Original · Unique Research · 2026-05-28
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, four insight sections, closing analysis, and the full panel transcript including all named speaking turns and their continuation paragraphs. Product, user, efficiency, cost, market and business figures are source or speaker claims, not independently audited findings. Product names, company names and named people are preserved as source attributions; official English forms are used where known and transliterated where unverified. The source is dated May 28, 2026; temporal markers are preserved as stated.
AI Industry Truths
Where Is AI Implementation Stuck? Four Frontline Founders Said the Same Thing
It's not that the model isn't good enough, it's not that the engineering isn't done well—what's stuck is who signs off
"
Four people, four different industries,
and the answers point to the same thing: AI can already do the work, what's stuck is who signs off.
It's not that the model isn't good enough, it's not that the engineering isn't done well.
It's that there isn't yet a mechanism for AI's output to be formally held accountable.
Another roundtable in Shenzhen, four people from completely different directions sat together: someone doing AI office, someone doing financial sales training, someone doing an AI collaboration platform, someone doing offline retail AI.
They chatted for nearly an hour, and finally the moderator asked a question: which AI capability in your product do you feel is already good enough, but you still don't dare to release?
The four people's answers were almost identical.
First, What Changed in the Metrics
On the AI office front, Zhou Ze'an (ChatPPT, 17 million users) put it clearly: speed is no longer a metric.
In 2023, they hit 1 million registered users in 45 days. Now, just in China, there are 372 standalone AI document-writing products that you can name. Everyone can produce a report in 10 minutes—speed has thoroughly become infrastructure, not a differentiator.
So what is? What they're watching now is called "document decision ratio"—when AI outputs a piece of content, how much does the user actively modify, how much do they confirm, how much do they skip. The higher this ratio, the closer AI is to the user's true intent. Currently, the direct-output rate they measure is around 50%. For documents, having this number is already not low, because documents often need to incorporate personal content, and most of what AI produces can only be called reference.
There's another thing more easily overlooked: the value of a document isn't decided by the creator—it's decided by the audience.
The business plan (BP) he shows investors is now made in Live format. What investors get is a version that clones his voice and speaking rhythm, and they can ask questions directly. He collects what investors really care about, then goes back and revises. PPT is no longer one-way output—it becomes a tool for collecting audience feedback.
In the past, all office software was designed with the creator at the center. The real value lies with the audience, but software has never solved this.
Li Shouguo from Beta Data does AI sales training for finance. His north-star metrics fall into two categories: for service-communication roles, the core is pass rate—after training, can they pass industry certifications or internal institutional assessments; for mid-to-senior sales roles, the core is cost—compared with human teachers, how much has the cost dropped to achieve the same training effect. Originally, hiring a human teacher cost 20,000 to 30,000 yuan a day; the AI version is roughly 2%-3% of the original cost. The metrics themselves aren't innovative, but if you can break through, the business works.
Howard from COCO AI offered a judgment: the metric for measuring individual output should now be "output per unit of attention." Not human efficiency, not work hours—attention. AI can help you execute, but direction, standards, and acceptance are still human, and they consume attention. Back at the organizational level, the word he uses is "throughput"—with the same number of people, how many product lines can you simultaneously manage.
He also gave a counterexample: many domestic companies use Token consumption to assess employee performance, then found that employees purelyscam (game) Tokens with no actual output, and this set of metrics quickly became ineffective. Ultimately you still have to look at output itself.
Chris Yang from Aimo Technology does offline retail AI. He said the biggest internal change is: he now communicates most with young people and the students he mentors, not wanting past experience to hinder organizational development. The biggest external change is: helping major clients shift from experience-driven to data-plus-AI-driven.
Why Big-B Implementation Is So Hard
After stepping into many pitfalls, Li Shouguo summarized a shortcut: to push an AI project, the transaction decision must be singular—the department paying the bill must be able to decide on its own.
He put it bluntly: if you want a project to fail, the best way is to drag several more departments into meetings together.
Another condition is: this thing, in the original process, must require cross-team, multi-role collaboration to accomplish. The greater the friction, the more obvious AI's replacement value, and the faster the push. They launched a product for large-ticket insurance sales in April, and in less than two months it was close to profitable—precisely because both conditions were met.
Chris Yang, who does offline retail AI, talked about another dimension: within big-B clients, you must find an internal "coach" willing to polish the product with you.
This coach doesn't necessarily have a high rank, but must be familiar with business processes and willing to hand over years of accumulated experience to you, so you can judge what can be reshaped with AI. He doesn't care much whether you're an AI company—he only cares whether you can truly solve his problem. Finding this person is more important than landing a big-client contract.
He also told a specific case. You walk into a store to buy mineral water—you don't care whether it's Nongfu Spring or C'estbon, you just grab the bottle closest to your line of sight. All FMCG brands spend marketing budgets fighting for this position. But after the budget goes to the agency, how well it's executed—previously you could only rely on people visiting stores one by one. Now they use AI for store audit: after each marketing action is completed and AI verification passes, payment is made. The budget shifts from upfront to post-payment.
Trying to push this through the Marketing department got them turned away every time—this tool essentially audits their execution results, so of course they oppose it. Later they went directly to the boss, who welcomed it very much.
This kind of thing only moves forward when you find the right person.
How to Give Tacit Knowledge to an Agent
Howard told a detail.
His own Agent can now make some decisions that are very similar to what he himself would make. Inside this Agent, about 35 of his thinking principles have been deposited.
But these 35 weren't written in by him—he himself didn't even know there were 35.
How did they come about? AI, in the collaboration records of his daily work, watched how he evaluated whether something was good or bad, how he made choices, and slowly distilled them.
His conclusion: tacit knowledge can't be relied on to be summarized and written in by yourself—you have to let AI extract it from daily collaboration. The more you try to write rules clearly, the more you can't finish writing, and the less accurate they become.
Managing Agents and managing people, he feels, has one essential difference: managing people requires caring about psychology, motivation—you have to coax them; Agents don't—give them the work and they go do it. But for an Agent, you have to give sufficient context, set clear standards, and be responsible for acceptance. What's consumed differs, but both require human attention.
That Last Question
The moderator asked: which AI capability in your product do you feel is already good enough, but you still don't dare to release?
Zhou Ze'an: Targeted delivery. The output of specific solutions—behind it are issues of responsibility and expectation, which can't yet be fully handed to AI.
Li Shouguo: Compliance review. Financial AI has reached the capability at this step, but regulation requires the final link to be a person—this cost can't be brought down.
Howard: Same as what Li Shouguo said—it's the issue of responsibility. So they set a final responsible person for each Agent, which must be a specific person to sign off.
Chris Yang: What I most want to do is directly help clients operate stores and share in the profits. If one store makes an extra 20,000, just give me 10,000. The capability is there, but it can't be pushed through.
Four people, four different industries, answers pointing to the same thing: AI can already do the work, what's stuck is who signs off.
It's not that the model isn't good enough, it's not that the engineering isn't done well. It's that there isn't yet a mechanism for AI's output to be formally held accountable.
Until this mechanism emerges, all the most valuable features can only be used internally and can't be released externally. This isn't an AI problem—it's an institutional problem, and it's also where the next real opportunity lies.
More Conversation Details
Guests:
Biyou Technology Founder — Zhou Ze'an
Beta Data CEO — Li Shouguo
COCO AI CTO — Howard
Aimo Technology Founder & CEO — Dr. Chris Yang
Moderator: Unique Research Partner — Xue Qian Amber
Xue Qian Amber: The theme of our Panel today is actually a very open, phased topic—about Harness Engineering, the implementation and experience of human-machine collaboration. Actually, last year when we chatted with everyone, it was a lot about prompt engineering, then context management, and now at this point, the term has evolved to "Harness Engineering." Actually, all the guests here who do delivery are very familiar with Harness, especially because everyone's key focus this year is how to deliver, how to let AI and humans enter the organization and workflow, and truly deliver the product. The four guests here today are actually from different fields. This Panel is also a good opportunity to learn more from their experience in their respective business scenarios. First, could each guest give a self-introduction, then in relatively simple terms tell everyone what the most critical or painful specific problem you're solving is. Alright, let's start with Mr. Zhou.
Zhou Ze'an: Hello everyone, I'm Zhou Ze'an, founder of Biyou Technology. First, a brief introduction—Biyou Technology is actually a team very vertically focused on the AI office field. Our team, including myself, is probably like the previous panel said, considered "veterans" of this industry—we've been in the AI office industry for almost 13 years. What we do is very simple. Compared with others who face the B-end, we're more scenario-oriented—providing documents that let you quickly create something that meets the desired effect. We have several products, including our enterprise star product ChatPPT with about 17 million users, and our projects like YOO Resume adding up to about 80 million users—all serving a very vertical document collaboration product, and of course we have both To B and To C. Back to today's theme—what problem are we solving? Actually, our definition of AI differs from other teams. Although we do AI office and AI documents, we're not trying to let you write things out quickly. We're more—consistent with today's theme—enabling you to produce and deliver documents to your clients, interacting employees, and reporting scenarios.
Li Shouguo: Hello everyone, I'm Li Shouguo from Beta Data. We're making an AI sales training product, because the skill of sales is like swimming or driving—no matter how many times a teacher explains it, you're better off just getting in and practicing, because it's fundamentally a muscle-memory thing. But the cost of role-play practice between people is actually very high, especially in complex sales. So we use AI to create various virtual clients, letting these virtual clients accompany salespeople in skill practice. Currently we're mainly in the financial industry—banks, securities, insurance are relatively more common, and we've also done some in industries like power, telecommunications, and healthcare, but the main body is still banks. The core is actually cost reduction. Because in these large state-owned enterprises, improving sales skills and cultivating employees' sales capabilities used to rely more on human teachers to solve problems. But in the past two to three years, cost reduction and efficiency gains have been severe—mainly cost reduction, not much efficiency gain, and training budgets have been cut quite heavily, so many things can't keep up. But AI may only cost 2% or 3% of the original, so it still brings a relatively large cost reduction.
Howard: Hello everyone, I'm Howard, CTO of COCO AI. What we're building, COCO AI, is actually an AI Native collaboration platform. We have two layers: the first is our self-developed OS framework, similar to what you may be familiar with as OpenClaw—it gives our AI memory and the ability to autonomously schedule work 24/7. The second is something we call C-Workspace. Its core problem to solve is enabling our real humans and AI to collaborate in a mixed way, making AI's Harness more controllable, and in the process depositing our enterprise's digital assets.
Dr. Chris Yang: My name is Chris. Listening to the previous guests' sharing, I feel quite emotional—I'm probably the oldest one sitting here. Because I've been doing artificial intelligence—from my first paper to now—it's been exactly 20 years. The first ten years were in academia, because at that time the conditions for industry didn't exist yet, so I became a professor—from undergraduate, master's, PhD to professor, I've always been in the AI direction; the past ten years have been in industry. Our current core direction is actually a bit different from everyone else—most AI directions you hear about are industries that have already been digitized, like documents, marketing, e-commerce. We focus more on AI plus physical industries. Our world—actually 70%-80% of GDP—is still supported by physical industries, including retail, cinemas, shopping malls, logistics, and so on. How these industries, which naturally haven't been well digitized, combine with current AI to reshape their workflows and help reduce costs and increase efficiency—the difficulty may be higher, but at the same time we feel that once you get into this, the barriers will be higher. After 20 years, I personally feel, including what we're talking about today with Harness Engineering—in the virtual world, making a mistake and running the program again at most consumes some Tokens; but in the physical world, the cost of your mistakes can be very, very large, so Harness Engineering becomes even more critical for us.
Xue Qian Amber: Our four guests actually have businesses leaning toward To C and some toward To B. I want to help everyone grasp some key information: now that each of your AI has entered business scenarios, it must have brought some changes to these traditional clients or to your own measurement metrics. For example, those doing offline retail must be watching certain metrics that have changed; and Mr. Li doing financial sales training is mainly about cost reduction. Also, like many large foreign companies, even employee assessment metrics have become Token consumption—these are things we could hardly imagine before. So now I'd like to ask each guest: among the clients you serve, or yourselves, do you feel the core north-star metric has changed? Let's start with Mr. Zhou again.
Zhou Ze'an: Regarding the north-star metric of a product or from the client perspective, let me first trace back—because we do office documents, this industry has developed for nearly 40 years, a very typical scenario product strongly bound to the software era. The previous generation of all office document scenarios solved the tool attribute itself—you have an idea, now you need to output something the client can understand, using software tools to help you accomplish this. So in the past, the core metric all our office software solved was completion rate—previously it took five hours or three days to finish writing, now AI helps you solve it in ten minutes. This is the change we might understand at a shallow level.
But the real change, I think, has now undergone an even greater transformation. Because I truly felt it—from when we made the first ChatPPT in 2023, hitting 1 million registered users in 45 days, we were ecstatic; to now, recently joking with colleagues, let's see how many standalone AI document-writing products there are—not counting those with built-in capabilities under ecosystems—there are 372 you can name in China. This means a more important metric has changed: speed is no longer a metric, because everyone gets a report in 10 minutes. It becomes a very important signal: what is the quality of what's written? This means in our definition, we've quantified a concept called "document decision ratio." What is decision ratio? When AI outputs a piece of content at once, compared with two other people simultaneously giving articles, how do you show differentiation? It's not in the model itself, because models at a certain stage are more or less the same. This process greatly tests the human-AI interaction—can I discern what you want, should I give you Options, do you choose A or B, or just skip directly. There's a ratio issue here—in each communication with the user, the activated parts have decision components, which can confirm whether the user actively modifies, or Yes or No. When the weight of your entire decisions gets higher and higher, it means the entire scenario chain can better discern the effect—that's the first point.
The second point may be another dimension we've recently started exploring. When documents develop to a certain stage, besides the so-called decision key points, there's a particularly easily overlooked signal. In the past, all office software stayed at efficiency, with the creator himself deciding value—"today it saved me three hours" is actually all about oneself. But the true value of a document doesn't lie with the creator at all. The PPT I present on stage has its value decided by investors; the report I give to my boss is whether the boss finds it understandable. This means its audience decides the document's value, but in the past all software chains didn't solve this. If AI has solved efficiency and activation, the real next step should be paying attention to the audience's Yes or No for this document. So we advocate that our current products should be as Live as possible. Like my own BP (business plan) for investors now—I always let investors directly take the version that clones my voice and my speaking rhythm to view themselves, they can ask questions at any time, and I need to collect what the client actually cares about, then I go modify. There's also something called user engagement in all this, which may determine the value manifestation of your entire document—these are two very important metrics.
Xue Qian Amber: Understood—let me use a slightly crude understanding: can we understand it as my "direct-output rate"? That is, from what I want to what I finally get, the part that can be used directly—what range is this value in now?
Zhou Ze'an: Actually, overall we looked at it, and the direct-output rate should only be around 50%. This is a very real number. Maybe for many creative categories like posters and videos it's already 80-90%, but in the document industry 50% is already very high, because documents often serve oneself and need to incorporate one's own things—most of what AI produces can only be called reference and needs activation, so having 50% is already very high.
Li Shouguo: My feeling is that the north-star metric doesn't necessarily come from our Agent and software itself—it may more come from the client's own feelings. For some service-communication roles, the core focus is pass rate. Because the financial industry has strict industry certifications or internal institutional certification systems, with strict position pass rates. For these types of roles, the core metric is whether, after training, they can pass the internal test or exam. But for mid-to-senior sales roles, especially like wealth management advisors, the focus is more on cost. That is, compared with human teachers, to achieve the same sales skill level, how much has the cost dropped? Originally using a human teacher might cost 20,000 or 30,000 yuan a day; today to achieve the same training level, what is the software cost roughly? I think the business result you care about gives birth to the north-star metric—it actually doesn't come from the software itself.
Howard: I think I'd look at it from two dimensions. The first dimension is from the individual—like I myself have several Agents or Bots on hand, and in a few days I can build out a system that natively might take one to two months. In this process I participate throughout—although AI can replicate execution, my attention is in it, I'm responsible for discussing solutions with it, deciding direction, giving standards, and acceptance. So from an individual perspective, the metric in the current AI era should be "human output per unit of attention." Previously we might talk about human efficiency; now for the individual it's actually output per unit of attention. If we go back to the organization, it might be "throughput." An organization with the same number or fewer people—how many product lines can you simultaneously manage, how much output can you deliver. Now there are also many domestic companies using Token count metrics to evaluate employee performance, but later I heard this evaluation method also failed—many employees purely game Tokens, meaningless and wasting money. More likely it's still about looking at final output.
Dr. Chris Yang: I'll also share my own views from internal and external perspectives. Internally, actually the biggest north-star change is: previously as company management I always communicated most with the core management, but in the past two years I've actually communicated most with the students I mentor. Because I still write code and do algorithms, so I feel this is a big change—I don't want past experience to hinder the company's organizational development.
From the client side, the biggest change I feel is the same—letting many major clients shift from experience-driven to data-driven. Most of what we serve are physical industries, so they're certainly more conservative. For example, in the FMCG industry, no one has seen a new FMCG brand able to build up offline in three to five years—they all start at 50 or 100 years. The retail industry is the same—to open a thousand stores, it can't be as easy as two days online. Previously almost all their processes relied on experience-based decisions.
Let me give two specific examples. First, helping FMCG brands do offline marketing budgetad placement management. Each partner FMCG brand has a budget of around 1 billion yuan. Previously the experience was "I think this place's marketing is good so I'll invest more"—very head-based. But now we evaluate every marketing phenomenon, every effect of everything done, through AI, and even market insight through AI perception, finally using data to tell them what's good and what's bad. Then when they invest budget next time, they're very clear that every penny goes where it should. Previously most of the budget went to local agencies—today you give 10 million, tomorrow he buys a BMW 7 Series. This is an example of shifting from experience-relationship-driven to data-plus-AI-driven.
The second example is: when a brand has 1,000 stores, the hardest role to find is a store manager. A well-managed store manager makes money; a poorly managed one not only can't make money but also brings many negative effects. Now among the brands we partner with, the store manager's decision-making ability for the store has been reduced to the minimum. We've delivered a so-called AI store manager to each store—it has 100 pairs of eyes, can simultaneously know what to do next when seeing this situation, who's doing well or poorly, and finally form decisions. This is like our OpenClaw entering the physical world, but much more difficult. Because all of OpenClaw's input-output and intermediate decisions are completed in the digital world, while we must complete them in the physical world, and with human participation. Humans have uncertainty and individual differences, and only then can business objectives be achieved. In summary: internal organizations should learn from fresh blood; externally, shift from experience or relationship-driven to a data-plus-AI-driven model.
Xue Qian Amber: I feel the challenge of evaluating brand marketing effectiveness on your side is very large. Because in the past, company marketing departments were often challenged by the boss—spending so much money, how do you evaluate the effect? Now if using AI, how do you make it data-driven?
Dr. Chris Yang: This is a very good question. Looking around the world, we're probably the only company that has implemented this very solidly, which is also why we've survived and done well after so many years in AI. For FMCG, the most core thing is shelf display. You walk into a store to buy mineral water—you don't care that much whether it's Nongfu Spring or C'estbon, you just grab the one in the best position within your line of sight. All brands spend large marketing budgets to do this. But the problem is whether the agency can execute the SOP requirements—probably after execution you don't know what the situation is. Previously you could only rely on people visiting stores one by one. At the beginning we were also a tool—I don't trust the results you give me, I only trust the videos and photos you take; if it's not done well, AI will tell you in one second that you should adjust before going to the next store. This is typical human-machine collaboration. Now we've gone deeper—with a large amount of FMCG brand and store information, we directly send marketing tasks to this store, because China has a large number of mom-and-pop stores that help us execute. Previously money was given to them upfront; now it's post-payment: after each marketing action is completed and AI verification passes, I then distribute the budget. I guarantee the effect is audited by AI before the budget is sent. But this promotion is a bit difficult—going through the Marketing department can't get it done, they all oppose this tool; we can only talk to the boss, and the boss will probably like our product very much.
Xue Qian Amber: Then I'll also take the opportunity of Chris's talk about thepredicament of offline execution to ask Chris and Beta Data's Mr. Li. Because both of you serve many big-B clients, when really using AI to implement into business, you'll definitely encounter difficulties. Many enterprises say they want to implement AI, but the path from 0 to 0.8 is very fancy, and when it reaches 0.8 to 1, difficulties pile up—whether stuck by data, by internal organizational inertia, or by permissions—anyway it's not so smooth to finish. Among the clients you each serve, how do you help everyone truly reach the last step? Let me start with Mr. Li.
Li Shouguo: First, big-B or the super-large-B central enterprises and state-owned enterprises we work with are definitely not the best customer group for validating AI products. But there's no way—having chosen this most difficult path, due to path dependence we still have to bite the bullet and do it; their organizations are too complex. If under the unavoidable big premise you have to force it, the path we choose is: first, the business result that the AI tool ultimately achieves should preferably be decidable by the one department paying the bill. The decision must be relatively singular—if you want a project to fail, the best way is to drag several more departments into meetings together. Second, in the original business process, it should preferably be a business result that requires cross-team, multi-role collaboration to accomplish. Originally, cross-role collaboration actually had very large friction and was very painful. They must have pain, but also want this business to have results. So your transaction decision should be simple, and the friction in originally achieving the result should be relatively large—in such fields the push is relatively much faster. We launched a product for large-ticket insurance sales in early April, targeting financial institutions, and in less than two months now, we predict this month we can almost achieve profitability—the speed is very fast, precisely because it fits this point. This is a small shortcut summarized after stepping into many pitfalls.
Dr. Chris Yang: Very, very good experience. If Mr. Li and I were to pour out our grievances, it might take a whole day. What I want to tell everyone is: although doing big-B is difficult, once you get in, the future barriers will also be very high, and it's hard for others to replace you. If there are entrepreneurs around, I encourage everyone—some difficult things, when done, build very high barriers. In this process, for example, I myself previously didn't like teaching at university, but when I first started doing AI I was forced to give lectures at central enterprises and state-owned enterprises—they like to invite experts. You're forced to do things you don't like, and what you ultimately discern is: you first have to solve the To B pain point, even a very small pain point, and you have the opportunity to cut in as a new supplier. After biting off a piece of the cake, you have the opportunity to keep expanding the cake. In this process, you have to find such a Coach within the B-end client, who then in turn leads our team. He actually doesn't care that much whether you're an AI company—he'll tell you what he truly wants to do. It's very critical that in this company you must find a Coach to polish the product with you, ultimately solving its business problem, rather than you thinking what needs he has and then selling.
Xue Qian Amber: Yes, I'm still very curious—offline retail execution is such hard, tiring work, having to take care of so many mom-and-pop stores and get them to cooperate. What's the internal structure roughly like in your company, can you help us understand roughly how much energy you spend solving the problems of so many offline retail stores during implementation?
Dr. Chris Yang: Actually, in our company there's a very important role called project manager, morebias a product manager role. This is also our company's most important asset. For any client, as long as the payment ability they bring is strong enough, we basically assign a dedicated project manager. His role, besides delivering the product, is to hope he stays with the client's Coach every day. This Coach is very critical—he's not only someone willing to help you, he must be someone very familiar with the industry and business processes, willing to hand over years of accumulated experience to you, and then you think about whether you can reshape all processes through AI. The project manager role is very critical, but it's hard to recruit in the market—most need to be cultivated by ourselves. He must both understand the industry and at the same time have very strong sensitivity to the boundaries of AI technology. If you let a pure AI engineer do it, he definitely can't; if you let a pure salesperson do it, his understanding of AI isn't that deep. At the same time, this Coach will also pick people—if he feels communicating with you has no value, he won't talk to you again. So the positioning of this role is very, very critical.
Xue Qian Amber: Alright, thank you Chris. Then let's turn back to our office side. I'd like to ask Mr. Zhou—you also mentioned there are too many intelligent office software now, over 300 you can name, and they may be updating and iterating every day. I myself also like using various AI office software. From your observation now, with ChatPPT's user base already very large, what do users who are truly willing to pay and continuously use the product care about most?
Zhou Ze'an: Actually, in our scenario, delivering results is the most important. Regardless of efficiency, large models, whether general or professional documents, results come first—what users care about most is actually the best effect. To achieve good results, it may actually be a mindset issue. If you have sufficient engineering capabilities to adapt to the scenario and let AI produce good effects, that's a good approach. If it's just general large-model capability, why should users choose you? First point, you must adapt to the current technical paradigm and framework for rapid response—general model capabilities are already OK, and you need to use more Skills to refine your things into output in this scenario. Second point, you must strengthen and refine the functional points in your scenario. Many people say it's exceeding expectations, but actually it's more called "not disappointing"—letting him not be disappointed at the moment of use is OK.
For example, our ChatPPT has a category of users who particularly like using it for stage presentations. After writing the PPT, what's the hardest part? Not being able to deliver it. To amplify the value of the document itself, we did two things. The first is we made a presentation-assisting Coach that can help you record rehearsals, directly clone your voice to present, collect client feedback and then revise. The second scenario we refined to—starting the year before last, collaborating with RayNeo on the teleprompter in AI glasses. Previously it might have been simple translation, but now in RayNeo's X3 it's basically fully adapted—stage presentations don't need a page turner, it can automatically scroll words according to your rhythm; even based on vision, a gesture can zoom the page in or out. This actually solves delivery, amplifying the capability in the scenario itself. Third point, you must do a good job of data closed-loop feedback. The best insight is what users naturally produce. So the advantage we advocate for being as Live as possible is not only convenient for sharing, but also getting feedback, and quietly using last time's feedback to produce better effects.
Xue Qian Amber: Next I'd like to communicate with Howard again—roughly how many people does your company have now, and how many Agents?
Howard: Our company currently has a little over ten people, and probably 30+ Agents.
Xue Qian Amber: That's two Agents per person, right?
Howard: After our employees join, the company assigns an Agent to them, paired with the world's best model—the most expensive one. Some individuals on our team may have several Agents.
Xue Qian Amber: Then what do you feel is the biggest difference between using people in the past and using Agents now—the biggest difference between managing 30 people and managing 30 Agents?
Howard: What we're building is an AI Native collaboration platform, and at the same time we also sell digital employees. The concept of digital employees was proposed two or three years ago, but it only truly landed this year after Claude 3 Opus came out—because the model is powerful enough that you have 90% confidence to hand a problem to it and it can complete it. Previous models basically had to be corrected midway. Back to how to manage dozens of Agents—it's actually similar to managing people, because it's already very much like a person, and the methodology of managing people in the past applies to Agents the same way. Where does it differ? When managing people, you care about their psychology, motivation—you have to coax them a bit; but Agents don't—you give them the work and they go do it. But for Agents, maybe we need to care more about context—they need to know information, you need to set standards for them and do acceptance. This is where it differs a bit from collaborating with people.
Xue Qian Amber: I previously saw in your product introduction that you have roughly a hundred different role functions—are these part of your digital employee capabilities?
Howard: It's equivalent to some preset Agents. We deposit the experience of managing people into Skills or role settings. Now AI's capability is sufficient to well follow planned SOPs for execution.
Xue Qian Amber: Right, this is actually the question mark we raised when seeing "can give you a digital employee": maybe some surface-level work can be done well, but if used as a senior employee with five to ten years of experience, there's a lot of tacit knowledge (Know-how) behind it. From the perspective of your framework building, how do you solve this?
Howard: I might be able to tell a small story—my own Agent can now make some decisions very similar to what I would make. How does it do it? It actually distills my way of thinking, with about 35 principle rules inside. This way of thinking is very tacit knowledge—I myself can't even realize there are 35 rules, and can't write them down for it. How to put it on the Agent? It's because in the daily work environment group, when I make many decisions I'll say what's bad and what's good, it forms Cases, reviews every day, and then deposits them with a certain theory. So how to distill tacit knowledge? It should be letting AI extract it in daily collaboration, rather than you writing rules yourself—it can't be written.
Xue Qian Amber: I have one last quick question here. I'd like to hear everyone give a brief one-sentence conclusion: currently, regarding your product's AI capabilities, which capability do you feel it should have, but you don't quite dare to release yet? Let me start with Mr. Zhou.
Zhou Ze'an: I think it's in some targeted delivery stages. Now most are general documents, and some concrete, specific-solution outputs after chatting may not be able to be handed to AI—behind this is one issue of responsibility and one of expectation.
Xue Qian Amber: Targeted delivery.
Li Shouguo: If I have to limit it like that, it might be compliance. Actually, AI's capability to do financial review is mostly in place, but because someone needs to "take the blame," financial institutions always require the final link to be a person. This cost can't be brought down, but the capability has actually arrived.
Howard: The part that can't surpass humans is the "taking the blame" capability—I also agree with what Mr. Li said about the responsibility issue. So like us, for every Agent we set a final responsible person who is that person.
Dr. Chris Yang: Because we've experienced particularly many physical industries, what I most want is to directly hand this store over to me to operate, and when we make money we split it—this is what I most want to do. Because one client has thousands of stores, and assuming I can help each store make an extra 20,000 yuan, it's OK if he gives me 1 yuan—this is the feature I most want to push.