In the previous article, we looked at how a company could get a grip on the eel: start from its processes, own its knowledge and adapt its software. Let's assume this transformation succeeds. That leaves the most personal unknown in our equation: the human being who works inside this transformed company.
The first article introduced Peter Drucker's Knowledge Worker, the professional who creates value by collecting, interpreting and passing on information, and explained why this figure matters so much for Tunisia: IT services, engineering, outsourced accounting, skilled customer relations. The knowledge worker is at once an export engine, a pillar of the middle class and a key contributor to public finances.
For decades, computing was the knowledge worker's best ally. The spreadsheet helped the accountant; the search engine helped the analyst. These tools were bicycles for the mind: they amplified the effort, but the human was still doing the pedalling.
Generative AI changes the physics of cognitive work. For the first time, the tool no longer just structures or retrieves information. It produces content. It can now pedal for us.
Schumpeter comes to the office: the optimisation imperative
Schumpeter's creative destruction, also mentioned in the first article of this series, is now moving through our workplaces. AI is rapidly eroding the economic value of isolated, repetitive cognitive work: the standard summary, the first draft, routine code, everyday translation.
A misunderstanding needs clearing up here. In Schumpeter, destruction doesn't come from within: it comes from competitors. No company reinvents itself for fun; it does so because a rival, somewhere, has already cut its costs or raised its quality. Optimisation is therefore not a choice: it is a constraint the market imposes on every player, all the time.
That constraint now reaches all the way down to the individual desk. When an AI-augmented consultant, developer or lawyer delivers in two days what a peer delivers in two weeks, they aren't just redefining their own productivity: they are helping to redefine the market price of the service for everyone else. The knowledge worker who doesn't change their methods therefore doesn't keep their old level of competitiveness. They lose it, because the benchmark moved without them.
The imperative is asymmetric: doing nothing is not a neutral position but a relative decline. The only variable still in the professional's hands is not whether the transformation will happen, but who will steer it: the professional, by rethinking how they work, or the market, at their expense and on its own timetable.
And the constraint won't stop at the individual. It travels up the chain: the company that doesn't optimise loses ground to the one that does. In the next article we'll see that it eventually reaches states themselves.
So the immediate threat is not necessarily being abruptly replaced by the machine. It is something more insidious: deskilling, the gradual loss of skill through blind delegation. In other words, being transformed instead of steering the transformation.
The chess lesson: birth and death of the centaur
In 1997, Garry Kasparov lost to Deep Blue. Instead of declaring chess dead, he went on to help popularise “advanced chess”: humans playing alongside machines.
The freestyle tournaments of the 2000s then revealed something remarkable: players who weren't necessarily the best, but who had a machine and, above all, an excellent collaboration process, could beat stronger opponents, human or computer. The centaur was born: the human supplies strategy and judgement (the torso) while the machine supplies computing power (the horse's body).
Intellectual honesty demands an admission: in top-level chess, the centaur is now practically extinct. Modern engines have become so strong that human intervention tends to make their results worse rather than better.
Does that mean the centaur model has a sell-by date? Not necessarily, and the reason is decisive.
Chess is a closed system: perfect information, fixed rules, an unambiguous goal. These are precisely the conditions under which the machine can end up pushing the human out.
Knowledge work is generally the opposite: an open system. What “winning” means is up for negotiation. It depends on context, stakeholders, ethical constraints and goals that may shift along the way. A machine can offer an excellent answer to a given question. But deciding which question is worth asking, whom the answer should serve and what collateral damage is acceptable is not merely a matter of computation: it is a matter of framing and responsibility.
So the lesson from chess is not that the centaur always wins. It is something more precise: The centaur stays relevant wherever the payoff function is not fully defined.
That is the space the knowledge worker needs to move their value into: away from tasks that can be locked inside closed systems, a terrain on which they will probably lose in the end, and towards defining goals, contexts and guardrails.
Three stances towards the machine
This analysis yields three modes of collaboration between human and machine. What sets them apart is how each one treats the payoff function of its own work.
1. The Self-Automator: passive delegation
They hand most of the work to the AI and implicitly treat their job as a closed system, as if every task had a “right” answer that the machine would inevitably find.
That's the trap: by defining their work this way, they have already moved it onto the terrain where the machine can push the human out. The result is a double deskilling: they unlearn their craft and gradually lose the judgement needed to check the results.
The Self-Automator doesn't get rationalised away only because AI is too strong, but because they themselves have stopped asking the right questions.
2. The Cyborg: fused co-creation
Human and machine work in constant dialogue: every paragraph, every analysis, every line of code emerges through iteration.
This mode shines on partially closed tasks: the overall framework is set, but execution requires continuous adjustment. The risk lies elsewhere: whoever works in permanent fusion with the machine can lose the distance needed to ask whether the whole thing is still heading in the right direction.
3. The Centaur: a curated division of labour
The dividing line stays sharp: the human keeps control of the what (goals, context, guardrails, trade-offs and evaluation), while the machine takes over much of the how.
This mode is especially suited to situations where the payoff function remains open: strategy, prioritisation, ethical trade-offs, understanding stakeholders or choosing which problem to solve.
Here is the key point: Cyborg and Centaur are not competitors; they are attitudes. The sovereign knowledge worker switches between them deliberately: Centaur for framing, Cyborg for execution, never Self-Automator by default.
The danger arises when the switch becomes unconscious. Whoever sinks into Cyborg mode without ever climbing back into the saddle slowly drifts towards self-automation.
Knowing how to change modes deliberately thus becomes one of the core skills of our time.
AI as a decision amplifier
When the marginal cost of cognitive execution drops sharply, the bottleneck moves.
Yesterday, the scarce resource was often production: drafting the analysis, writing the code, building the slide deck. Today, the scarce resource is increasingly the decision: what to produce, within what framework and towards what goal.
And value tends to pile up wherever the bottleneck sits. Economically, nothing new: the scarce factor tends to set the price. What's new is that this scarce factor is now called judgement.
But this thesis has a darker twin: AI is also a decision amplifier.
Good framing gets multiplied by the machine. So does bad framing.
Yesterday, someone who slipped a faulty assumption into an analysis produced one flawed document. Today, someone who slips the same assumption into an automated process can produce flawed documents in bulk, at machine speed, professionally formatted and with every appearance of rigour.
That is the Self-Automator's real weakness: not that they no longer decide, but that they decide unconsciously. Every unthinking prompt, every implicit assumption, every result accepted without checking is a decision that isn't recognised as one.
They aren't just delegating their work: without always noticing, they are delegating their judgement.
Production competence and verification competence: the birth of the Creative Generalist
To understand the next evolution of the knowledge worker, let's start from a simple observation: until now, a professional's value rested largely on their disciplinary silo. The accountant knew accounting; the developer, code; the lawyer, law. Crossing the boundaries between these fields was expensive: it took years of study, long hands-on experience or hiring a second expert.
AI slashes that cost of access dramatically. Specialist knowledge from an unfamiliar field doesn't become instantly mastered, but it becomes much easier to consult, question and put to work.
Two concrete examples:
- A business strategist is preparing an overhaul of a customer service operation. Yesterday, they would have set out their vision in a slide deck, then waited weeks for the technical team to turn it into specifications. Today, they can explore different software architectures with AI, understand their main constraints and walk into the technical meeting with sharper questions, or even with ideas their original training would never have let them formulate.
- An economist wants to test a hypothesis on export data. Yesterday, they would have called in a statistician or dropped the analysis. Today, they can run a first exploration themselves, guided by the machine, and save the expert for validating methods and results.
This is what we might call the Creative Generalist: a professional whose value no longer lies only in accumulating specialist knowledge, but in their ability to connect fields nobody thought of linking and to draw new solutions from them.
For a country like Tunisia, whose strength may lie more in the versatility of its graduates than in the depth of its R&D budgets, this figure deserves particular attention.
But an objection follows immediately, and it's a serious one: isn't this generalist simply a Self-Automator with a better self-image?
A strategist who adopts AI-generated technical concepts without being able to assess their validity is not a generalist. They're an adventurer with an oversized ego.
To answer that objection, we need a distinction that becomes central in the age of AI. Let's break down what it means to “master” a field. Two distinct competences hide inside it.
Production competence means being able to do it yourself: write the code, build the statistical model, draft the contract. This is the one that traditionally required years of training, and AI can now take over a growing share of it.
Verification competence means being able to judge what has been done: recognising that a software architecture is fragile, that a statistical correlation is misleading or that a contract clause is risky, even without being able to produce everything yourself at the same level.
The distinction isn't new. A plant manager can't necessarily do every weld, but must be able to spot a faulty manufacturing process. An editor-in-chief hasn't written every article, but knows how to catch a dubious claim.
What changes is that this asymmetry could become the standard working model for a growing share of knowledge workers: the machine produces more; the human frames, verifies and owns the decision.
The legitimate Creative Generalist can then be defined more precisely. They don't need full production competence in every field they connect, but they must build a minimum level of verification competence in each: enough depth to notice when the AI hands them a plausible absurdity. Not to practise fully in that field, but to know when to decide, when to doubt and when to call in an expert.
Which leaves a practical question: how do you acquire this verification competence without going through all the years of production on which it was traditionally built?
Symbiotic learning with the machine provides part of the answer: arguing with the models rather than passively consuming their answers, challenging their claims, asking for their sources, looking for counter-arguments and turning every error you catch into a chance to refine your own mental model.
But AI cannot be at once the sole producer, the sole teacher and the sole judge of its own mistakes. This learning also has to rest on external references, real cases, supervised practice and the occasional intervention of experts.
So the honest formula is this: Creative Generalist is not a status that access to AI confers automatically. It's a discipline you build.
Whoever skips the verification-competence stage is not a centaur with a wider range: they're a Self-Automator rationalising themselves away in several disciplines at once.
And in a system that amplifies my judgements, the quality of those judgements becomes the main lever I still hold.
The Gen Z dilemma: the flaw in the centaur factory
This analysis, however, runs into the Gen Z Dilemma identified in the first article: if the machine does the first analyses and handles the simple cases, how will young professionals gain the experience needed to handle the complex ones? We can now state the stakes more precisely. Verification competence, the very thing that separates the centaur from the Self-Automator, is forged by confronting mistakes, by repetition, by practice and by doing the basic tasks.
The centaur of 2035 has to start training today. But on which tasks, if the machine has already swallowed them? Answers exist, but they require deliberate redesign:
- reinvent the junior role: stop asking juniors only to produce a first draft, and also have them audit the AI's draft under the supervision of an experienced professional;
- reorient education: assess less the mere ability to produce an answer, which the machine handles better and better, and more the ability to formulate a relevant question, make one's assumptions explicit and critique the answer obtained;
- institutionalise a “culture of override”: explicitly train people to change gait, that is, to know when to frame, when to iterate, when to verify and when to take back control;
- preserve spaces for hands-on learning, even when they seem less efficient in the short term, so as not to sacrifice the training of tomorrow's experts on the altar of immediate productivity.
None of this will happen on its own. It is a choice about organisation, investment and the passing on of knowledge, not a mere by-product of technology.
Creative destruction changes scale: from the individual to society
So far, we've been reasoning at the level of the individual. Let's widen the frame, because this is where the Tunisian equation gets tricky. Every optimised process mentioned in the second article (specialised document processing, augmented customer service, agent-assisted development) means the same volume of work can be done by fewer people. If a team of ten analysts produces what twelve produced yesterday, the productivity gain is 20%. Multiplied across a whole sector, that gain can feed three scenarios, which are not mutually exclusive.
Scenario 1: growth absorbs the gains The company, now more competitive, wins new markets and keeps its headcount, but transforms the content of the work. This is one of the scenarios observed in earlier technological revolutions. It does, however, require expanding markets, demand that can pay, and a successful move upmarket. Productivity doesn't automatically create its own demand.
Scenario 2: the silent reduction The company produces as much with less. Not necessarily through spectacular layoffs, but by not replacing people who leave, freezing junior hiring or never creating roles that would otherwise have existed. This is probably the most discreet form of destruction, and the hardest to measure. It leaves few traces in layoff statistics, but shows up later in the employment figures for young graduates.
Scenario 3: the relocation of destruction The first article already flagged it: the automation threatening Tunisian knowledge workers doesn't need to happen in Tunisia.
If a European client automates part of its support or development, it can cut back on outsourcing. The destruction then happens in Tunis, while the creation of value, capital or intellectual property concentrates elsewhere.
And creative destruction never promises that what gets created will appear in the same place, for the same people, or at the same pace as what gets destroyed.
A 45-year-old consultant whose standard analysis tasks disappear doesn't automatically turn into a supervisor of AI agents. A 24-year-old graduate doesn't automatically find the junior job that no longer exists.
This is where the loop closes on society as a whole.
The knowledge worker is not just a producer: they are also a taxpayer, a contributor to social insurance and a consumer. Their income funds pensions, health insurance and their children's schooling.
Which brings us back to the paradox set out in the first article: just when society would most need resources to train, protect and retrain, the traditional sources of those resources could weaken. But we can now watch the precise mechanics at work, process by process, job by job.
The individual centaur may be able to save itself. But who's going to pay for the stable?
That will be the subject of the next article: the governance puzzle. How do we adapt taxation, social protection and regulation to an economy in which digital capital produces a growing share of value? And how can Tunisia turn this constraint into an opportunity to rethink its social contract?
Because a society is measured not only by the productivity of its machines, but by what it chooses to do with the people they free up.
Or leave standing on the platform.