Six Organizational Design Principles When Production Is Cheap

For more than a century, organizations were designed around a single constraint: human labor. Hierarchy, layers and process discipline existed to coordinate large groups of people doing cognitive work at scale. Almost every organizational design principle we treat as natural — the span of six to eight direct reports, the shape of the graduate intake pyramid — solves that one problem.

That assumption is breaking. Generative and agentic AI change the economics of work: the cost of producing analysis, content, software and draft decisions is falling towards zero. As AI absorbs coordination and routine execution, the bottleneck moves from labor capacity to judgment, verification and accountability. That shift requires a redesign of the organization itself.

We published a short note on this recently, and the volume of requests to expand it told me how live the question is. What follows is the longer argument: six organizational design principles, with the evidence for and against each. They are hypotheses to test, not prescriptions.

Managers must become producers again

In the legacy model, promotion to manager is a one-way exit from craft. The manager coordinates and reviews; producing the work directly is treated as a poor use of expensive time. That logic holds only when output is constrained by human labor. With modern tooling, a senior practitioner can produce in a day what recently took a small team weeks. Many managers now create more value by doing than by supervising the doing.

The practical moves: redefine the manager role as 50 to 70 percent individual contribution, remove the pay and status penalty for managers who keep producing, and build dual ladders that stop treating the IC track as a consolation prize. The answer to who then develops talent is to separate real leadership from administrative overhead. Hiring, calibration, difficult feedback and conflict resolution stay human. Status reporting and workflow coordination increasingly do not.

The entry tier changes shape rather than disappearing

Entry-level roles did two jobs: absorb high-volume, lower-judgment work, and develop the next generation of senior talent. AI disrupts both. The temptation, already visible, is simply to hire fewer graduates. That is a strategic error — it liquidates the pipeline for a short-term saving and concentrates institutional knowledge in a shrinking senior cohort.

The durable response reimagines the tier: new hires who evaluate, stress-test and red-team AI output instead of drafting it, and junior work judged on the quality of problems framed rather than artifacts shipped. The risk of getting this wrong is visible in the Big Four, where graduate intake cuts of 6 to 33 percent across KPMG, Deloitte, EY and PwC between 2023 and 2025 removed the work that produced senior professionals. Median tenure to partner track is about seven years, so the analysts who would have been senior managers in 2028 were never hired. Cuts may be structurally correct. The timing and shape of the rebuild matter as much as the cut.

Layers lose their economic justification

Hierarchy has always carried a cost: review cycles, translation loss, slower decisions. We accepted it because managers moved information upward and decisions downward. When summarization and routing become inexpensive, layers that existed mainly to do that lose their justification. Decision rights have to migrate downward, because compression without real delegation just bottlenecks decisions at the top.

The right number of layers is not a constant. Risk, compliance and safety-critical engineering may justify more. Ask, layer by layer: what does this level decide that no other level can? If there is no clean answer, it is a candidate for compression. Bayer's 2024 move to Dynamic Shared Ownership is the most ambitious live test at Fortune 500 scale — half of all management positions eliminated, twelve or thirteen layers cut to six or seven, about 2,000 autonomous ten-person teams across 100,000 employees. The early results showed product development time-to-launch reductions of up to 70 percent, alongside real costs from self-organizational ambiguity.

Verification becomes a function, not a habit

The least discussed consequence of AI in the enterprise is the inversion of the cost curve between production and verification. Producing a plausible draft is cheap; deciding whether it is correct, complete and safe is expensive. Few organizations have anyone explicitly responsible for that decision.

Verification deserves to be a first-class layer, not a side-task of QA. Its mandate: define what good looks like for AI-assisted output in each workflow, build evaluation harnesses and review protocols, calibrate accuracy continuously, and own the escalation path when verification fails. Expect to write a job family that does not exist in your architecture today. The cost of its absence is already documented. Air Canada was held liable in a 2024 tribunal after its chatbot gave incorrect bereavement-fare guidance; the tribunal rejected the argument that the tool operated autonomously. Klarna rehired human staff after replacing 700 service agents with AI degraded quality in complex, emotionally sensitive interactions.

Small cells become the default operating unit

Amazon's two-pizza team was a constraint dressed as a culture statement: keep teams small enough that communication overhead does not eat the productivity gain. AI sharpens that logic, because each additional human adds coordination cost while each additional AI capability typically does not. Expect much of the work now done by 8 to 12 person teams to migrate to 2 to 5 person cells that own outcomes end to end. Smaller is not universally better: incident response and safety operations draw their reliability from redundancy.

Cursor is 18 months of evidence — a billion dollars in ARR and a 29.3 billion dollar valuation on roughly 300 people, with agents creating over 30 percent of internal pull requests. Compelling, but not longitudinal. For that, look at Haier, which has run RenDanHeYi since 2012: 80,000 employees reorganized into around 4,000 self-managing micro-enterprises, each with its own P&L. One thing there needed constant reinvention — the coordination layer between teams, which Haier rebuilt as explicit ecosystem contracts. Removing layers does not remove the need for coordination. It makes that need visible, and the vacuum fills itself poorly if you leave it.

Designing for a workforce that is not entirely human

The sixth principle binds the other five. Enterprises are no longer managing only people; they are managing combinations of humans and agents. The agent workforce is elastic and constantly changing, so it needs its own governance. For every agent in production: which human is accountable for its outcomes, under what conditions is it retrained or retired, and what is the escalation path when it fails? At function level the planning metric becomes workforce composition — agent-heavy for tier-one service, human-heavy for high-stakes credit or clinical decisions, mixed for engineering. Verification is the natural owner of those calls. Without a named owner they default to whoever built the agent, the condition behind both Air Canada and early Klarna.

Span of control then gives way to span of supervision: human reports plus agents under a leader's accountability. Frameworks that reward output volume misfire once volume is no longer scarce. The job catalog also gains archetypes it does not have — the producing manager, the verifier, the cell lead and the agent owner.

Where these organizational design principles land first

Adopting any one principle alone will disappoint. Together they reshape the firm into a network of small accountable cells, supervised by senior practitioners who still produce. Start in one business unit where adoption is underway and leadership is willing. In the first 90 days, map span of control, layer count and team sizes against these principles. Over the next two quarters, pilot one cell of three to five people on a real outcome with explicit AI augmentation and a defined verification protocol, measuring cycle time, quality and engagement. Then refactor job families and promotion criteria against that evidence. Resist scaling early; the second pilot teaches more than the first repeated ten times.

Throughout, stand up an agent register: every agent in production listed with a named owner, an objective, metrics and a decommission criterion. And watch the pipeline health ratio — your entry-to-senior ratio and projected senior cohort in five to seven years — more closely than anything else here. It is the one indicator where the damage compounds silently and is expensive to reverse once visible. Hold the principles firmly and the implementations loosely. Organizational design changes are slow to take effect and costly to undo, and AI capability is still moving under us.