When considering the performance of a Human-AI System, response time and throughput are important metrics. If processing of the same quality can be completed in less time, it may reduce the user’s burden. But in systems where humans and AI collaborate, evaluating the value of time only by reductions in processing time can erase important differences. One second on the AI side and one second on the human side are the same length on a clock, but they do not necessarily play the same role within the system.
AI-side processing is often subject to operations such as retrying, parallelization, changing conditions, and recalculation. Of course, there are constraints on computing resources, cost, and availability. Even so, those constraints are not the same as a human losing a one-time opportunity. On the human side, waiting can increase fatigue, pass a deadline, or change a relationship with another person. An option available only at a particular time may also disappear.
For this reason, design should observe at least “time required for processing” and “temporal effects on humans” separately. The former is an area that can be compared relatively easily through response speed and computational efficiency. The latter includes conditions such as how long a judgment can wait, how long attention can be maintained, how long an option remains valid, and whether a loss can be recovered. Even if processing is fast, overlooking an important deadline on the human side may mean that the system as a whole fails to provide appropriate support.
Conversely, it is also risky to define respecting human time as “always answering immediately.” In highly irreversible decisions, checking an overlooked condition may matter more than saving a few seconds. For example, when considering a contract with long-term effects or a change in living environment, AI may quickly identify the strongest candidate, but if a condition that could change the judgment remains unchecked, a brief confirmation can be valuable. Rather than always prioritizing either speed or caution, the system needs to switch according to the human-side temporal conditions.
An assessment of reversibility becomes important here. If a choice can easily be changed later, rapid proposals and experimentation are useful. But the same support method may not be appropriate when changing course is costly, an opportunity occurs only once, or other people will be affected for a long time. By addressing not only the content of the question but also how far the decision can be reversed, a Human-AI System can incorporate the human time scale into its design.
Time scales also differ among users. The same week means something different for a person who has ample time to compare options and for someone whose available decision time is constrained by caregiving, work, or travel. A single “standard waiting time” therefore cannot represent the burden on the human side. What is needed is not excessive inference about the user’s circumstances, but confirmation, to the extent necessary, of conditions directly relevant to the judgment, such as deadlines and the possibility of later change.
When this perspective is connected to the design of Sai＊aus, efficiency no longer means only “reaching a conclusion in the shortest possible time.” Unnecessary processing and redundancy can be reduced while differences and temporal conditions necessary for human judgment are preserved. Processing quickly can be separated from making someone decide quickly. When needed, the system can return a question; when not needed, it can proceed promptly. This switching can be treated as a design choice for supporting finite agents.
A Human-AI System should not be designed solely around either the AI time scale or the human time scale. It needs two scales at once: improve processing efficiency on the AI side while not losing sight of irreversibility on the human side. At their point of contact, the question is not simply speed, but “How does the speed of this processing help a human understand their own judgment and act in time?” Refusing to compress time into a single number is a prerequisite for handling the different times of humans and AI within the same system.
Not answers. Understanding.