What AI application companies fear most is not being asked about the strength of their models, but the "human involvement" behind them.

Author: Yan Jun

When AI applications are still confined to demos, press conferences, and financing news, the most frequently discussed topics are model parameters, response speed, and product experience.

However, once a product truly enters the stages of customer on - site deployment, procurement acceptance, financing due diligence, listing disclosure, and exit, a more sobering question arises: For a system that seemingly operates automatically, how much of its operation is actually completed by the system itself, and how much is maintained by human intervention?

This is the "human - involvement ratio" issue that AI application companies will inevitably face.

The so - called "human - involvement ratio" is not a form of sarcasm, nor is it a negation of AI. Instead, it is a very practical assessment of product delivery: The more automated a result appears to be, the more we need to question how much of it relies on human review, remote takeover, quality control, deployment, operational support, and on - site customer maintenance.

The focus here is not on large - scale foundational model companies, nor on chip, computing power, or underlying infrastructure companies. Instead, it is on AI application companies that are closer to the customer's business processes, such as intelligent customer service, enterprise knowledge bases, sales outbound calls, content review, AI agents, intelligent office tools, industry - specific large - model solutions, and various vertical - scenario automation products.

Actually, many people are not unaware that there might be human involvement behind AI applications.

Investors have come across projects where the demos run smoothly and there are customers. However, when they dig deeper into the gross profit margin, delivery cycle, and service personnel configuration, they find that there is a significant implementation and delivery team behind the revenue growth.

Industrial customers have also encountered similar situations. During the pilot phase, the supplier's team is almost always online, ready to address issues promptly. But after the official launch, they realize that the system still highly depends on human backup. Questions such as who can access the data, who is responsible for errors, and how to write the acceptance report become more pressing.

Founders themselves are often more aware of this. In the early stages of a product, relying on human intervention to fill in the gaps, handle on - site issues, and collect samples is a very common approach when AI applications enter real - world scenarios. The real challenge is not whether there is human support in the background at the beginning, but whether, after six months or a year, this human input has been integrated into the product to become an automated capability, or if it continues to exist in the delivery chain, becoming new service costs, operational costs, and organizational burdens.

Two Types of Human Involvement

The automation of AI applications is never just a pure model - related issue.

In real - world business delivery, an AI application product is often composed of models, rules, data, quality control, human review, exception handling, remote takeover, customer processes, and an operational team.

Almost all areas, such as customer service, knowledge bases, review, sales, office automation, enterprise agents, and industry solutions, go through a stage where AI first handles standardized tasks, and humans take over abnormal, sensitive, low - confidence, or high - risk tasks.

The real key is not whether there is human involvement, but where this human involvement ultimately leads.

The first type of human involvement is from those who help the product learn.

They perform tasks such as data annotation, sample screening, model evaluation, quality control sampling, collection of abnormal cases, human feedback, and process training. Their value lies not in performing long - term tasks on behalf of the system, but in gradually integrating abnormal inputs, process breakpoints, long - tail problems, and review experiences from the customer site into models, rules, SOPs, evaluation mechanisms, and product boundaries.

The second type of human involvement is from those who have become part of the delivery structure.

They answer customers on behalf of the system, take over tasks remotely, manually complete orders, handle processes that were originally promised to be automated, or rely on the project team to maintain the delivery results in the long term.

This type of human involvement does not mean that the company has no value, but it will change the company's business nature. Customers may think they are purchasing an AI application, but in reality, they may be getting a human - powered service system with an AI front - end. Investors may think they are investing in a high - margin, replicable software company, but in the end, they may find that behind the revenue growth, there are more deployment, service, quality control, and project operation personnel.

The underlying logic of this judgment is actually quite simple: The human involvement that helps the product learn will ultimately make the system more automated; the human involvement that becomes part of the delivery structure will ultimately make the company more resource - intensive.

This distinction determines not only the cost structure but also the pricing and valuation logic.

A more resource - intensive AI application company can still be valuable, especially in high - threshold industries, complex organizational processes, and highly customized scenarios. However, if a company is essentially more like an "AI - enhanced service," it cannot rely on the story of a standard SaaS or a lightweight software company to discuss revenue quality, gross profit prospects, and replicability in the long run.

Beyond Costs

When many people talk about the "human - involvement ratio," the first thing that comes to mind is cost.

This is, of course, correct. Human review, remote takeover, on - site delivery, long - term maintenance, quality control teams, and outsourced review all directly impact the cost structure and are ultimately reflected in the gross profit margin, delivery cycle, and organizational scale.

However, what is often underestimated is that the "human - involvement ratio" is also a data - related, responsibility - related, and compliance - related issue.

In many AI application scenarios, human review, quality control sampling, customer service escalation, exception takeover, and outsourced processing all involve access to user inputs, voices, images, corporate documents, business records, order information, and even sensitive personal information.

If the system is only used for navigation, marketing, or simple Q&A, the issues are relatively controllable. However, once it enters scenarios such as finance, healthcare, government affairs, insurance, enterprise knowledge bases, legal consultations, elderly care companionship, and customer service, human intervention directly changes the level of responsibility.

At this point, what customers really need to know is not just "what is the accuracy rate," but a more specific set of questions: Can the background human staff access the original data? Has the data been anonymized? Who has the access rights? Is outsourcing involved? Is the access process recorded? Have users been informed? Can the complete processing chain be traced after a complaint?

The deeper an AI application penetrates into the customer's processes, the more it cannot rely solely on "model capabilities" to prove itself. It must also be able to answer questions such as how to write the contract, how to disclose the privacy policy, how to agree on the data processing agreement, how to stratify access rights, how to retain audit logs, and how to hold someone accountable in the exception - handling process.

What customers are buying is not just an interface with a strong "AI feel," but a system with clear automation rates, human takeover rules, data access policies, and responsibility boundaries.

Therefore, for industrial customers, the real questions to ask might be: Are the records of automated and manual completion clearly distinguished? Is human takeover written into the SLA? Can the background human staff access customer data? Is the automation rate and takeover rate clearly stated in the acceptance report? Can the decrease in the takeover rate, increase in usage rate, and reduction in delivery hours be seen during the renewal period?

These questions may not be glamorous, but they are closer to the real purchasing decision than "how many model parameters" or "is it multi - modal." Because once AI is truly integrated into the business system, customers bear the responsibility, not just the imagination.

Entry Point for Due Diligence

In the past, when many AI application companies had not entered the public capital market, external judgments about them often focused on model capabilities, financing news, product experience, and demo results.

It seemed that as long as the product was smart, fast, and seemed "automated," the story was almost complete.

However, as more and more AI companies face a more rigorous information - disclosure environment, many issues that could previously be packaged as "technological progress" will be re - framed as financial, delivery, and revenue - quality issues.

Public reports on the breakdown of prospectus materials have given the market a wake - up call. 36Kr's reports on Zhipu AI and MiniMax mentioned that there are significant differences in the gross profit margins of the two companies, and their cost structures are also different. The reports also pointed out that items such as salaries of service and deployment personnel and computing service fees will create different cost pressures and delivery profiles for different companies.

Another report on the two companies' efforts to enter the capital market mentioned that MiniMax experienced rapid revenue growth from 2023 to 2025, but the industry as a whole was still in a stage of high growth, high investment, and high losses.

These data cannot be simply interpreted.

The cost of service and deployment personnel does not mean that the background human staff is replacing the system in all tasks. Enterprise - level AI applications naturally require deployment, implementation, maintenance, customer support, and local adaptation. The cost structures of B - end, G - end, local deployment, C - end AI - native applications, and API platforms are inherently different.

However, these disclosures at least make one question unavoidable: How is the revenue of AI application companies actually delivered?

If human involvement is mainly part of the product - learning mechanism, as the number of customers increases, the unit delivery cost should theoretically decrease, and the delivery method should become more standardized.

If human involvement is a prerequisite for delivery, then as the revenue grows, the labor cost will also increase proportionally.

The former indicates that the product is becoming stronger, while the latter means that the organization is becoming more resource - intensive.

This is why the "automation rate" has become the most important entry point for due diligence of AI application companies.

The automation rate is not just a marketing term. It ties together product capabilities, labor costs, revenue quality, and capital pricing.

What investors really need to look at is not what the AI has accomplished in the demo, but after the customers' actual use: What is the automation rate? What is the human takeover rate? Are the reasons for takeover recorded? Has the takeover experience been fed back into product iteration? Has the delivery hours for similar customers decreased? Does the improvement in the gross profit margin come from the reduction in model costs or the standardization of delivery? Does the customer renewal come from the system's value or the long - term on - site presence of the service team?

In more rigorous procurement and acceptance processes, customers can even write the upper limit of the human takeover rate, classification of takeover scenarios, and log - retention methods into the terms. Although the specific thresholds cannot be copied across industries, "whether the takeover rate continues to decline" should definitely be an indicator to be continuously tracked during acceptance and renewal.

These questions rarely appear in financing press conferences but will definitely be in the due - diligence questionnaires.

What Founders Should Be Wary Of

For founders, using human resources in the early stage is not a cause for alarm.

Often, human involvement is an essential part of a product entering real - world scenarios: it helps the team understand customers, identify high - frequency problems, recognize boundary scenarios, train processes, and establish a replicable delivery method.

The real danger is when the team itself starts to lose sight of the boundary between which human involvement is training the product and which has become part of the delivery structure.

Once this boundary becomes blurred, salespeople will make over - optimistic claims, customers will purchase as if it were a software product, investors will value it as a high - margin product, while the team is still using a project - based, on - site, and highly operation - dependent delivery method. In the short term, the revenue may grow, but in the long term, the financial statements will gradually reveal the real situation.

If an AI application company raises funds with the story of a software company but delivers in the way of a service company for a long time, when it comes to the next round of financing, listing disclosure, or merger and acquisition due diligence, the capital market will re - evaluate it based on revenue quality, gross profit structure, and delivery replicability.

The problem is not the "resource - intensity" itself, but whether the company has accurately defined its business nature.

Therefore, the earlier founders establish the following internal indicators, the more they can avoid being in a passive position later: Is the human takeover rate continuously decreasing? Is the delivery hours for similar customers continuously decreasing? Is the gross profit per customer improving? Have the abnormal samples processed manually really been incorporated into product iteration? Does the customer renewal come from the system's value or the on - site presence of the team? Has the deployment experience been transformed into standardized modules? Has the manual quality control been upgraded to an automatic evaluation system?

A good AI application company is not one that has no human involvement from the start, but one that can gradually transform human intervention into product capabilities.

Transforming customer - service takeover into intention - recognition optimization, manual quality control into an automatic evaluation system, deployment experience into standardized modules, abnormal samples into model boundaries, and on - site customer problems into the next version of the product.

The real question is whether human involvement can make the product less reliant on human intervention.

Conclusion

When AI applications are still confined to demos, press conferences, and financing news, it is natural to mainly discuss model capabilities.

However, when AI applications enter the customer site, contract acceptance, data processing, financial statements, and the capital market, the questions will become more specific.

Customers will ask: Who is responsible when an error occurs? Regulators will ask: Who has accessed the data? Investors will ask: Is the gross profit margin real? Procurement departments will ask: How should the acceptance indicators be written? The capital market will ask: Can the revenue be replicated? Exit due diligence will ask: Where are the labor costs hidden?

The real world does not reward a friction - free story. Instead, it rewards systems that can gradually establish clearer product boundaries, more stable delivery capabilities, more reliable customer relationships, and a healthier gross - profit structure in the face of data, responsibility, process, cost, and regulatory frictions.

Therefore, it is not surprising that there is human involvement behind AI applications. The key is whether this human involvement can be transformed into product capabilities, making the system more automated, stable, and replicable.

What AI application companies really need to prove is not that they don't need human support from the start, but that as the number of customers increases, scenarios become more complex, and revenue grows, their reliance on human backup decreases, product boundaries become clearer, and the unit economic model improves.

The "human - involvement ratio" is an issue that AI application companies will have to clarify sooner or later. The only difference is whether they clarify it proactively today or have it dug out layer by layer by others during acceptance, due diligence, and disclosure tomorrow.

References:

36Kr: "Breaking Down the Prospectus: Understanding the Two 'Ways of Life' of Chinese Big - Model Unicorns"

36Kr: "Zhipu and MiniMax Rush to the Hong Kong Stock Exchange: AI Unicorns Seeking Capital under the Pressure of Giants"

This article is a contribution from the author, Yan Jun, and is published by 36Kr with authorization.