One day in the 1990s, Stan Shih, founder of Taiwanese computer maker Acer, drew a curve to convince his employees. The horizontal axis showed the flow from product planning through manufacturing to delivery to the customer, and the vertical axis showed the added value created at each stage. The resulting line was high at both ends and dipped in the middle, shaped like a smiling mouth.
The diagram, later called the “smile curve,” became a map showing at a glance where value resides. More than 30 years later, AI is now quietly but boldly redrawing that map.
What a “Smiling Mouth” Teaches About Where Money Is and Is Not Made
The smile curve reveals a simple but harsh fact. High added value emerges upstream, in planning, R&D, and the production of key devices, and downstream, in branding, sales, and after-sales service. In the midstream center—processing and assembly—profits tend to be squeezed to the limit.
Why does the valley form? Upstream R&D and intellectual property require expertise and major investment and cannot be easily imitated by competitors. Patents and technological advantages therefore support high profits. Downstream brands and customer relationships connect directly to the “reason for being chosen,” making it easier to retain pricing power.
The manufacturing processes in the middle, however, are readily standardized and modularized. Companies in emerging countries can enter them if they have equipment and labor. The result is intense price competition and thin margins. Apple was quick to exploit this structure.
By concentrating resources on planning and design and on sales and service through its own stores and ecosystem, while assigning low-margin assembly to Taiwanese EMS providers—contract electronics manufacturers—it built a remarkably profitable model.
AI Deepens the “Valley”
This curve is no longer only about manufacturing. The same U shape appears in every kind of white-collar work, and generative AI is deepening the valley as never before.
AI is best at repetition, summarization, organization, classification, search, and writing formulaic text or code. These correspond exactly to the “midstream processes” of intellectual production: interpreting ambiguous instructions, researching information, and preparing a draft for comparison and consideration.
AI performs at vastly greater speed and lower cost in areas where junior programmers, administrative staff, and middle managers have earned salaries by “executing intermediate processes.” The standalone value of “being able to write code yourself” will decline sharply in relative terms.
What must not be overlooked is that AI does not flatten the curve. Instead, it sharpens the valley and raises the peaks at both ends. Value becomes polarized. At the upstream left, what matters is the ability to pose questions with no correct answer, determine a direction, and take responsibility for the results.
AI excels at returning the optimal answer to a given question, but cannot decide for itself “what should be asked in the first place.” What remains on the downstream right is “wet” work: facing customers, building trust, and delivering stories. In sales, AI may replace the function of explaining a product, but only people can stand beside someone, understand their circumstances, and take the next step with them.
Overflowing in Cities and Withering in the Regions
This change shakes the very shape of society. In his book *The Disappearance of White-Collar Workers*, Kazuhiko Toyama warns that if decisions are based only on economic rationality, many existing white-collar workers could lose their roles. He compares this with the historical decline of the samurai class during the Meiji Restoration.
More seriously, the change takes opposite forms in cities and regional areas. In urban global industries, AI replaces routine work and creates a “surplus” of white-collar workers. At local industries supporting regional communities—construction, healthcare, nursing care, and transportation—the declining birthrate and aging population combine to worsen a “shortage” of people to do the work.
Administrative jobs pile up in cities while irreplaceable frontline work is depleted. Toyama says this asymmetry is the true nature of the “valley” at the macro level. There are two paths to survival: become an upstream decision-maker who masters AI as a subordinate, or retrain as a highly skilled worker who combines AI with embodied knowledge from the field.
The Day the Curve Breaks into a Bracket Shape
Keio University Faculty of Environment and Information Studies professor Kazuto Ataka calls the future beyond an even deeper valley a “bracket-shaped, or inverted-pi-shaped, society” and sounds a warning.
In a world where autonomous AI agents move across multiple tools and complete work from end to end in one sweep, midstream processes do not merely become “thinner”; they effectively “evaporate.” The intermediate work of opening screens, entering forms, and pressing approval buttons is no longer necessary. What remains is no curve at all, but a right-angled bracket.
The upper horizontal line is “direction that decides where to go,” the empty middle is “generation and optimization that AI completes at the speed of light,” and the lower horizontal line is “critique that evaluates the output and bears responsibility to society.” Human work is compressed into these two ends alone.
This may look like a rational division of labor, but it contains a trap. The volume of information AI processes far exceeds what a person can digest at once. As a result, people may cease to participate meaningfully in judgment and become “responsibility terminals” hanging outside the loop.
They press “approve” without being able to verify the content and excuse themselves by saying, “The AI said so.” What happens is that the design of values—what is good and what should be discarded—is quietly handed to an external AI platform. What Ataka truly fears is not AI running out of control, but people outsourcing human judgment because AI succeeds too well.
Another Smile Curve
So far, this has been a story only of the deepening valley. But the same words have also begun to be used with hope in a completely different context: user retention in the world of software and SaaS.
For conventional SaaS, the ideal has been a “flat curve” in which users for whom the product is not a fit leave early and retention then stabilizes horizontally among those who remain. Generative-AI-native companies, however, are observing a phenomenon that goes beyond this.
After churn lowers retention, the curve reverses upward again at a certain point, creating a “smile” shape. A dreamlike state emerges in which customers who once left return.
The reason is that AI models themselves continue to evolve as they are used. Features that felt inadequate several months ago become capable of handling complex work before anyone notices and bring departed users back. Venture-capital firm a16z cites ChatGPT as a representative example of this behavior. Although it is also a “smile,” this is a brighter curve showing compounding growth.
How to Avoid Sinking into the Valley
What, then, should people possess to avoid sinking into the deepening valley and being reduced to AI “responsibility terminals”? The key lies in leaving the playing field of information processing.
In the era of the IT revolution, the source of value was “information.” But when information from around the world has been learned and AI instantly lays out optimal answers, information itself loses value. What rises to the top in its place is “will”: what future do you want to deliver, and to which customers?
Which risks are you willing to bear yourself? Ataka says that people need “gut” and “intuition” when engaging with AI under uncertainty. Do not wait for all the data to be complete; first choose a direction and start moving, then correct course along the way. This two-stage decision is the posture required to ride the fine horse called AI.
There is also another domain AI can never reproduce: childcare and nursing care, responses to sudden changes in medical settings, or a craftsperson responding to the fire in a kiln. Conditions fluctuate every time, matter is involved, and there are no do-overs. This singularity cannot be reduced to an AI objective function.
As digitization advances, the weight of such “open worlds” increases. Waseda Business School professor Akie Iriyama argues that people should deliberately use “paper” rather than digital devices to train upstream conceptual ability for the same reason. Nothing suits writing down vague sensations and encountering unexpected discoveries quite as well as an expanse of paper.
When we think about it, much of what today’s schools make students compete over—memorization, rapid calculation, and formulaic report writing—is the “midstream” ability at which AI excels most. Without realizing it, we may be training the next generation in skills intended to be replaced.
What matters is not scores on tests with correct answers, but the urgent sensibility of “what do I find beautiful, and what do I want to protect?” If a future arrives in which AI takes over all execution, what finally remains in our hands is the question itself: “Where do we want to go?” The ability to conceive that direction from our own gut, without entrusting it to an external algorithm, should become our strongest weapon for surviving the AI era.
