When Judgment Becomes Infrastructure
How software turns management decisions into operating conditions

- By
- Cam Myers
- Content type
- Point of View
- Reading time
- 6 min read
- Status
- Author’s perspective
- Topics
- Operating Models
In brief
- Executable rules give management judgments a continuing role across transactions.
- Observed facts, model estimates, approved policy, and exception judgment should remain distinguishable.
- A technically functioning rule can become commercially wrong as its underlying assumptions change.
The argument
Imagine telling a manager to stop advertising an item once its price reaches an acceptable level. The instruction seems straightforward. You want to avoid paying for demand after the immediate commercial goal has been met.
Now put that instruction into software. The system reads prices, compares them with targets, and changes advertising eligibility. It can do this across many items without another conversation.
You have done more than save the manager a task. You have given a judgment a continuing role in how the business operates. What once depended on a person noticing, interpreting, and acting has become an operating condition.
That can be valuable. It can also preserve the wrong assumption with remarkable consistency.
The important question is which judgment should become a rule, under what conditions it should apply, and who remains able to change it. As businesses use software and AI to handle more decisions, those choices become part of management itself.
The decision underneathWhich judgment should become a rule—and who retains the ability to change it?
A rule changes who decides
In work with a regional auction business in the northeastern United States, I explored connecting inventory age, product costs, target sale prices, and advertising decisions. The purpose of the design was to make commercial information affect repeated actions instead of leaving it in a report.
Consider the advertising instruction above as an illustration of that design problem. Someone must decide what an acceptable price means. A target based on recovering purchase cost serves a different purpose from one based on expected contribution after other costs. Someone must also decide whether reaching the target justifies stopping promotion. There may still be worthwhile additional demand, obligations to a seller, or effects on related items.
These are commercial choices. A developer can implement them, but implementation does not settle them.
Once the rule is running, it changes the distribution of authority. The person who sets the target and the person who defines the rule influence transactions they may never see individually. The employee watching the campaign may have less discretion than before. If only a technical specialist can revise the logic, practical authority may move away from the commercial manager without anyone formally changing the organization chart.
Herbert Simon described organizational authority as extending beyond direct orders to the premises people use in making decisions. Software gives some of those premises a persistent, executable form.[1]
The benefit is that useful decisions can reach more of the business without requiring the same person to repeat them. An experienced owner no longer needs to remember every routine adjustment. Staff can act within clearer boundaries. The organization can preserve part of its operating knowledge when responsibilities change.
That benefit depends on making the logic available to the people accountable for its consequences. A manager should be able to understand why an item was excluded, what information was used, and which approved instruction produced the action. The explanation needs to be commercial enough to challenge, not merely technical enough to debug.
Keep estimates separate from policy
Automation becomes harder to manage when facts, estimates, and instructions are mixed together.
An item's recorded cost is information. Its expected selling price is an estimate. The minimum return management will accept is a policy choice. Whether an unusual case merits an exception may require judgment about circumstances outside the available data.
Those elements can work together, but they should remain distinguishable. Otherwise a change to a forecasting model can quietly alter the commercial policy. Or an old cost field can make a current decision appear more certain than it is.
Suppose an AI model estimates an item's likely sale price. A better model may improve that estimate. It does not decide whether the business should prioritize cash recovery, protect a seller relationship, or spend more to pursue additional demand. Those objectives must come from the business. Nor can a model reliably use a product's damaged condition if nobody has observed and recorded it.
A useful design therefore gives each element a defined job. The model estimates what may happen. Approved logic determines which actions are permitted under specified conditions. People handle decisions that require information, authority, or interpretation the system does not have.
This does not mean routing every transaction back to an owner. That would preserve the bottleneck inside a new interface. It means identifying where the rule is reliable enough to operate and where the business needs a different response.
Missing information, conflicting records, or an unfamiliar situation may justify holding an action rather than guessing. A large financial commitment may justify review even when the information is complete. The boundary should reflect the consequence of being wrong and the cost of waiting.
Exceptions also need a route back into the design. If staff override the same rule repeatedly for the same reason, the business has learned something. The policy may be incomplete, the data may be poor, or the supposed exception may be ordinary work the design overlooked.
Treating every override as employee noncompliance would destroy that source of learning. Treating every override as correct would make the rule meaningless. The pattern needs investigation.
Infrastructure must remain open to revision
A rule that made sense when inventory was scarce may be costly when stock accumulates. A threshold suitable for a growing market may be wrong after demand weakens. The software can continue functioning exactly as designed while the commercial decision becomes progressively worse.
This is a different failure from a broken integration. The system may be accurate, available, and faithfully applying an instruction that the business should no longer use.
Review therefore has to include the assumption behind the action. A meaningful review might examine items just above and below a threshold, the frequency of overrides, and the results of comparable decisions. It should ask whether the original reason for the rule still holds. The amount of scrutiny should match the exposure; a small, reversible adjustment does not need the same treatment as an expensive commitment.
Human involvement requires support, too. In her work on industrial automation, Lisanne Bainbridge explained how removing routine activity can leave people responsible for difficult situations while reducing the practice and understanding needed to handle them. Her argument concerned industrial control, but it gives business software designers a useful warning: assigning a person to intervene does not establish that the person will be prepared to do so.[2]
The employee handling an exception needs the relevant context, the reason for the automated action, and enough authority to respond. A button labeled override is only a small part of that capability.
For AI systems, the same continuing responsibility appears in NIST's 2023 risk-management framework: roles, monitoring, feedback, and mechanisms for changing or disengaging a system belong throughout its operating life.[3] These are management obligations as well as technical ones.
The opportunity is to free experienced people from repeatedly making decisions the business already understands, while preserving their ability to recognize when that understanding no longer applies.
When judgment becomes infrastructure, the business acquires a way to act without another conversation. It must also retain a way to question the conversation that started it.
Sources
- Herbert A. Simon. Organizations and Markets. Journal of Economic Perspectives, 1991, 5(2), 25–44.
- Lisanne Bainbridge. Ironies of Automation. Automatica, 1983, 19(6), 775–779.
- National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework 1.0, 2023. Govern 1.5, Govern 2, Measure 3, and Manage 4.1.
For the leadership team
In practice
- Keep commercial logic understandable to the managers accountable for its consequences.
- Treat recurring overrides as evidence to investigate, not automatically as either error or good judgment.
- Review the premise behind the rule and equip exception handlers with context and authority.
Questions worth asking
- Could a manager explain and challenge the instruction that produced this action?
- Which change in conditions would make this rule inappropriate?
Weighing this decision now? Tell us what you are considering and we will say plainly whether we can help.
Discuss your business

