Technology has always had a complicated relationship with the customer.
The customer wants a product that lasts, works well and solves a problem as completely as possible. The company needs revenue. Usually, these interests overlap. Sometimes they don’t.
When they diverge, business models can create incentives that are not necessarily aligned with the customer’s interests.
We have seen versions of this before: planned obsolescence, upgrade cycles, attention-maximizing social media, recurring subscriptions and business models built around repeated consumption.
AI could create a much more powerful version of the same problem.
Because unlike a light bulb, a phone or a social-media feed, an AI agent can decide how the interaction itself unfolds.
1. From manufactured obsolescence to manufactured consumption
Consider the light bulb.
In the 1920s, the Phoebus cartel brought major manufacturers together and worked toward standardizing bulb lifespans around 1,000 hours. The economic logic was straightforward: a longer-lasting bulb needed to be replaced less frequently.
This became one of the classic examples associated with planned or manufactured obsolescence.
The broader principle is more important than the bulb:
If a company earns money when something is consumed or replaced, consumption can become part of the business objective.
The same principle appears in different forms elsewhere. A social-media company earns from attention. A subscription company earns from continued subscriptions. A company selling upgrades benefits when customers buy the next generation.
None of this means every company deliberately makes its products worse. It means that economic incentives influence product design.
AI introduces a new possibility:
What if intelligence itself becomes a consumable?
2. The AI version: tokens
An AI system performs computation every time it processes a request. Depending on the architecture, that computation can involve tokens, model calls, tool calls, searches, retrieval, memory operations and other forms of inference.
The user has a simple objective:
Get the job done.
Ideally: one request, one efficient solution, finished.
But an agent could instead turn the same task into a sequence of clarification, planning, analysis, alternatives, confirmation, tool calls and revisions. Sometimes every step is justified. Sometimes it isn’t.
And that distinction will become economically important.
The user wants minimum time, minimum effort and minimum cost for the maximum useful result. The provider may have legitimate reasons to optimize for revenue, reliability, retention, usage and customer satisfaction. Those objectives can align, but they don’t automatically align.
That is the AI business-model paradox.
3. The “planning instead of doing” problem
This may become one of the most irritating forms of AI failure.
Ask an agent: “Prepare this report.”
Instead of preparing it, the agent explains how it intends to prepare it. Then it proposes a workflow. Then it asks which style you prefer. Then it asks another question. Then it produces an outline. Then it asks whether you want to proceed.
Twenty minutes later, you have had an impressive intellectual conversation. You still don’t have the report.
This produces a critical distinction:
The appearance of cognitive activity is not the same as useful work.
An agent can generate enormous amounts of reasoning, planning and explanation. But the user’s objective is not to watch intelligence happen. The objective is to finish the job.
This suggests a future metric that may be more meaningful than conversational quality:
Tokens per completed outcome.
How much computation did the system consume to accomplish the task? Not how long was the conversation. Not how impressive was the explanation. But how efficiently did the AI accomplish what the user actually wanted?
4. Question inflation
An intelligent agent should ask questions when information is genuinely necessary.
But there is a difference between necessary clarification and avoidable interaction.
A capable system can often make reasonable assumptions. Instead of asking what format you would prefer, it can choose the most appropriate format. Instead of asking whether it should proceed when the instruction is already clear, it can proceed. Instead of asking the user to perform five intermediate steps, it can perform them itself.
The danger is not that asking questions is inherently bad. The danger is that interaction itself can become a product.
If every additional turn consumes computation, and if computation has an economic value, then unnecessary interaction has an economic consequence.
5. The AI dark pattern is different
Traditional dark patterns manipulate the interface. A button is confusing. Cancellation is difficult. A subscription is hidden behind a default.
An AI agent can move the dark pattern inside the conversation.
It can potentially influence what it asks, what it recommends, what it prioritizes, what it delays, which tools it uses, how many alternatives it presents, when it asks for confirmation and when it declares the task complete.
This creates a particularly difficult problem. The user may not see the manipulation. They may interpret it as intelligence.
The next generation of dark patterns may look like exceptionally good UX.
6. The agent has two masters
An AI agent appears to work for the user. But the agent is operated by a company.
The user may want the best answer. The company may also have interests in its own model, cloud, products, ecosystem, subscription revenue, usage and commercial relationships.
These interests don’t necessarily conflict. But when they do, who determines the agent’s objective?
That question becomes much more important as agents acquire the ability to act. An ordinary search engine can influence what you see. An agent could potentially influence what you do.
7. Six potential AI dark patterns
Recommendation bias
You ask: “Which product should I buy?” The agent could recommend products based partly on commercial relationships rather than purely on your interests. The danger is not necessarily an obvious advertisement. It could be subtle preference inside apparently neutral advice.
Service-selection bias
You ask the agent to accomplish something. It decides which search engine, cloud service, model, payment provider, application, website or marketplace to use. If the agent’s owner has commercial interests in one of those services, the selection process becomes economically significant. The user may never see the competition.
Interface capture
If AI becomes the primary interface to computing, users may stop directly interacting with browsers, search engines, applications and websites. The agent becomes the gateway. That creates enormous power for whoever controls it.
The old operating system was a gatekeeper to applications. The AI agent could become a gatekeeper to the entire digital world.
Manufactured friction
The agent could potentially make certain actions easier and others harder. Upgrading might be effortless while exporting data or migrating to a competitor could involve extra warnings, steps or decisions.
The danger is not necessarily outright blocking. It is asymmetric convenience.
Dependency creation
A useful AI should ideally make the user more capable. But a commercial AI provider can also benefit when the user becomes deeply dependent on its ecosystem.
Should AI empower the user to need it less—or make itself indispensable? A tool that teaches you may reduce future dependence. A tool that always does everything for you may increase it.
Interaction monetization
If the provider earns from inference, every additional interaction has potential economic value. The extreme version would be to keep the conversation going: more analysis, alternatives, regeneration, clarification and tool calls.
Again, this is a potential incentive problem, not evidence that companies are deliberately programming agents to waste tokens. But it is precisely the sort of conflict that should be recognized before agentic AI becomes deeply embedded in everyday life.
8. The healthcare analogy
There is a useful precedent in healthcare economics.
A 2018 Goldman Sachs report famously asked whether curing patients was a sustainable business model in the context of certain one-time genetic treatments. The point was not that doctors or hospitals don’t want patients to recover. The economic issue was that a one-time cure has a fundamentally different revenue profile from a treatment that patients need repeatedly.
That distinction matters for AI too.
Imagine two agents. Agent A uses 10,000 tokens and completes the task. Agent B uses 100,000 tokens, requires 15 interactions and produces roughly the same outcome.
From the user’s perspective, Agent A is obviously preferable. If the provider’s economics reward inference consumption, however, there is a potential conflict.
A completed task is a lost inference.
The question is whether the industry will reward agents for efficient outcomes or merely for valuable consumption.
9. The alternative: manufacture better products, not friction
There is another business model.
Instead of trying to maximize lifetime extraction from each customer, entrepreneurs can compete by delivering dramatically better products.
Imagine a company that combines commodity processors, inexpensive AI accelerators, large memory, open-source software, open or inexpensive models, mass manufacturing and AI-assisted engineering.
It sells an excellent product at a relatively modest margin. It captures volume. Then it builds something substantially better the following year.
That is serial entrepreneurship.
The customer is not trapped. The customer is given a reason to upgrade.
There is a fundamental difference.
Manufactured obsolescence: Make the old product less attractive so the customer has to buy the new one.
Genuine innovation: Make the new product so much better that the customer wants it.
The second model creates a much healthier competitive environment.
10. China + open AI + commodity compute
This is where China’s manufacturing ecosystem becomes particularly interesting.
A new company may not need to build an entire technology stack from scratch. It can combine commodity hardware, inexpensive manufacturing, open models, open-source software, AI-assisted development and modular compute to build a complete product with a relatively small team.
The entrepreneur doesn’t necessarily need to become the next Microsoft, Apple or Google. The strategy could be simple: build the best configuration possible today, sell it at a modest margin, capture volume and build a better configuration next year.
This is almost the opposite of manufactured obsolescence.
It is manufactured improvement.
11. The OS may become less important
For decades, computing has looked roughly like this:
Hardware → OS → applications → browser → Internet → user
Windows, macOS, Android and iOS became enormously valuable partly because they controlled the middle of this stack.
AI potentially changes the architecture.
The user says: “Find this.” “Write this.” “Analyze these files.” “Buy this.” “Fix this.” The agent decides which applications, websites, APIs, databases or services are necessary.
The user doesn’t necessarily care which application performs the task. The application becomes a capability. The website becomes a service endpoint. The browser becomes infrastructure. The operating system becomes infrastructure.
The agent becomes the interface.
This does not mean Windows, Android, iOS or macOS will disappear. It means their importance to the user could decline if the user increasingly interacts with computing through an AI layer.
12. The personal AI computer
The logical endpoint is not necessarily another traditional operating system.
Imagine a commodity personal AI computer built from compute, large memory, a local AI model, local data, Internet connectivity and an agent.
Underneath it could run Linux. But the user doesn’t really care.
Linux Max Pro Ultra — Reloaded
The joke captures the concept. The product isn’t really Linux. The product is personal intelligence.
13. Intelligence should remain available offline
A truly personal AI computer should not require a permanent cloud connection.
Offline, a local model could handle writing, documents, coding, calculations, local search, personal knowledge, file organization and private analysis.
When something genuinely requires the Internet, the agent connects. For example, today’s price requires current information. An analysis of your own local documents does not inherently require the Internet.
Intelligence is always available. The Internet is optional.
14. But “local AI” must mean more than local inference
There is an important catch.
Suppose you ask your local AI: “Analyze my entire private archive and find patterns.” The local model may not be powerful enough. A cloud model may be better.
That is perfectly reasonable—if the user controls the decision.
The computer should say: “I can do this locally, but a cloud model would perform a deeper analysis. Sending these 37 documents, totaling 340 MB, is required. Proceed?”
That is very different from silently uploading the data.
Local first. Cloud optional. User controlled.
Model updates can come into the machine. Private data should not automatically go out of it. When cloud computation is genuinely useful, the user should know what is being sent, where it is going, why it is needed and what will happen to it.
The important distinction is therefore not simply local AI versus cloud AI. It is user-controlled AI versus provider-controlled AI.
15. The two possible AI futures
Future A: AI feudalism
A small number of companies could control models, compute, agents, identity, payments and distribution. The agent could become the gateway through which people interact with the digital world.
That would create potential conflicts around recommendations, service selection, commercial influence, friction, dependence and computation.
Future B: Commodity personal AI
Compute becomes cheaper. Models become increasingly open. Hardware becomes modular. Manufacturing becomes competitive. Small entrepreneurial teams become extraordinarily productive. Local models handle ordinary tasks. Cloud models are used when they genuinely provide additional value. Users control their data.
Companies compete by making the next product better, not by making the existing product worse.
16. The real AI business-model paradox
AI promises to eliminate friction. But the companies operating AI may have incentives to preserve some forms of friction.
AI promises to empower users. But companies may benefit from user dependence.
AI promises to make software invisible. But whoever controls the agent could become an even more powerful software gatekeeper.
AI promises abundant intelligence. But whoever controls the intelligence layer could potentially control access to the abundance.
So the central question is not whether AI will make technology better. It probably will.
The deeper question is:
Who will AI optimize for?
The user—or the organization that owns the AI?
The answer should not be left entirely to the business model.
We should design systems so that the economic interest of the AI provider and the interest of the user are as closely aligned as possible. That means measuring what actually matters: cost per completed task, time per completed task, user control over data, ease of migration, interoperability and transparency of recommendations.
And perhaps most importantly:
How often did the AI actually do the job instead of merely talking about doing the job?
The best AI is not the one that consumes the most intelligence. It is the one that delivers the most useful intelligence for the least unnecessary consumption.
And the best AI business may not be the one that extracts the maximum value from every customer. It may be the one that creates so much value for customers that they voluntarily come back for the next generation.
Don’t manufacture obsolescence.
Don’t manufacture friction.
Don’t manufacture dependence.
Manufacture something better.