AI Agents Working on Performance-Based Pay vs. Japanese Workplaces Where Approval Processes Stall: A 57-Year-Old Supervisor's Last Weapon Was 'Mud'

I read the article written by paiza Chairman Ryohei Katayama on note titled 'The Arrival of AI Agents That Sell and Earn: Sierra's Commission-Based Model'.

Sierra, an American AI company introduced in the article, provides AI agents for customer support. However, they are completely different from what you might imagine when you hear the word 'chatbot'.

A customer sends an image saying, 'The product is defective.' The AI looks at the image and determines whether it is defective. If it confirms the product is defective, the AI completes the arrangements for shipping a replacement. No human intervention is required at all.

That is not all. When a customer calls to cancel, the AI analyzes their usage history and prevents the cancellation by suggesting, 'It would be more cost-effective to switch to this plan.' It even performs cross-selling by recommending other products. And if that generates sales, Sierra receives a commission.

AI handles customer service, makes decisions, processes requests, and even generates sales. This is the reality of 2026.

And the clincher is the billing model. Sierra's pricing structure is 'performance-based pay for completed tasks.' Fees are only incurred when the AI autonomously resolves an inquiry; if it escalates to a human, it is free. According to Mr. Katayama's article, the inquiry resolution rate for AI agents at one brand has already reached 70%.

Reading this article, I saw the 'death' of the industry I am in.

I am a 57-year-old call center supervisor. I manage about 30 operators and handle around 60 client projects.

The business model of the Japanese call center (BPO) industry is fundamentally built on 'number of seats' and 'operating hours.' How many operators are deployed and how many hours they work. In other words, we are selling 'human labor.' Clients pay for the time that people spend answering phones.

What Sierra is selling is 'results.' Whether an inquiry was resolved or not. That is all.

In a world where AI agents that provide 24-hour support without wait times, instantly grasp customer purchase history, and even handle cancellation prevention and cross-selling work on a 'performance-based' model, a proposal that says 'we will line up dozens of people at such-and-such hourly rate' cannot possibly compete.

Mr. Katayama's article states: 'What overseas AI application layers are targeting now is not software costs, but labor costs and outsourcing expenses.'

The revenue of call center BPOs is exactly that 'outsourcing expense.' We are right in the middle of the market that AI agents are targeting.

So, are we, the people on the front lines, just waiting to be replaced and discarded by AI?

I do not think so.

Because when you shift your gaze from the cutting edge of Silicon Valley back to the Japanese workplace, you see an endless expanse of 'mud' spreading out there.

At my workplace, even now in 2026, clients still request reports to be submitted by fax. This is mainly for projects involving local governments. I think, 'Well, since the other party is asking for it, it can't be helped.' But it is also true that the feeling of 'Can't we do something about this?' arises.

I pay out of my own pocket for three AIs—ChatGPT, Claude, and Gemini—and use them as sounding boards for my work. However, these self-funded AIs are what you call 'shadow AI.' I cannot feed them confidential information or client customer data. That is why I proposed that the company let me use a paid license for Microsoft Copilot, and I received approval at the end of February this year. I was able to use it for work for about a month.

In that month, I deployed Copilot for the automatic generation of flashcard Q&As for operator training and for report creation, achieving a result of approximately 800% in ROI terms.

However, after the license expired at the end of March, I submitted a request for approval to renew it in April, but my shifts have been missing my boss, who is the decision-maker, and it remains stalled at this very moment.

In his opening address for the term, the president of our parent company clearly stated, 'AI is not just a trendy buzzword; it is crucial that we actually apply it in the workplace.' In that same company, the approval process for an AI tool has been stalled for days due to a physical disconnect. I swallow this contradiction in silence.

That is not the only contradiction.

At my company, we use Google Chrome as our browser. Yet, Google services are prohibited for security reasons. We spend every day navigating the situation of being unable to use Google while having Chrome open.

There is also the barrier of operator IT literacy. We have old manuals and talk-flow documents. Ideally, we would migrate them to digital and make them searchable. But there are voices saying, 'Paper is easier to read' or 'I'm used to the current materials.' We still haven't been able to abolish these moldy manuals.

This is the reality of the workplace in large Japanese corporations.

No matter how excellent a Sierra AI agent may be, or how quickly it can process customer complaints in seconds, it cannot persuade a client representative who insists, 'You must submit the report by fax.'

AI also cannot sit next to a veteran operator who complains, 'I can't do my job without paper manuals,' and listen to them while looking at the screen together.

Japanese business workplaces are filled with 'entanglements' and 'local rules' that cannot be explained by rationality. Simply introducing excellent overseas AI models as they are will never purify this mud.

Towards the end of Mr. Katayama's article, it is written as follows:

'There are countless non-general tasks in every domain of every industry, and it is practically difficult for foundation models to optimize for them.'

Even Sierra does not just drop their product in; they embed themselves in the workplace, grasping the nuances of operations—such as 'it's a 30-day return policy, but if the customer has a purchase history, 45 days is actually okay'—before incorporating them into the AI.

The mud in Japanese call center workplaces is on a different level.

This is where the real issue begins.

I believe that the path for us call center supervisors to survive in the coming era is not 'managing human operators to earn an hourly wage.'

The path to survival is to become a 'translator (adapter)' who adapts slick, excellent AI to the muddy, hands-on operations of the Japanese workplace.

I have a feel for that.

During the one month I was able to use Copilot, I had Copilot read the client's operational manuals and worked on automatically generating Q&A for operator training.

What came out at first was a mountain of useless Q&A.

'What is the official name of this service?' 'What are the reception hours?' It was just the table of contents of the manual converted into questions. It didn't even graze the points where operators on the floor actually get stuck.

What new hires at call centers get stuck on is not the phone handling itself. It is the 'after-call work' after hanging up the phone. How to input data into the system. When and how to relay information to the client. It is written in the manual, but they don't know where it is written. That is where their hands stop.

AI can read the text in a manual. But it doesn't have the on-site intuition of 'how many seconds a new hire freezes after hanging up the phone.'

That's why I created a workflow where I narrow down the prompts I feed into Copilot to 'post-processing steps' and 'handover methods,' strictly specify the output format for Excel pasting, and have it output five items at a time for human review.

Even so, half of the Q&A that Copilot produces cannot be used as is. There is a deep chasm between 'what is written in the manual' and 'what can actually be done on the floor.' The AI's output is correct, but it lacks specific file names, recipient addresses, and subject lines. Even if a new hire reads it, they won't know how to proceed.

Bridging that gap is the job of a supervisor who knows the floor. This is what I call 'translation.'

Feeding the AI the mud of the workplace. Taking the 50-point deliverables the AI spits out and finishing them into 100-point work within the context of the floor. Standing between the client who demands faxes and the AI. Dodging the boss's shift to get approvals passed.

No matter how much AI evolves, the final role of 'taking the mud' can only be performed by a human who knows the floor inside and out.

The business model of lining up operators by hourly wage is undoubtedly heading toward an end. The Sierra case study clearly demonstrates this.

But I have no intention of being on the side that gets weeded out.

As a 57-year-old non-engineer, what I am aiming for is to stand on the side that 'uses AI to the fullest in the gritty, muddy workplace.'

There are approval documents that remain stalled. There are moldy manuals. There are report requests that arrive by fax. There are environments where you can't use Google even while having Chrome open.

The mud won't dry tomorrow, either.

But I believe that a person who can make AI work within that mud is the most needed existence in the coming era.

【Read also】 Even if you know how to use AI tools, 'approval' won't pass in old Japanese companies. I have published here the 'actual proposal for guaranteed approval (a hack for office politics)' that I used to successfully win a Copilot ID extension.