Series note: This is the second article in a five-part series on AI, economics and advisory work.
Much of the public debate on AI begins with the wrong question: is AI really intelligent? It sounds deep, but it is analytically careless. Often it is a status dispute dressed as philosophy.
Intelligence is not one thing. A person may have strong memory and weak judgement, strong verbal fluency and poor self-monitoring, strong reasoning and weak social perception. Firms, meanwhile, do not buy intelligence in the abstract. They buy combinations of cognitive functions that help them frame problems, interpret evidence, challenge assumptions and make decisions.
That is why the most useful comparison between human and artificial intelligence is not a philosophical duel over whether machines 'really think'. The more practical question is which layers of intelligence AI can already reproduce, how reliably, and which layers still carry a human premium.
My own interest is practical rather than academic. In strategy and transaction work, intelligence is tested by decisions, not definitions. The question is not whether an answer sounds clever. It is whether the work helps a management team, board or investor decide what to do under uncertainty.
This article argues four things. First, intelligence is better understood as layered than as a single quantity. Second, the layers are moving: AI is progressing from language fluency to deliberative reasoning, longer context, memory-like retrieval, tool use and bounded agency. Third, many valuable business outputs depend less on individual brilliance than on collective intelligence. And fourth, as reproducible cognition becomes cheaper, the human premium moves again - toward salience, target-setting, judgement under uncertainty, persuasion and accountable action.
The first companion essay supplies the technical capability map beneath this argument. This article changes level: from what AI systems can do to how human, artificial and collective intelligence combine inside organisations - and where the human premium moves as reproducible cognition becomes cheaper.
AI does not need a human mind. It needs a role in the work.
1. Intelligence is layered, not scalar
Psychology has long resisted reducing intelligence to the everyday notion of being 'smart'. For business purposes, the more useful move is to decompose intelligence into capabilities: selecting what matters, holding context, retrieving knowledge, reasoning from evidence, learning from feedback, monitoring one's own thinking, reading social reality and initiating action.
The point is not to settle a century of psychometric debate; it is to create a better lens. Once the layers are visible, the debate becomes less mystical. AI can be strong in one layer and weak in another. Humans are the same. The economically relevant question is configuration and direction of travel.
A simple map helps.
| Layer | AI/current capability | Human or system premium |
|---|---|---|
| Problem selection and attention | AI can prioritise, rank and flag anomalies inside a supplied frame | knowing what matters before the frame is settled, framing trade-offs and mobilising attention. |
| Memory, recall and synthesis | AI is increasingly strong across long context, text, code, data, tools and prior patterns | salience, source judgement, confidentiality, context and the meaning of exceptions. |
| Reasoning and abstraction | AI is increasingly strong in bounded frames, multi-step reasoning, issue trees, hypotheses and first-pass analysis | causal judgement, assumption quality and knowing when the objective or frame is wrong. |
| Hypothesis and idea generation | AI can generate options, alternative explanations, risks and value-creation angles at scale | selecting which possibility deserves belief, resources and action. |
| Operational execution | AI can increasingly write code, draft procedures, navigate tools and execute bounded steps | live context, permissions, sequencing, verification, rollback logic and risk ownership. |
| Self-monitoring and challenge | AI is improving when review roles, tool checks and uncertainty flags are explicit | knowing when confidence is misplaced because of incentives, silence, politics or hidden constraints. |
| Social understanding | AI is uneven because it does not inhabit the relationship | trust, persuasion, empathy, negotiation and stakeholder reality. |
| Agency and accountability | AI can recommend and execute bounded steps, but it does not own consequences | commitment, responsibility, courage, legitimacy and consequence-bearing. |
This map should be read as a moving frontier, not a fixed inventory: from fluency, to synthesis, to reasoning, to memory, to tool use and bounded agency. The question is not whether humans or AIs are smarter in general. It is which layers are being reproduced, which are becoming cheaper, and which still depend on human and institutional judgement.
2. AI is becoming stronger at reasoning, memory and bounded agency, but still under a frame
Once intelligence is decomposed, current AI looks less mysterious and more economically legible. The frontier has moved beyond fluent text. Earlier systems could compress, draft, classify, translate and synthesise. Current frontier systems increasingly combine pattern induction, deliberative reasoning, longer context, tool use and bounded agency. That is a real capability shift.
But it should not be misunderstood. Reasoning inside a frame is different from choosing the frame. Optimising for a stated goal is different from knowing whether the goal is the right one. Producing code is different from making a system safe, deployed and institutionally acceptable. Retrieving a similar case is different from knowing whether the analogy should influence the decision. AI is becoming stronger at cognition under specification; humans and organisations remain scarce in salience, target selection, social interpretation, verification standards, legitimacy and accountability.
Human and artificial intelligence also differ by architecture. Human intelligence is embodied, socially situated, motivated and consequence-bearing. AI intelligence is symbolic, tool-mediated, objective-specified and cheap to replicate - powerful and scalable, but dependent on context, permissioning and governance. The point is not that one kind is real and the other is not. Their strengths and failure modes are different.
The key point is economic rather than philosophical. AI does not need to reproduce the full human mind in order to reshape work. It only needs to reproduce enough of the relevant task bundle at acceptable quality and cost. That threshold has already been crossed in more places than many incumbents are comfortable admitting.
This is also why the common retort - that AI still makes mistakes - is less decisive than it sounds. Humans make mistakes as well. The relevant comparison is not machine perfection versus human perfection; it is one production system against another. How such a system surfaces, challenges and contains error is the subject of the next article.
3. In firms, intelligence is collective
The next conceptual mistake is to focus only on individual intelligence. Many of the most valuable outputs in business are produced by teams, routines and institutions that allow different people to think together. A good board, investment committee or consulting team is not simply a pile of IQ points. It is a mechanism for combining differentiated perspectives, storing shared context, contesting weak reasoning and converging on a decision.
Consulting has long been a practical case study. The industry did not reduce mistakes by assuming that one partner, however bright, would always be right. It reduced mistakes by building structured problem solving, role differentiation, challenge, review and synthesis into the production model: analysts gather and process facts, managers challenge structure and integrate streams, partners reframe, stress-test and connect the work to client stakes and decision rights.
The same is true in transactions. A strong diligence process is not just a spreadsheet or a memo. It is a system for pressure-testing a thesis: commercial logic, market structure, management claims, downside risk and the decision a buyer is actually being asked to make. The intelligence is in the interaction between roles, not only inside one head.
This is where AI becomes more interesting. The issue is no longer only comparison; it is combination. Human and artificial capabilities can be arranged in ways that improve the quality, speed and consistency of decision work. That does not remove the need for human ownership. But it does make the organisational question more important than the model comparison.
4. The human premium: operational judgement and accountable action
This brings the argument back to scarcity. If more cognition becomes reproducible, what becomes scarce next? My current answer is entrepreneurship, properly defined. It is not merely having ideas - models can already generate many. Entrepreneurship is choosing which uncertain possibility deserves commitment, mobilising resources around it, persuading others, absorbing doubt and bearing the consequences of being wrong.
Economists have long distinguished measurable risk from true uncertainty, and the distinction matters here. AI is often strongest where the task can be framed, the objective stated, the data supplied and the evaluation rule approximated. Entrepreneurship begins earlier and reaches further: it selects the frame itself, and it acts before the world has yielded enough data to make the choice comfortable.
In PE and M&A, this is often the whole game. The stated task may be "produce a market view", but the real optimisation target is whether the investment thesis survives pressure. The stated task may be "build a value-creation plan", but the real target is a sequenced intervention management can execute under capital, timing and capability constraints. The scarce skill is knowing what the work is really optimising for.
That is why agency, judgement and accountability remain central. Not because machines are incapable of novelty, but because they do not yet own conviction, responsibility or consequence in the human sense. They can expand the search space, improve the analysis and strengthen execution. The act of commitment still sits elsewhere.
Closing thought
The phrase artificial intelligence has always encouraged a misleading comparison: machine versus person, as though intelligence were a single contest. A better frame is layered, social and dynamic - and many of the most valuable outcomes in modern firms arise from how capabilities are combined across people, tools and institutions.
The next question is therefore practical rather than philosophical: how should organisations design the conditions under which human and artificial capabilities actually work together, remember what matters, challenge one another and improve the quality of decisions? That is where the next article goes.
Disclosure. This article was AI-assisted. The argument, judgement and conclusions are my own.