AI Needs a "Call Button"

11/19 2026 477

Original: Shenmou Finance (chutou0325)

In 2026, the AI industry finds itself in a counterintuitive and awkward phase.

On one hand, demand for AI is skyrocketing. According to OpenRouter data, the Token processing volume dispatched by its platform grew roughly 250-fold in 20 months, reaching 128.4 trillion; the computing power rental market has risen in tandem, with H200 rental prices up 160% over the year and B200 up 120%.

Yet the profits that actually land in companies' pockets are unusually bleak. Also according to OpenRouter data, open-source weight models processed 72% of Tokens in the second week of September but captured only 27% of total platform spending.

Not only do platforms fail to get their share of the money, but model companies themselves are bleeding even more. Take Zhipu as an example: its adjusted net loss in the first half of 2026 was nearly 2 billion yuan; and across the ocean, OpenAI's operating loss rate in the first quarter was as high as 122%.

Surging demand still failed to lift model companies' profits. This inevitably prompts deep reflection: what exactly is blocking AI from turning a profit?

01 Replaced Old Jobs and Unclear New Accounts

When AI entered the market, the first to be hurt were highly standardized specialized tools.

For example, services that previously had to be purchased separately—such as translation, basic customer service, and code completion—are now being swallowed whole by AI large models.

According to an Anthropic report, the task automation coverage rate for computer programmers has reached 74.5%, and for customer service representatives 70.1%; in translation, it is even more severe. According to earlier Microsoft research, AI is already capable of handling 98% of interpretation and translation work.

(Source: Anthropic)

AI has directly taken away these old jobs that were originally supported by a vast number of small and medium-sized service providers. The original old market truly existed, so people's demand for AI is naturally also real.

But although the demand has been captured, the accounting model has not been inherited along with it.

The traditional specialized market bought "results."

For example, buying translation was priced per thousand characters; buying programming was quoted per project. This transaction was like buying clearly priced goods in a supermarket: how much the customer paid and what they got were clear in advance.

Today, however, most mainstream large model companies have switched to another model: like water, electricity, and gas, they adopt a strategy of purchasing by "consumption," and the pricing unit has become the Token.

As a result, what customers face has become a cost account that is vague before the fact.

The first issue is that Token consumption is invisible.

When users use AI to complete tasks, they cannot know in advance how many Tokens will be consumed. A seemingly simple task may have a very long call chain behind it. This "invisibility" can easily lead users to generate a large number of calls without realizing it, and when the bill arrives, the cost is often far beyond expectations.

For example, according to a real case cited by Axios from an AI consultant, a company opened Claude permissions to all employees without setting a budget cap. Because no tool could estimate in detail the Token consumption behind each conversation and each call chain, the company may not have had, and could not have measured in advance, an approximate cost level.

As a result, when the monthly bill came out, it directly created a shocking 500-million-dollar hole. For many listed companies, 500 million dollars is enough to wipe out their entire annual net profit, and here it was only one month. For any company, this can be called a sudden financial disaster.

Similarly, after deploying Claude Code, Uber also underestimated call volume, causing it to exhaust its entire annual AI budget in just four months.

This experience of "impossible to calculate beforehand, shocking afterward" is likely to make customers more cautious in use, for example by imposing usage limits on permissions, budgets, and call scope.

A Cloudera survey once pointed out that because costs are difficult to estimate, 89% of respondents said their companies were forced to shrink or redesign their AI plans this year.

Moreover, this opacity also deprives users of the basis for price comparison.

When buying a bottle of water worth about two yuan, everyone has a sense of it, and cost-effectiveness is clear at a glance; but with Token-billed AI, because consumption varies for each task, users may simply be unable to judge whether it is more cost-effective to use AI for this task or to go through traditional channels.

And even if users grit their teeth and accept this billing method, or see every bill clearly, Token consumption may not necessarily bring users equivalent business value.

Data compiled by Entelligence AI shows that for every 1 dollar of Token investment, only 18 cents generates actual value, 44 cents is used to fix bugs, and 27 cents goes to rework. Accenture research also shows that less than 20% of Token spending can be linked to quantifiable business outcomes.

In other words, companies spend a lot of money buying AI, but more than half the budget ends up paying for its "immaturity" and "mistakes." With this input-output ratio, companies that originally hoped to "lie flat" with AI end up "paying out of pocket" instead.

This imbalance may not only directly drive away old customers and lose repurchase rates, but may also scare off new customers and limit future expansion.

It can be seen that although AI has seized the rice bowl of the old market, unclear pricing and inability to deliver equivalent value make users unwilling to pay. The profits of the traditional market have therefore not been transferred to large model companies.

02 The "Results" That Cannot Be Written Into Contracts

If selling by consumption cannot bring in money, the natural "solution" seems to be going back and inheriting the traditional specialized tools' method of "charging by results." But AI cannot go down this road either.

In traditional tools, results are concrete. Translation delivers a translated manuscript, customer service closes a ticket, code completion generates a program, and so on. Their boundaries are clear, they can be accepted, and they can be priced.

But AI has broken the boundaries of capability. It can write poetry, write code, and even conveniently analyze financial reports. This makes it difficult to use a standard contract to stipulate "what counts as finishing the job" and "how much value there is when so many jobs are done together." This is the dilemma of "attribution."

Attribution is difficult because achieving a real business result is often jointly brought about by multiple links.

For example, a company launches AI marketing, and conversion in the quarter increases by 20%. But of this 20%, how much belongs to AI capability, how much belongs to the assistance of human operations, and how much belongs to a possible cyclical recovery in the market?

Cloud service provider UCloud said this very directly: a business result is often the product of multiple links working together, and it is difficult to attribute the final effect simply to a single call of a certain model. That is to say, when AI, machines, people, the environment, and other factors work together, the slicing of value becomes a dead knot in AI billing.

Stripe specifically issued guidance on outcome-based payment for this, clearly warning that if attribution standards are not clearly defined in advance in the contract, customers and vendors may dispute "whose result it is."

If there is no unified standard for judgment, then charging by results cannot even be discussed.

Even if there are vendors in the market that charge by results, they are limited to single closed-loop scenario areas. Simply put, they generally let AI do only one kind of job, and charge only when AI does the job.

For example, Salesforce, which does CRM overseas, charges for its Agentforce in customer service scenarios only when AI independently completes the closed loop without transferring to a human; domestically, Bairong Intelligent also settles in customer service scenarios by "tickets successfully completed without transfer to a human."

AI has been applied and has created value. A single application may be one solution, but it also narrows AI's capabilities into one scenario and turns it back into a specialized tool.

It can thus be seen that results are hard to define and attribution is hard to split. When general-purpose large models sell "results," they still cannot fully present the market with an operating model that both preserves AI's ability to "do anything" and keeps the accounts clear.

So if general-purpose large models want to break this deadlock, they may need to find a new solution that allows value to truly land.

03 First Establish the Anchor, Then Build the Entrance

When both tactical charging models—selling "consumption" and selling "results"—fail, the key to the problem points to a strategic "absence."

Large model companies have been trying to use AI's general capabilities to enter the market "on multiple fronts," hoping to use AI's cross-domain capabilities to carry out a "great unification" of the originally fragmented specialized markets.

Although AI's general capabilities can indeed swallow large amounts of demand from specialized markets like a black hole, precisely because it "can do anything," it instead appears to lack irreplaceability in specific business scenarios.

In fact, this dilemma of "how general-purpose tools take root in the market" is not unique to AI. Looking back at the evolution of technology products, those general-purpose tools that eventually grew into giants often started out very concretely.

For example, mobile phones. They are themselves general-purpose devices integrating multiple functions, and now people can use them to do many things on the internet. But in the early days, the way they entered the market was very direct: using a piece of physical hardware to bring people the convenience of "being able to make calls anytime, anywhere."

Take the Nokia 1100 as an example. From its launch in 2003 to discontinuation in 2009, it sold more than 250 million units in total, becoming the best-selling mobile phone in history.

Looking at it today, this phone's functions are so rudimentary that they are not even comparable to the little Genius phone watch on an elementary school student's wrist. It had no camera and no app ecosystem. Its 400-hour standby time and slip-resistant, dust-proof shell were designed entirely to guarantee the single need of "being able to dial a call at any time."

As for functions such as checking the time, checking the calendar, and playing Snake, they were added only after the call entrance had been established. In other words, these were incidental functions of Nokia rather than core selling points. Users would not buy it separately just to check the time or calendar, but they were willing to pay for the convenient transmission of information.

Later, Nokia was eliminated by smartphones. That was the result of hardware forms changing in technological iteration, not a failure of this playbook itself. If we strip away the rise and fall of the product itself, the inspiration it leaves for AI remains valid: to gain a foothold in the market, AI must also first rely on a high-frequency, essential anchor.

In other words, today's AI is like a super smartphone stuffed with all kinds of apps. It can write poetry, draw pictures, and write code, but it also needs to find a "call button" of its own.

Admittedly, AI will not, like Nokia, become a "small rectangular object" that users can hold in their hands, but AI can "embed." Whether this "embedding" node is in B-end corporate expense reimbursement and equipment maintenance, or in C-end users' AI search and travel planning.

And once AI is embedded as a node, its output becomes a deliverable that can be clearly recorded by the process. You do not need to see how much credit AI takes; you only need to see whether the efficiency of this node itself has changed.

For example, Lenovo's financial Agent directly covers 32 core financial processes, increasing invoice processing efficiency by 50%. Dongfeng Yipai's intelligent agent "Equipment Master" shortened machine fault response time from 10 minutes to 4 minutes, directly reducing corporate losses...


(Source: Feishu)

It can be seen that once AI is embedded in a specific task node, it gains a container that can clearly demonstrate delivery value. Only on this basis can large model companies rebuild a charging structure that makes people willing to pay.

As for what this charging structure is, it is not yet known. In the future, it may be a fixed system subscription fee, a commission based on saved costs, or some evolved dynamic pricing model.

But no matter which choice is made, if the AI business is to truly become profitable, AI value must be "calculable" and "visible"; otherwise, even the most perfect charging model is just empty words on paper.

After all, in the business world, value must come first before talking about collecting money. If a solid value anchor has not even been identified and one rushes to settle accounts with customers, it is destined to be nothing but a cruel "unrequited love."

* Images are from the internet. If there is any infringement, please contact us for deletion.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.