Organising Intelligence

Most people use AI through one exchange, usually inside a chat: a person asks, and AI answers. When a difficult question has you stuck, the experience resembles Phone a Friend, the lifeline from the quiz show Who Wants to Be a Millionaire? Someone clever is called to help, but the help remains a call; the person at the other end can advise you, while the work stays on your side.

That is the distinction I want to make in this first article: the important unit of AI may increasingly be the organisation around the conversation, more than the intelligence answering inside it.

A brilliant analyst with only a phone

Imagine that the best real-estate analyst you know has moved to the other side of the world. You can contact her whenever you want, but only through WhatsApp—or Signal, which you should probably use instead. She can explain how to compare two properties, suggest which figures to check, or draft a note for an investment committee. She is capable, available, and often extremely useful.

Still, she only has a phone. She cannot open your company database, inspect the latest reports, run a calculation on the full portfolio, or update your dashboard. If you need any of that, you must find the information yourself, send it to her, and then perform the action she recommends. That creates friction at every step: the analyst may provide the intelligence, but you are still carrying the work.

This is roughly how a chatbot works. It can reason about what you give it, but it remains confined to the conversation and cannot see anything else. We tend to judge its intelligence by the quality of the answer, even though answering is only one small part of useful work.

No company would give such an analyst nothing but a messaging app and expect the full value of her expertise. It gives her a desk, access to the relevant information, and tools suited to the job; once we do the same for AI, the relationship changes.

Giving the analyst a desk

Here, I use agent in its broadest useful sense: an autonomous unit that receives an objective, acts within some boundaries, observes what happened, and adjusts. An agent can be a human or an AI; the word describes its role in the work, not what it is made of.

When the agent is an AI, its desk consists of access and tools. Instead of only explaining how to calculate a rental yield, it can open the data, perform the calculation, check whether the result makes sense, and try another route if something fails. The tools do not need to be exotic: a document library, a calculator, a search function, a database, or the ability to create a chart may be enough to move from talking about the work to doing part of it inside a controlled environment.

This is the first step from a chatbot to an agentic organisation, yet one capable agent with a desk and tools still does not amount to a company: attention remains finite, delegation has limits, expertise is uneven, and some tasks simply require more than one pair of eyes.

One brilliant person is still one person

Now ask our analyst to review an entire property portfolio. She must find the latest information, compare it with last week, identify unusual movements, verify each important figure, decide what matters, and prepare the result for several audiences. She may be excellent, but the assignment bundles several kinds of work; some require concentration, some can progress in parallel, and others benefit from knowledge or skills she may not have.

Humans are not especially reliable when we work this way. We forget checks, lose information, and become worse as our attention is divided. Companies respond by specialising. One person gathers information, another analyses it, a third reviews the result, and a manager keeps the work pointed toward the decision.

Marvin Minsky offered a similar picture of intelligence in his 1986 book, The Society of Mind. His agents were not modern AI assistants, but small and limited mental processes, none of which contained the whole mind. What looked like one intelligence emerged from the way these limited parts worked together, using different forms of organisation for different problems.

Managers already work with similar raw material: no employee holds the whole picture, everyone has limited attention and different skills, so responsibilities, information, decisions, and checks are arranged around a common objective. Agentic organisations start to follow the same principle.

Léa's Monday review becomes a team assignment

Consider Léa, a portfolio analyst who begins every Monday with one simple question: what changed across our buildings, what deserves attention, and why? The question sounds small, yet answering it requires data collection, comparison, interpretation, verification, and presentation.

In an agentic structure, Léa does not give every responsibility to one AI agent. A coordinating agent can divide the assignment. One specialist finds the relevant reports and portfolio data. Another compares the buildings and flags unusual changes. A third checks the calculations and traces important figures back to their sources. A final specialist prepares the result for Léa.

These specialists do not need invented names, faces, or artificial personalities. Each is an agent because it owns a responsibility and can act within it; the role may be filled by a human, an AI, or a combination of both, depending on the process and the consequence of the decision. Keeping the responsibilities separate helps each part of the work remain focused and makes failures easier to locate.

Seen this way, resolving the problem starts to look remarkably familiar to anyone who has designed a team. Agents have limited capacity and partial responsibilities, so we use organisational principles to coordinate them: finding the data, analysing it, checking the result, and preparing an answer become distinct responsibilities rather than one long improvisation. The value comes from arranging limited capabilities into useful work, rather than pretending that any single agent knows everything.

Coordination is part of the work

A company does not improve each time it adds another employee. Put twenty people in a room without clear responsibilities and they will repeat tasks, wait for one another, and produce incompatible answers; more people can make the result worse because coordination has a cost. A recent Anthropic experiment found a similar effect among AI agents, with coordination becoming much harder once their tasks depended on one another.

A useful agentic structure therefore needs more than a collection of agents. It needs a clear objective, separated responsibilities, shared information, and rules about what must be checked or escalated. The coordinating layer keeps the work aligned, while each specialist retains freedom over how to complete its own part.

Iteration matters for the same reason. If the reviewer finds that a vacancy figure uses the wrong period, the work can return to the analyst, be corrected, and be checked again. Reliability grows from this ability to revisit the work, because a useful organisation is built around the expectation that first attempts may be incomplete or wrong.

The human can sit in different places

It is tempting to draw a simple hierarchy with a human at the top and AI agents underneath. Sometimes that is the right picture. Léa can define the objective, set the boundaries, and approve the final conclusion while agents handle the intermediate work.

Other processes require a human in the middle. A valuation specialist might review one assumption before the system continues, while a compliance expert approves the use of sensitive information. In more exploratory work, a human and an agent may work side by side, each reacting to what the other finds.

The aim is to place human attention where judgment, responsibility, and experience add the most value. In Léa's case, the wider organisation can collect, compare, check, and prepare, while she remains responsible for what the analysis means and which decision it should inform.

From a collection of features to an organisation of capabilities

Most digital products are presented as a collection of features. The user learns where each function sits, performs the steps, and moves information from one screen to the next; even when AI is added, it often appears as another feature inside the same structure.

Products built around AI point toward a different arrangement. The user describes an objective, and the system brings together the right combination of information, tools, specialists, and checks around it. The design question shifts from “which button should the user press?” to “who should do what, under which rules, and how will the result be checked?”

Zoom out once more and the pattern can repeat. A portfolio review unit can sit beside market monitoring, research, and client delivery. Their internal shapes can differ because their work differs. Shared rules and human direction keep the larger organisation coherent.

Not every task needs this structure. A simple question is often better answered by one agent, just as a short assignment does not require a new department. The organisational approach becomes useful when the work is too broad for one actor, when different skills are needed, or when the result is important enough to verify independently.

Put the whole progression together. A chatbot is a brilliant colleague with only a phone. An agent gives that colleague a desk and tools. An agentic structure adds specialists, responsibilities, shared rules, and a way to correct the work.

The intelligence at the centre may be exactly the same at each step; what changes is the organisation around it, and that organisation determines what the intelligence can actually accomplish. The chat window remains useful, but it is only the front door.

References

• Minsky, M. (1986). The Society of Mind. Simon & Schuster. See the MIT Media Lab overview.
• Anthropic (2026). “Patterns and problems in emerging multiagent systems.”