Four Design Moves for Building a Humans + AI Organization

The question that is moving to the heart of my work is: what will be the best-performing Humans + AI organizational structures? To embark on this journey we need to lay out an often-unacknowledged truth.

Every organization has a documented structure. This includes the org chart, the process maps, the workflow definitions inside the software, and the role descriptions often written far too long ago and underperforming throughout.

Every organization also has a structure of the realities of how the work actually gets done, the networks and patterns of:

who people really go to when something breaks

what gets escalated, and what quietly does not

which numbers are trusted, and which are checked twice

what is easy

what has simply always been done this way

That second structure is not what anyone necessarily wants or would design. It is what has accumulated, out of what worked, what was quick, who was trusted, and what became habit.

New people learn it within weeks, mostly by observing. AI systems only ever see the documented structure.

That gap sets the limits on the return on almost every AI investment now being made. Give an agent the documented process and it will execute the documented process, at speed, at scale, with perfect fidelity to a description that was never accurate.

The race to make the implicit explicit

A wave of companies has formed around this problem. Their method is to observe rather than to ask. They instrument the tools, watch how work actually flows, mine the messages and the meetings and the handoffs, then reconstruct the real process from evidence rather than from whatever somebody wrote down. Substantial near-term value sits there. Much of what we loosely call tacit knowledge is only undocumented knowledge, and some, but importantly not all, undocumented knowledge can be captured.

Getting organizational structure right carries a large payoff. Researcher Canhui Liu ran 8,000 identical knowledge-work tasks through seven organizational forms and found agent-native structures 395% more efficient than the forms that imitate how humans organize. The committee structure finished last, which will surprise nobody who has sat in one. The strongest result came from a meta-organization that switched form according to the task. Human organizational structures exist for good reasons, and almost none of those reasons apply to agents.

What AI cannot currently access

Observation captures what people do. It struggles with why they did it, and it cannot capture what they chose not to do.

The most valuable thing an experienced person brings to a role is a sense of when to slow down. They notice a request from a client who never makes requests, and they know it matters far more than its content suggests. They know which colleague will quietly fix a problem and which one will turn it into a meeting. They know the number in the report is technically correct and practically misleading, because they know what happened in the warehouse in March. None of this sits in the data. Very little of it survives being written down, because writing it down removes the context that made it a judgment rather than a rule.

Yale School of Management professor K. Sudhir states the design implication precisely in a recent Harvard Business Review article. Ask the people in a role what they notice that never reaches the data, what they care about beyond their job description, and when they slow down. The distance between their answers and the documented process is your design specification.

Friction is doing work

Organizations are slow because people hesitate. They interpret instructions rather than executing them. They push back on things that look wrong, ask a clarifying question, sit on a decision overnight, occasionally refuse outright. All of this is expensive. It is also a safety system, since a badly specified instruction inside a human organization usually passes three or four people who each think it looks odd before it does real damage.

Agents do not hesitate. They execute a misspecification at full speed and full fidelity, a thousand times over, before anyone notices. INSEAD organization design professor Phanish Puranam names the consequence directly, arguing that multi-agent AI systems are organizations. Human friction is a hidden buffer. Decades of effort have gone into engineering it out, with little recognition of what it was holding.

Of course, a good deal of what passes for judgment is unexamined habit. Experienced people are often slow for reasons that have nothing to do with wisdom, protecting a routine or a relationship rather than an outcome, and a well-instrumented system would improve it substantially. What survives that objection is the asymmetry of scale. A person acting on a bad habit gets it wrong one case at a time, while a misspecified system gets it wrong everywhere at once.

Four design moves

To respond to these challenges and accelerate the opportunity of high-performance Humans + AI organizations undertake these four simple design moves:

Map the actual organization before redesigning anything. Ask the people in the role what they notice that never reaches the data, what they care about beyond their job description, and what makes them slow down. Treat the distance between their answers and the documented process as your brief.

Make escalation a teaching loop. Route the hardest and least confident cases to humans whose decisions refine the system's policy, so that reviewers work as teachers rather than gatekeepers. Keep enough ordinary cases flowing past them to preserve their feel for what normal looks like.

Govern the system rather than the individual agents. Assign responsibility for what agents collectively produce, monitor the outcomes that no single agent owns, and watch for errors compounding at machine speed.

Protect the pipeline that produces judgment. Before automating a body of work, check whether that work is how people in your organization build senior judgment. If it is, build the replacement training ground first, or reconsider how you redesign work, or you will find in five years that you automated the apprenticeship and kept the job description.

I suggest you start this week with one decision that genuinely matters, and map how it is actually made rather than how it is documented. Design from the distance between the two. Your organization runs on a network nobody drew, and the parts of it no instrument can read are the parts worth protecting.

We are entering a new era where the most successful organizations will be those that effectively integrate the distinctive capabilities of humans and AI. I am working hard on researching and mapping out the potential and pathways and will be sharing more on the journey. Please let me know what you learn along the way!