The Limits and Possibilities of AI Demand Forecasting as Seen from the Local Frontlines

AI does not know the 'morning at the produce market'

4:00 AM. The lights in the auction area turn on.

The voices of wholesalers overlap, and a unique tension runs through the floor. In that atmosphere, a veteran buyer intuits, 'Radishes will move today.'

If asked for the basis, the answer is this:

'The way it rained yesterday, the movement of temperatures this week, and—since next Monday is the beginning of the month, orders for facilities will increase.'

Much of this information is not in the data that AI has learned.

I have watched the frontlines in this industry for nearly 20 years. In a world of fresh food wholesale and retail, where the 'gap between freshness and demand' is a matter of life and death. And in the last few years, AI-based demand forecasting tools have quietly permeated the industry. I am writing this record as someone who has observed this change from the frontlines.

What AI is good at—the accuracy of reading past 'waves'

To be honest, AI demand forecasting is 'amazing'.

When you have it learn structural data such as weekly and monthly sales data, seasonal fluctuations, correlations with temperature, and purchasing patterns by day of the week, there are times when the average prediction accuracy exceeds human rules of thumb.

It is particularly effective for ordering management of staple items. Eggs, milk, tofu. Items that move stably every week are a perfect match for AI. It can accurately read patterns such as increased demand on weekends, decreased customer traffic on rainy days, and last-minute shopping before long weekends, thereby reducing waste loss.

In this area, AI will certainly surpass humans. I admit that.

However, when you are on the frontlines, there are definitely moments that AI cannot read.

What AI cannot see—the blind spot called 'context'

The 'AI blind spots' I have observed on the frontlines can be organized into seven major categories. All of these are cases that are actually happening in the industry.

1. Local events and festivals

At one supermarket that pilot-tested AI demand forecasting for ordering seasonal hot pot ingredients, the system recommended 'average' order volumes during the week of a large local festival. The AI did not have event data, so front-line staff had to manually revise the figures upward. The same applies to AI vending machine demand forecasting by beverage manufacturers; there is a structural limitation where one-off events, such as 'a specific local festival held every year at the vending machine location,' cannot be incorporated unless parameters are set manually.

Large festivals and local events change in scale, venue, and attendance depending on the year. In the case of new events or changes in scale, there is no historical data for the model to reference.

2. Elections and political events

In the Economy Watchers Survey, qualitative insights are recorded as testimony from taxi drivers that 'there is a tendency for foot traffic in nightlife districts to decrease in months when elections are held.' Front-line managers intuitively incorporate political and event calendars into demand, but this is not conveyed to AI systems. This is because elections are held at different times and with different issues every 4 to 5 years, and the impact on consumer psychology is different each time.

③ Factory and Office Operation Schedules

In factory-based towns, the operating schedules of nearby large facilities are directly linked to local consumption patterns. During a week when a major manufacturing plant set a collective paid leave, local supermarkets saw significantly higher sales of evening side dishes and bento boxes than in previous years. Because the AI's ordering recommendations were based on a 'normal Thursday,' frequent stockouts occurred. The factory operating calendar is private information and is a purely local variable that cannot be reflected in AI.

④ School Events and Enrollment/Admission Seasons

During long holidays when school lunches are not provided, demand for lunch at home increases. One supermarket attempted to incorporate local school events as parameters, but unstructured information such as 'this year, the sports festival for XX Junior High will be held at the nearby park' was not reflected in the model, and staff had to make manual adjustments each time. School events do not always occur at the same time every year, and manual work is essential to integrate the event calendars of multiple schools for demand estimation.

⑤ Food Safety Incidents and Recall Reports

During the early stages of the COVID-19 pandemic, there were reports of cases where AI forecasting models completely malfunctioned, triggered by rumors on social media that 'toilet paper would run out.' Because the AI had learned from previous 'normal' purchasing patterns, it could not detect this abnormal demand spike in advance. COVID-19 was so effective at rendering existing forecasting models useless that it was called 'AI's kryptonite.'

⑥ The 'Context' of Weather

While weather data itself is one of the external variables that AI handles best, the problem of context remains: 'even with the same temperature and precipitation, demand differs depending on the situation.' Demand fluctuations in a complex context like 'rainy day + weekend + distribution of sale flyers' are difficult to incorporate into a model. In these days of volatile weather, while it can incorporate 'this summer is 3°C hotter than usual,' the model has no precedent to refer to for a context like 'the first summer festival held during a heatwave.'

⑦ Physical Changes such as Road Construction and Neighborhood Development

Cases are frequently reported on the front lines where the number of customers drops significantly due to poor access during weeks when the main road in front of a store is closed for construction. Although this information is recorded in government public notices, it is almost never linked in real-time as input for demand forecasting AI. When a new housing complex is completed, the customer base and demand volume of nearby stores change, but this is not reflected in official data until resident registration is complete, and the AI continues to output incorrect values based on the old trade area population.

In quantum mechanical terms—'unobserved information' is as good as non-existent.

Let me add a perspective that is uniquely mine here.

In the world of quantum mechanics, there is a principle that 'the state of a particle is not determined until it is observed.' When we overlay this onto AI demand forecasting, the structure becomes clear.

AI can only handle 'observed data.'

Recorded things, quantified things, structured things—it sees the world only within that scope.

Local festivals, the atmosphere of an election, factory holiday calendars, school event schedules, news of food safety incidents, road construction—all of these influence the real market, yet are treated as 'non-existent variables' by AI. Information that is not observed does not exist in the AI's universe.

And these 'unobserved contexts' are held as tacit knowledge by veteran staff on the front lines.

'Side dishes sell well the week before that factory goes on summer vacation.' 'Customer traffic drops the week after an election is officially announced.'

This knowledge is not entered into the database. It is a type of information that AI cannot possess.

Technically, this is called 'concept drift.' It is a problem where the 'normal patterns' learned by a model become rapidly obsolete due to changes in context. COVID-19 was an extreme example of this, but on the local front lines, small-scale 'context changes' occur daily, and each time, the model continues to output incorrect forecasts based on past patterns.

Hybrid design is the realistic solution

Given these structural limitations, hybrid operation is attracting attention as a practical response.

Paltac's AI automated ordering service explicitly adopts a hybrid operation that uses AI automated ordering and conventional ordering for different categories, noting that "AI is strong at stable demand fluctuations for daily consumer goods, but is not good at products that are easily affected by external factors." The AI ordering system at Halo-Day also adopts a method where "ordering tasks are not completely automated, but the starting point for judgment is calculated by the AI system," designing it so that human staff confirm the AI's recommended values and make the final decision, thereby taking on the role of contextual correction.

I believe this direction is correct.

AI handles "accuracy during normal times," while humans handle "detection of discontinuous changes."

This division of labor is the current optimal solution.

And what is interesting is that when this "context" is successfully verbalized and given to the AI, the AI's prediction accuracy increases. Just by inputting, "There is a big event in the region next week. Higher attendance is expected than in previous years," the AI's recommended quantity becomes closer to reality.

Humans "observe the context" and pass it to the AI. The AI processes the data and improves accuracy. This form of collaboration is quietly beginning in the field of fresh food distribution as well.

However, one point requires caution. If the order quantity presented by the AI "exceeds the processing capacity of the site, and staff are forced to manually correct it," the perception that "AI is useless" will spread, hindering its adoption. Organizational design that correctly understands and utilizes the limitations of tools determines the success or failure of implementation.

Knowing the limits is the gateway to possibilities

Managers who think of AI as an "all-knowing prediction machine" will inevitably be betrayed. Conversely, sites that dismiss AI as "useless" will lose their competitiveness.

What is truly necessary is to keep asking the question, "Am I seeing things that the AI is not seeing?"

AI processes observed data precisely. Humans perceive reality that has not yet been observed. Design roles with an understanding of this asymmetry.

I believe that is the "power to see as a structure" needed to survive in local operations.

A must-read book for practitioners involved in demand forecasting and the
director-level management of such roles.
A practical book on demand forecasting that covers both theory and practice.
Includes the AI prediction that won the Special Prize at the 38th Logistics Awards! (Explanatory version, Chapter 11)
Includes a paper published in the Journal of Business Forecasting, the world's highest peak in demand forecasting!
(Japanese explanatory version, Chapters 13-14)

[Please leave comments and impressions in the comment section]

Thank you for reading!

是非noteを読んで、
・気づいたこと
・感銘を受けたこと、
・実践したいと思ったこと
を自分でメモにまとめるたり、
生成AIで壁打ちして振り返ると
より深い学びになります！
 
また、読んだ感想を一言でもいただけると、
励みになります！是非コメント欄でおしえてくださいね！

Z Observer | Observing the universe, AI, management, and life through the gateway of quantum mechanics. A record of a human continuing to hone the 'power to see' from the frontlines and practical work.

📡 Recommended reading

Click here for a note article on cultivating the power to see as a structure.

The concept of the 'KPI Observation Sheet.' In it, when an anomaly in the numbers is spotted, the structure is designed to break it down by time of day, day of the week, weather, customer demographics, product groups, promotions, and competitors, and then go further to examine frontline contexts such as stockouts, traffic flow, customer service, inventory, POP displays, and layout. This serves as a direct path to applying the latter half of this article to practical work.

It is highly compatible as an article that reads the 'context that AI cannot read,' mentioned in this free article, one step deeper through 'retention conditions.'

What a manager sees creates the reality of the company. The essence of organizational development, deciphered through the quantum mechanical concepts of 'observation,' 'entanglement,' and 'superposition.'

In this free article, the scenario of 'AI recommendations look correct, but there is a sense of discomfort on the frontlines' appears many times.
In this syntax, discomfort is defined not as noise, but as a notification of a discrepancy, preventing intoxication with numbers and the confusion between superficial evaluation and essential evaluation.
This is defined as a way to prevent being misled by numbers and confusing superficial evaluation with essential evaluation.

The conclusion of this article is not to use AI for everything, but to divide roles: AI for standard items, and humans for non-continuous changes. This is a matter of resource allocation. In this syntax, resource allocation is organized not as how hard to work, but as deciding where to place one's efforts, serving as a path to bring the discussion of hybrid operations into management practice.

#AIUtilization #DemandForecasting #GenerativeAI #DX #FoodLoss #Retail #Distribution #LimitsofAI #FrontlinesandAI #TechniquesofObservation #ZObserver #BusinessThinking #LocalManagement #note #TechniquesofObservation
#ZStyleSyntax #ZObserver