When Gutenberg set up his workshop around 1450, his contemporaries predicted the death of the scribe. Why copy by hand what a press could now reproduce? Five centuries later, humanity has never written, read, edited, translated, or published more. The technology that seemed to threaten the work of the mind did not abolish it. It displaced it. And it made it more demanding.

Something similar may be happening to collective intelligence today.

The question deserves to be asked without flinching. If artificial intelligence gives each of us instant access to nearly universal knowledge, why still gather experts, run workshops, debate for days?

If a machine can synthesise ten thousand scientific papers, compare strategies, simulate objections, translate entire bodies of knowledge, generate scenarios and produce recommendations, what is left for the human collective? The conventional answer would be to say that artificial intelligence and collective intelligence are complementary. True. But it stops short.

The harder question is this: does AI make part of collective intelligence obsolete? Yes.

But not all of it. And not where one might think.

The end of a certain kind of collective

Organisations long needed collectives because no individual could know everything. Knowledge was scattered across functions, geographies, experts, frontline teams. The fragments had to be reassembled into a coherent picture. Collective intelligence served, in large part, to compensate for the dispersion of knowledge.

That is precisely the function AI now dismantles.

A single person can obtain in minutes a first-cut analysis that once demanded weeks of meetings or a consulting firm. Much of what organisations called collective intelligence was, in truth, a slow way of sharing information. A meeting to catch up. A workshop to surface ideas already available elsewhere. A committee to reformulate what no one had bothered to read.

These forms are dying. And, frankly, that is no bad thing.

Knowing is not deciding

The misunderstanding begins when we confuse knowing with deciding.

Generative AI responds to prompts: it answers questions, synthesises information, compares options, reformulates ideas and proposes outputs. Autonomous agents go one step further. Rather than waiting for a question, they pursue an objective: they plan the next steps, call on tools, make decisions and adjust as they go. Where generative AI assists, agents act. On bounded, repetitive, measurable tasks governed by a clear objective, they can already outperform us. Faster, more consistent, never tired.

But neither generative AI nor agents abolish collective intelligence. They shift its centre of gravity. Generative AI displaces the collective from informational catch-up. Agents displace human decision from a layer of operational execution. Together, they force collectives to climb one floor higher in the decision chain: from sharing knowledge to framing, arbitrating and assuming responsibility.

Optimising is not arbitrating

The real frontier no longer runs between humans and machines. It runs between two families of decisions.

In some cases, the objective is clear. Reduce a delay. Avoid a stockout. Detect fraud. Sort cases by urgency. The success criterion is known, measurable, tracked. Here, AI often outperforms a collective. The human task is to define the frame: what to optimise, within what limits, with which exceptions, against which alert thresholds?

In other situations, the objective itself is contested. Growth or profitability? Personalisation or fairness? Speed or safety? Margin or accessibility? Here, AI can clarify. It can compare, model, simulate. But it cannot decide alone, because the difficulty is no longer cognitive. It engages legitimate, but divergent, interests. It demands that someone say what matters most, what to sacrifice, what to refuse to optimise.

AI excels when the question is well-posed. Collective intelligence becomes indispensable again when the question itself is in dispute.

Where power now resides

This is where the real work of leaders shifts. Human decision does not vanish. It moves upstream. It relocates into the mandate.

Picture a major retailer deploying a dynamic pricing agent with a single mandate: maximise margin. In some neighbourhoods, the algorithm notices that customers keep buying despite price hikes. Not because they are wealthier, but because they have fewer alternatives: fewer competitors nearby, less mobility, less time to compare, less access to other channels. The system raises prices accordingly. Mathematically flawless. Socially, it charges more to those with the fewest choices. The objective function was defined. It was not legitimate.

The Amazon case is now a textbook lesson. Between 2014 and 2017, the company built an AI tool to rank job candidates, trained on a decade of resumes. Because tech was male-dominated, most resumes came from men. The system learned to favour male candidates, penalising CVs containing the word women’s. Amazon tried to neutralise the algorithms, failed to guarantee fairness, and quietly abandoned the project (Reuters, 2018). The algorithm had perfectly learned the past. That was precisely why it was about to obstruct the future.

One might object that the argument runs the other way. Why insist on collective intelligence when one-person companies, fully augmented by AI, are already changing what an enterprise can be? Base44 is the case everyone cites. Launched in February 2025 by Maor Shlomo, the app-development platform was sold to Wix six months later for 80 million dollars. One founder. AI-native tools. A swift exit. A glimpse, perhaps, of what comes next.

The closer one looks, the more the case turns against itself, though. By the time of the sale, Shlomo had already brought in eight collaborators. He did not scale alone. The sale came at the moment when solo decision-making appeared to reach its limits. AI lets one start faster, work alone, last longer. Scaling demands something else: trust, expertise, governance, distribution.

The pure solopreneur, today, exists as horizon and as story, rarely as a durable organisation built for scale. AI does not abolish collective intelligence. It pushes back the threshold at which collective intelligence must engage.

What the collective still owns

Four functions remain decisively human.

The first is legitimacy. A technically sound decision can fail when it is experienced as imposed, foreign or misunderstood. Machines improve the analysis. They do not, by themselves, generate the trust that makes action possible.

The second is the emergence of what is not yet formulated. AI is valuable at detecting weak signals once they have left a trace in the data. But not all signals are yet data. Picture a sales committee where the AI confirms every indicator is green. Then an account director quietly says: “My main contact at our biggest client doesn’t return my calls as quickly. He used to lead our meetings, now he sends a junior. Something is going on, I don’t know what.” No data confirms it. Six months later, that client leaves for a competitor. The signal was there, six months earlier, not in any system, but in the trained perception of an experienced professional, surfacing only because a collective took the time to listen. AI catches the faint traces. The collective surfaces what has not yet found its form.

The third is responsibility. Neither generative AI nor agents answer for what they produce. They do not carry the consequences. On August 1, 2012, the U.S. broker Knight Capital deployed a new high-frequency trading algorithm. A single line of faulty code was enough: in 45 minutes, the system executed over four million erroneous orders across 154 stocks, accumulating 440 million dollars in losses, roughly three times the firm’s annual profit. The seventeen-year-old company never recovered (SEC Form 8-K, August 2, 2012). Servers did not bear the consequences. People did.

The fourth is ethical arbitration. AI executes whatever objective it is given. It does not ask whether the request was wise, nor whether it captures what the institution truly intends. Think of an autonomous vehicle facing an unavoidable collision: protect the passenger or the pedestrian, swerve toward the elderly woman or the child? MIT’s Moral Machine experiment put such dilemmas to more than two million people across 233 countries (Awad et al., Nature, 2018). The findings were unequivocal: moral preferences diverge sharply by culture, by age, by economic context. There is no universal answer encoded in the machine. There is no universal answer at all. Yet the car must drive, and the algorithm must be written. The model does not choose. The carmaker and the legislator do, and must stand behind that choice. This case is only one example. Every institution that deploys AI faces an equivalent arbitration, whether it knows it or not. A bank ranking loan applications, a hospital triaging patients, an insurer pricing risk, a recruiter screening candidates, a platform moderating content: each defines, by the objective it gives the machine, what it considers worth optimising and what it accepts to sacrifice. Ethics, in the age of AI, is knowing what to optimise, and what to protect from optimisation.

What the decade will demand of us

This analysis could sound abstract while layoffs and reorganisations linked to AI multiply. It would be indecent to write about AI without facing that reality.

But look carefully. Which activities are most exposed? Those that aggregate knowledge, coordinate information flows, and synthesise data for higher layers of the organisation. In other words, what generative AI now automates. Which responsibilities resist, or even become stronger? Those tied to revenue, hard-to-encode expertise, human judgement, negotiation, legitimacy and risk.

The distribution is not accidental. Roles built around defined objectives are most exposed. Roles that operate in zones of arbitration, ambiguity, responsibility and legitimacy endure. AI is not killing collective intelligence. It is stripping away its weakest tasks and forcing it to perform its highest functions.

As printing did not end writing, AI will not end collective intelligence. After Gutenberg, copying was no longer enough. One had to edit, choose, translate, criticise, publish, interpret. After AI, gathering minds to share knowledge will no longer be enough. We will gather them to decide what that knowledge obliges us to do.

We will need fewer meetings to know.

We will need better ones to judge, arbitrate ethically and bear responsibility.

The challenge of the decade will not only be to learn to use artificial intelligence. It will be to raise collective intelligence to the level of power that AI has just placed in every individual’s hands.