Questions

Why this boundary?

Delegate what you can check and undo.

Haier, Buurtzorg and Zappos show different ways to hand authority to small teams. What makes delegation work, and what carries over to agents?

Kind

Hypothesis

Status

Working

Published

17 October 2026

Reading time

About 3 min

The position above, in one picture: where work leaves the company and where a person signs it out.

Three experiments in distributing authority

Before software agents, firms tried to push decisions down to small units. Three cases are especially well documented.

Haier reorganised a large appliance maker into many microenterprises. They are coordinated through internal markets, shared platforms and contracts. The model has proved operationally viable at scale. It has not been shown to cause Haier’s results. Its core difficulty is that autonomous units sit in an interdependent value chain, and each can optimise locally at the expense of the whole.

Buurtzorg runs Dutch home care through teams of about ten to twelve nurses with no line managers, supported by coaches, a small back office and shared IT. Our interpretation: part of why it works may be that each team delivers a nearly complete service, so most of the value is created inside the unit. Even so, the cost evidence is mixed. Direct costs were low, but a later evaluation found total cost per client roughly average once wider healthcare costs were counted.

Zappos adopted Holacracy, which separates roles from people and defines each role by its purpose, accountabilities and domains. After a severance offer in 2015, about 14% of staff left within a month. The lesson that held up is that removing managers does not remove the work managers did: allocating resources, resolving priorities and developing people. The formal system created a large coordination load of its own.

What the cases share

Our interpretation, building on Malone: decentralisation buys responsiveness and local knowledge at the cost of consistency and cross-unit coordination. The deciding variable is how much coordination has to happen outside the unit that creates the value. Buurtzorg’s teams need little; Haier’s microenterprises need a lot.

Applying it to agents

Our interpretation: an agent can be treated as an extreme case of an autonomous unit. It can execute quickly at low marginal cost. It also carries a specific risk: it may satisfy a stated rule while missing the purpose behind it. The technical evidence is narrower than the organisational question. MetaGPT studies software-development tasks and reports that assigning agents defined roles and standard operating procedures reduced cascading errors on its benchmarks. TheAgentCompany measures agents on tasks in a simulated company and reports a gap between completing individual tasks and handling organisational work reliably.

Hypothesis: the boundary question for agents has the same shape as for nursing teams.

• Delegate where the unit of work is nearly self-contained, its output can be checked, and a mistake is recoverable.
• Retain or gate where the work depends on many other units, quality only shows up later, or a mistake creates a commitment the firm cannot undo.

How authority moves over time

Several 2026 practitioner frameworks describe graduated autonomy: an agent moves from shadowing a human, to working under supervision, to working within guidance, to full autonomy in a bounded domain. BCG’s “harness” adds the scaffolding that makes each step auditable: a specification of scope, explicit rules and escalation triggers, an audit trail, and quality gates with a named human signatory.

These are proposals from consultancies, not measured results. They do give a testable vocabulary. For any delegated activity we can ask what stage it is at, what evidence moved it there, and what would move it back.

Agrawal, Gans and Goldfarb add a distinction. Dropping AI into an unchanged process is a point solution. Point solutions can be useful, but the authors argue that most of the additional value comes from system solutions that redesign the decisions around the technology. Redesign also shifts power between whoever owned those decisions before and whoever owns them after.

Our working position

Hypothesis, not finding. The line between delegated and retained work follows verifiability and reversibility, not job function. “The finance agent” or “the marketing agent” says little. The useful description is what the system may decide, below what threshold, with which escalation path, and how anyone would notice it drifting.

What would change our mind

• Agents reliably handling tightly coupled, hard-to-verify work without a human gate, over months rather than demos.
• Evidence that the cost of the harness (specification, monitoring, exception handling) grows faster than the value delegated, which would make the boundary an economic limit rather than a design choice.

Open questions

• Is there an agent equivalent of Buurtzorg’s team: a unit of work complete enough to run end to end with light coordination?
• Rule violations can be monitored. Intent drift, a sequence of individually permitted actions that adds up to something nobody wanted, is much harder to catch. Who is responsible for noticing it?