In a monthly planning meeting, the number discussed most often is service level, and the number discussed least often is the share of system-generated order proposals that were rewritten by hand. An automated replenishment or production planning engine produces thousands of output lines overnight; part of that output flows through on straight approval, while another part is edited the following morning in a spreadsheet opened alongside the planning screen. The proportion of edited lines appears in almost no management report, because the system does not record the rejection of its own recommendation as an error — it absorbs the user's correction as an ordinary transaction. What surfaces in the meeting is whether the month's service level held, and for as long as it holds, the question of how much of the system is actually running is never posed.
A second pattern runs quieter than the first: the exceptions tend to be resolved by the same individual. Somewhere in the organization sits a planner who knows the delivery behavior of a particular supplier, the ordering rhythm of a particular customer, and the typical customs delay attaching to a particular product group, and who each month expresses that knowledge not as a parameter change but as a manual correction. The service-level deterioration observed while that person is on leave, or in the quarter following a resignation, is better read not as random variation but as a measurement of dependency. What is commonly discussed as founder dependence has an exact operational analogue here; the only difference is that the dependency has been concealed behind a system.
The name of the pattern is automation brittleness — the inability of an automated system to generate adaptation once conditions move outside the distribution against which it was calibrated. The defining property of the mechanism is that the break is silent rather than noisy: an experienced planner encountering an unfamiliar situation knows that it is unfamiliar and asks, whereas a rule engine does not report ignorance, forces the input into the scope of the nearest defined rule, and presents the result in the same format, and with the same apparent confidence, as every other line it produced that night. To the extent that correct and incorrect outputs are formally indistinguishable, the error signal is lost, resurfacing weeks later as a symptom on the inventory side or at the customer interface.
A second layer of brittleness is skill erosion. As automation absorbs routine judgment, the reasoning capacity expected to engage at the moment of exception goes untrained; a competence exercised nowhere while the system runs cannot reasonably be assumed available the moment the system stops. The third layer is the loosening of input discipline. Safety-stock coefficients, lead-time assumptions, minimum order quantities, lot sizes and supplier calibration tables are entered once during configuration, and for as long as output remains satisfactory, no one asks which period's operating conditions those parameters were derived from. Parameter aging is the most common and the least expensive form of automation brittleness to correct; because ownership is rarely assigned, it is also, typically, the last to be noticed.
Characterizing this behavior as a defect would be misleading. Where the input distribution is stable, automation lowers transaction cost by an order of magnitude, standardizes the decision and reduces person-dependence; under the conditions in which it was installed, it is entirely rational. The problem lies not in the shortcut itself but in the shortcut persisting after the conditions change — when the supplier base shifts, when the product mix widens, when transport mode or customs regime is altered, the engine continues applying rules calibrated to the old distribution against the new one. Brittleness is less a flaw in automation than a delayed measurement of the gap opening between calibration and reality.
That gap seldom appears in the total inventory line on the balance sheet; the total holds while the composition degrades. When dead or slow-moving stock and shortages in critical items expand at the same time, average inventory turns remain within a reasonable band and generate no warning at the level of the financial statements. The remainder of the cost disperses into overhead: expedited freight, unplanned overtime, higher cost per shipment driven by partial deliveries, replanning and rework hours, and the commercial concessions extended to customers for late delivery. Because none of these items posts to an account bearing the name of the problem, its magnitude becomes visible only when they are deliberately assembled and read together.
At the review table the same phenomenon returns in a considerably sharper form. A buyer or a lender assessing whether the planning process is systematic asks not whether a system exists but how much of its output is corrected by hand; where the override rate is high, the conclusion drawn is that the process is person-dependent in practice, however thoroughly it has been documented on paper. That finding opens the question of whether service-level-driven costs are recurring or one-off in the EBITDA normalization, moves key-person retention onto the negotiation agenda, and is typically priced through earn-out structure, escrow ratio, or the scope of operational representations and warranties. What determines valuation here is not the performance itself, but whether that performance can be shown to be reproducible independently of a particular individual.
A third institutional surface of brittleness lies in the external interfaces on which the automation depends. Where the planning, ordering or shipment flow is tied to a supplier's data format, a carrier's tracking interface, a customs broker's system, or the release policy of a single software vendor, that dependency generates no cost in normal operation while converting directly into bargaining asymmetry in the contract renewal window. The counterparty, knowing that switching cost and integration rebuild time run in its favor, shows correspondingly less flexibility on price and terms — a technical dependency translated into commercial leverage. Maintaining no integration inventory alongside the contract renewal calendar amounts to not knowing on whose side that leverage sits.
The mechanism that neutralizes this tendency is not vigilance or training but a four-component system design. The first is override telemetry: how many of the system's recommendations were changed, in which product group, for which supplier and on what stated ground, captured at the moment of correction and tracked as a management indicator, on the understanding that a rising override rate signals not a working system but one whose calibration has drifted. The second is parameter ownership: every critical parameter carries a named owner and a defined review interval, with the review running on the calendar rather than on the system's apparent performance. The third is degraded-mode operation: a written specification of which decision is taken by whom, on what data and against which approval threshold when automation is unavailable, rehearsed at regular intervals. The fourth is the decision boundary — an explicit separation of decisions delegated wholly to the system, those requiring human approval, and those the system should not take at all.
BEIREK's intervention on this line is not to propose a new system but to make measurable the conditions under which the existing one actually works. The first record established when taking over an operation is the override log; once the volume, distribution and stated rationale of corrections have been observed across several planning cycles, the product groups, suppliers and seasonal conditions in which brittleness concentrates emerge without requiring estimation. In parallel, a parameter ownership matrix is constructed — each assumption's owner, last update date, underlying data window and review interval consolidated in a single table — because the cost of parameter aging grows in direct proportion to the ambiguity of ownership.
The operating rhythm is established at three scales: quarterly review of parameters, annual rehearsal of degraded-mode operation against a realistic interruption scenario, and yearly refresh of the external integration inventory mapped against the contract renewal calendar. In capital-intensive projects this work is pulled forward ahead of commissioning, since decision boundaries and exception procedures defined after a facility has begun operating will invariably ratify whatever habit has already formed, whereas the same definitions established before commissioning shape that habit instead. This distinction is among the small number of variables that determine how quickly a given investment reaches operational maturity.
The value of automation lies not in its avoiding error but in its capacity to report error in time; any system that produces no such report accumulates unmeasured risk under the appearance of efficiency. The maturity of an operation is measured not by how many processes have been automated, but by whether the question of what happens when the automation stops has a written, tested and owned answer.
One question remains: is the share of planning-system output currently corrected by hand recorded anywhere within the organization today?