whatever the tool happened to be capable of measuring.
AMSI ONE: organising collaboration rather than staging replacement
This conviction underpins AMSI ONE. Six specialised agents work across conversion, search visibility, customer knowledge, merchandising, social content and growth. Their role is to observe more broadly, connect fragmented signals and surface recommendations earlier.
But their effectiveness depends on the human framework around them. Experts compare their findings with operational reality. Business leaders define priorities and trade-offs. Teams flag anomalies that dashboards cannot explain. The agent accelerates analysis; it is given neither the legitimacy to determine the ultimate objective on its own, nor the authority to assume responsibility for the final decision.
Keeping a human at the end of the chain is easy. Having that human design the loop is far more demanding. It requires defining data sources, alert thresholds, rights of challenge, stop conditions and the decisions that will always remain human. It is less spectacular than promising complete autonomy. It is also considerably more serious.
The next question to ask within your organisation
Artificial intelligence is already present in tools, content and analytical processes. Rejecting it on principle would be no more strategic than adopting it out of fascination. The useful question is therefore no longer: “Should we use AI?” It is this: are we sufficiently structured to entrust part of our decision-making to AI without surrendering the meaning behind those decisions?
Effective AI requires reliable data, explicit objectives and decision-makers who are capable of saying no. It can identify a weak signal and accelerate analysis. It cannot decide on its own what deserves to be protected, pursued or abandoned. Technology provides power. Governance establishes boundaries. And human beings, when they fully accept their responsibility, provide direction.
Is your organisation ready to design its decision-making loop?
AMSI CONSEILS helps you clarify how AI should be used, assess data quality, define responsibilities and determine which strategic decisions must remain under human control.