The Hierarchy Before AI: Expertise Above Execution
To understand what GenAI changes, start with the organization it enters: a knowledge-based hierarchy. This chapter focuses on the pre-GenAI division between workers who handle ordinary problems and experts who handle exceptions, and it gives the reader a way to judge how skill requirements and managerial capacity are connected without treating hierarchy as merely a chain of command.
The basic arrangement appears in many forms. An analyst resolves routine cases and asks a manager about an unusual one. A junior attorney handles a matter within their experience and escalates a difficult question to a partner. A software engineer works within familiar patterns and turns to a tech lead when a problem exceeds their knowledge. The titles differ, but the organizational logic is similar: not everyone needs to know everything if the firm can route unusual problems to someone who does.
That logic is the baseline for interpreting GenAI. If a new system changes the number of routine tasks, the cost of asking for help, or the knowledge required to check a decision, it changes the structure that allocates expertise. The important question is therefore not whether hierarchy exists, but what work each layer performs and what makes the arrangement economical.
A worker handles the ordinary; an expert handles the exception
Imagine that a firm receives one problem at a time and that problems vary in difficulty. A worker has a knowledge boundary: they can solve problems up to a certain level without help. Problems below that boundary are handled directly. Problems above it are not necessarily lost. The worker asks an expert for guidance, receives a solution, and completes the output. The expert’s role is to make the worker effective on the exceptional case, not to take over every piece of routine production.
This distinction between consultation and production matters. In the model underlying this explanation, the worker remains the sole producer of output even when an expert supplies the missing knowledge. The firm has not created two interchangeable layers of task executors. It has created a division of labor in which one layer is broad and numerous, while another is specialized and used when ordinary knowledge is insufficient.
The arrangement economizes on knowledge acquisition. Requiring every worker to master the entire problem space would make each worker more expensive and may waste expertise on routine cases. Requiring workers to know very little would make the expert layer absorb too many consultations. A hierarchy is useful in the space between those extremes: workers carry enough knowledge to resolve a valuable share of cases independently, and experts concentrate on the residual cases where their breadth is worth the cost.
The worker’s knowledge boundary therefore has two effects. It determines how much the worker can do alone, and it determines how often the worker must call upward. Raising the boundary reduces referrals. Lowering it increases them. Neither movement is automatically good or bad. A higher boundary costs more in wages or preparation, but it may save expert time. A lower boundary saves on the worker side, but it creates a larger load for the expert side.
The expert layer is not simply “management” in the broad everyday sense. In the formal model, senior experts structurally stand for functions such as exception handling, quality assurance, and oversight. Real managers also motivate people, allocate resources, resolve conflicts, and represent the organization. Those activities are outside this simplified structure. The model’s narrower focus is useful precisely because it isolates one question: how does an organization distribute problem-solving knowledge when cases differ in difficulty?
That question also explains why the hierarchy can be understood as a capacity system. Suppose several workers encounter difficult problems at the same time. Each consultation consumes some of an expert’s time. If there are too few experts, requests queue or receive rushed attention. If there are too many, the firm pays for capacity it does not use. The number of experts is thus related to the frequency of referrals and the time required to answer them, not simply to the number of people who carry a senior title.
Consider two firms facing the same stream of problems. The first hires workers with broad knowledge and routes only unusual cases upward. The second hires workers with narrower knowledge and relies on a larger expert layer. Neither design is universally superior. The first pays more at the frontline but may give workers greater autonomy and reduce delays caused by consultation. The second may lower the cost of entry into the role, but it makes expert availability and communication quality more important. If the cost of contacting an expert falls, the second arrangement becomes more attractive. If the premium for worker knowledge falls, the first arrangement becomes easier to sustain. A change in the surrounding conditions can therefore alter the appropriate shape without changing the firm’s underlying goal of producing correct work.
The model’s abstraction helps isolate this relationship. It treats problem difficulty as varying across a common range and assumes that the expert layer can solve the full range. Those assumptions remove many complications of actual workplaces—different specialties, incomplete information about who knows what, organizational politics, and managers’ nontechnical responsibilities. They do not erase the central dependency. Whenever a frontline role has a boundary and difficult cases cross it, the organization must pay either to move that boundary outward or to maintain a channel to someone who can handle what lies beyond it.
The same reasoning applies to the worker layer. If the firm needs one unit of output and each worker can process one problem per unit of time, the number of workers reflects the amount of work that remains at the frontline. In the pre-GenAI benchmark, one worker and a number of experts are enough to represent this relationship. The model is deliberately spare, but its central dependency is realistic: the frontline and expert layers are coupled through the problems that cross the worker’s knowledge boundary.
The diagram makes an easily missed point visible: escalation does not automatically mean that work leaves the worker layer. In this arrangement, the expert supplies knowledge and the worker turns that knowledge into the completed result. What changes when the worker becomes less knowledgeable is therefore not only the worker’s wage. The frequency and cost of interaction with the expert also rise.
The price of knowledge sets the shape of the hierarchy
The firm’s design problem can be expressed as a trade-off between two costs. The first is the cost of hiring or developing a worker with a larger knowledge range. In the formal model, the knowledge premium rises convexly: each additional increase in knowledge is more expensive than the last. The second is the cost of relying on expert consultation. Every referral consumes communication time and expert labor. The firm chooses a worker knowledge threshold that balances those costs.
This is why there is no universally optimal level of frontline expertise. If knowledge is inexpensive, the firm can hire more knowledgeable workers and reduce referrals. If expert communication is quick and inexpensive, it can tolerate a lower worker threshold and use the expert layer more heavily. If expert time is scarce or communication is costly, raising frontline capability becomes more attractive. Organizational structure responds to the prices and constraints surrounding knowledge, not to a fixed preference for tall or flat organizations.
A useful way to describe the first outcome is the worker knowledge requirement: the minimum level of knowledge the firm chooses for a frontline role. It is a statement about workforce composition and hiring or development standards. A second outcome is the span of control, defined here as the ratio of workers to experts. A larger span means that each expert supports more workers on average; a smaller span means that expert capacity is distributed across a narrower frontline base.
These measures are related but not interchangeable. The firm may hire more knowledgeable workers and still need many experts if the remaining problems are difficult or costly to communicate. It may also lower the worker threshold while retaining a broad span if consultations are unusually efficient. Looking only at average skill or only at the number of managers can therefore conceal the mechanism. A structural change should be described through at least two questions: how much knowledge is required at the frontline, and how much expert capacity is needed to support it?
The distinction becomes especially important when a change affects both sides at once. Suppose a new system lets a worker handle more routine cases but makes the difficult remainder more consequential. The firm might need fewer worker-hours overall, yet it might still need experts who can resolve the cases that now carry greater risk. Alternatively, if workers become able to handle a wider range independently, the expert layer may shrink even if the worker knowledge threshold stays constant. The direction of the headcount change does not reveal which capability moved. One must examine the boundary of independent work, the volume of referrals, and the time required for each referral.
The threshold is also an organizational choice, not a complete description of any individual. A worker hired at a given knowledge level may have tacit experience, judgment, or learning potential that the stylized model does not measure. Conversely, a formal qualification may not guarantee that the worker can solve the particular problems arriving in the workflow. The model abstracts from those differences to show why the firm’s chosen threshold and the expert layer’s capacity are jointly determined. In practice, the relevant evidence would include the kinds of cases workers resolve unaided, the kinds they escalate, how long expert responses take, and whether the resulting work is accurate—not simply the average credential level.
This perspective also prevents a false choice between a “flat” and a “tall” organization. The firm can change the worker boundary without changing the number of layers, change expert capacity without changing job titles, or alter communication without changing either formal layer. A technology may first make one margin more valuable and later make another more valuable. The baseline does not predict which change GenAI will cause; it identifies the connections that any such prediction must account for.
The ratio also needs careful interpretation. A wider span does not necessarily mean that managers have become less important. It may mean that experts can handle more consultations, that workers can solve more problems independently, or that the organization has shifted work away from the expert layer. Conversely, a narrower span does not by itself signal bureaucratic excess. It may reflect a period when expert judgment is unusually valuable or when the frontline needs more support. The metric captures a capacity relationship, not the full quality of leadership, employee autonomy, or workplace culture.
This distinction helps separate employer incentives from employee experience. A firm may prefer a lower worker knowledge threshold because it reduces wage costs, while workers may value deeper expertise because it improves autonomy, bargaining power, and resilience when the technology or workflow changes. A lower threshold can widen access to a role, but it can also narrow opportunities to build the expertise that makes someone independent of the support system. A higher threshold can make entry more selective, yet it can also provide workers with more control over difficult cases and a stronger basis for detecting mistakes.
The benchmark also clarifies what a technology must do before it can change hierarchy. It could reduce communication costs, allowing a smaller expert layer to support more workers. It could increase the capability of the workers, allowing them to resolve more cases without escalation. It could change the cost of acquiring knowledge, making a higher worker threshold economical. Or it could perform some tasks directly while leaving the organization to decide who checks the result. These are different mechanisms, even if all of them are described casually as “AI improving productivity.”
There is a second reason to keep the baseline in view: hierarchy can absorb a shock through more than one margin. When a task changes, the organization can adjust the number of frontline workers, the knowledge expected of each worker, the number of experts, or the communication process between layers. A headline such as “AI reduces junior work” says little about which margin moved. The firm may need fewer people to execute routine tasks but more expertise per remaining worker. It may lower the entry requirement and increase referrals. It may preserve the expert layer because checking and correcting new forms of work become more important.
In other words, hierarchy is not a solid block that technology either leaves intact or knocks down. It is a set of linked capacities. The worker layer determines how much work can be completed without help. The expert layer determines how much exceptional work can be supported. Communication connects the layers, and the cost of that connection influences where the firm places its knowledge threshold. A change in one part can create pressure in another.
That is the baseline needed for the next question. GenAI is unusual not only because it can perform cognitive work, but because it can produce an answer that sounds finished while still being wrong. If an organization routes such output into production, someone must decide whether it is correct, repair it when it is not, and escalate it when the current layer cannot judge it. The question is no longer just how much knowledge workers need to execute a task. It is who needs enough knowledge to validate what the machine has done.