The Risk of Automating the Organisation We Already Have
AI may allow us to automate large parts of knowledge work. But the organisation we are automating was designed around humans.
In 2026, OpenAI used thousands of AI agents to attack the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems.
The experiment is remarkable both for its mathematical result and for how the work was organised.
Thousands of agents explored different approaches in parallel. Promising ideas spread between them. Effort shifted as new information emerged. According to OpenAI, the Navier–Stokes effort involved around 10,000 concurrent agents, 2.7 million messages and 130 billion output tokens.
Try to imagine organising 10,000 human mathematicians in the same way. Coordinating them would be an enormous undertaking.
Why Organisations Look the Way They Do
Organisations are shaped partly by the limitations of humans.
Expert attention is scarce, communication takes time, and knowledge is distributed across individuals and teams. People can only follow a limited number of problems simultaneously.
So we divide work across specialised roles and teams. Requirements, design, analysis and review may be performed by different people, with information moving between them through documents, meetings, databases and formal handovers.
There are good reasons for much of this structure, particularly in a complex and regulated industry like aerospace.
Asking a thousand engineers to independently investigate every difficult design decision would be prohibitively expensive. The same is true of involving every relevant specialist continuously across the organisation.
So organisations create interfaces and processes to make scarce attention manageable.
AI changes some of those constraints.
Automating the Process Map
There is a natural tendency to introduce new technology through the organisation we already understand.
The most immediate use of AI is therefore to automate individual activities within this existing organisation. AI agents can review requirements, analyse test results, prepare documentation or search previous programmes for relevant information.
This can create substantial value.
But imagine taking the process map of an engineering organisation and gradually replacing each box with an AI agent. Eventually we might have automated much of the work while preserving the structure created to coordinate humans.
Agents can exchange information almost instantly, examine many alternatives in parallel and be replicated when additional capacity is useful. Some of the boundaries between today’s activities may therefore become less important.
That raises a larger question than which boxes we can automate:
Which of those boxes are still required?
Start With the Purpose
Engineering processes often serve essential purposes. Their current implementation reflects both those purposes and the practical realities of human organisations.
A review may be necessary. The same review could potentially happen continuously as the work develops rather than through a document passed between departments.
Traceable engineering evidence may be necessary. Future engineering systems could capture much of that evidence as the work happens.
This suggests another way to look at the problem. Instead of starting with the pieces into which we have divided the work, start with the activity end to end.
What needs to happen from beginning to end? Which capabilities are required along the way? Where is human expertise needed? Where can AI perform the work? Where must a person exercise authority or carry responsibility?
The work can then be routed between human and machine capabilities accordingly.
Today, much of that routing is embedded in the organisation itself. A task goes to a particular person because it belongs to their role, team or discipline. With AI, the route could depend more directly on the capabilities the work requires.
That does not make roles or organisational structures disappear. People still need homes for expertise, development, authority and responsibility. But the boundaries between roles need not determine the boundaries of the work.
A Different Kind of Organisation
The OpenAI experiment gives us a glimpse of what becomes possible when work can be organised differently.
A difficult problem might temporarily receive hundreds or thousands of parallel attempts. Different approaches could compete. Successful ideas could be challenged before a human ever sees them. Specialist agents could be created for a particular question and disappear once it is resolved.
The organisation of the work could change with the problem.
Doing this with humans is expensive because creating teams, establishing interfaces and coordinating large numbers of people takes substantial effort. AI changes that equation, but it does not make intelligence free. The Navier–Stokes effort alone consumed around 130 billion output tokens. Human attention is scarce and expensive, but machine intelligence has a cost too. The organisational problem is how to use both well, rather than assuming that anything AI can do should automatically be assigned to AI.
The opportunity is to preserve what the organisation must achieve while reconsidering both how work is divided and how it is routed between humans and machines.
If we were designing an engineering organisation today with AI available from the beginning, how would we design it?
References
AI helped to research and write this article.