Bairong Intelligence 2025 Annual Earnings Presentation
[AI Key Points Summary]
Financial Performance
- Full-year revenue for 2025 reached 2.92 billion yuan, roughly flat compared to the same period last year, with gross profit at 2.102 billion yuan and gross margin maintained at 72%
- Operating profit was 60.38 million yuan, net profit was 73.87 million yuan, operating profit margin was 2%, and net profit margin was 3%
- Under non-IFRS, net profit was 179 million yuan with a net profit margin of 6%, EBITDA was 244 million yuan, and EBITDA margin was 8%
- MaaS business revenue increased by 9% year-on-year to 1.019 billion yuan, while BaaS business revenue decreased by 5% year-on-year to 1.901 billion yuan
Business Progress
- Served over 8,000 institutional clients, with a core client retention rate of 98%, the highest in history
- Built a silicon-based workforce system covering both CX and EX categories, achieving commercial breakthroughs across hundreds of industries
- Established a joint AI laboratory with Gaoling School of Artificial Intelligence at Renmin University of China
- Accumulated 573 patents and software copyrights, with the Craftsman AI operating system and BRLM large model successfully registered with the National Cyberspace Administration
Guidance for next quarter’s performance
- 2026 is a pivotal year for the AI strategy as it transitions from capability building to value realization
- Will focus on two key areas: contact centers and professional services
- Continue to invest in core technologies and computing power, emphasizing improved marketing efficiency and control of administrative expenses
- Achieve scaled profitability through mature scenarios and forward-looking deployment via innovative scenarios, driving growth through a dual approach
Opportunity
- The contact center market size reaches trillions, with thirty million carbon-based seats across China offering significant replacement potential
- Expand from financial insurance into various industries, developing enterprise-level AI operating systems to build a silicon-based workforce ecosystem
- Co-invest and incubate AI ventures with top-tier capital to build an industry ecosystem
- Use AI technology to significantly reduce labor costs and improve operational efficiency
Risk
- Financial regulatory policies such as Document No. 9 have impacted BaaS business revenue to varying degrees
- Major tech companies launch general-purpose office AI agents, lowering technical barriers and intensifying market competition
- Price hikes on AI cloud platforms impact operating costs
[AI Conference Record]
Sandy
Dear investors, analysts, and media friends, greetings to all. I am Sandy Qin, Director of Investor Relations at Berrong. We sincerely thank all the online investors for their attention to Berrong. The company has released its 2025 annual earnings report yesterday after the market close on March 26.
In this earnings call, we will present and share the company's operating results for 2025 and provide an outlook for 2026’s performance, as well as answer questions of interest to you. This earnings call is divided into three parts: the first part features the CEO discussing business progress and outlook, the second part involves the CSO reviewing financial performance, followed by a Q&A session.
Online participants can submit questions in the question bar of the live stream interface or ask questions by dialing in via phone. For mainland China, dial 023627371234008063263; for Hong Kong, dial 85230183602 or 800961505; for international calls, dial 862362737100. The Chinese participation code is 595868414, and the English participation code is 291375421.
After the earnings presentation, management will address the questions. This earnings call contains forward-looking statements that reflect the company’s current beliefs and expectations about the future. These statements include terms such as expectations, believe, intend, estimate, anticipate, and similar expressions. All statements other than statements of historical fact contained in this conference are forward-looking statements.
These forward-looking statements only reflect the views of the company’s management as of the date of this document and are not guarantees of future performance. No member of the company group or any other relevant party, including their directors, senior management, employees, consultants, or representatives, assumes any obligation, and explicitly disclaims any obligation or commitment to disseminate any updates or revisions to any forward-looking statements.
Next, allow me to introduce the company’s management team attending this earnings communication meeting. They are Mr. Zhang Shaofeng, Founder, Chairman of the Board, and Chief Executive Officer of Berrong; Mr. Kelson Chen, Executive Director and Chief Strategy Officer of Berrong; and Mr. Duan Ying, Partner and Senior Vice President of Berrong. Now, let us invite Mr. Zhang Shaofeng, Founder, Chairman of the Board, and CEO of the company, to introduce the business progress in 2025. Thank you.
Zhang Shaofeng
Thank you, Sandy. Dear shareholders, investors, analysts, and media friends, good morning to all of you. Welcome to Bairong's 2025 earnings release conference. It is a great honor to report to you on Bairong's performance progress and new growth, and to share with you some of our thoughts on the long-term development of the agentic AI era.
The first part is about the company introduction. First, let me briefly introduce Bairong. We are a leader in the enterprise-grade intelligent agent industry, and we are the inventors of the currently very popular results-as-a-service (RaaS) business model both domestically and internationally. We were also the first in the world to formally propose such a business model, making us a leading AI enterprise.
Through the RaaS business model, we deliver silicon-based employees or silicon-based assistants to our enterprise and individual customers. Our positioning is to design, develop, dispatch, and train various types of silicon-based employees or assistants for enterprises and individuals.
Relying on our result cloud, which is oriented toward enterprise outcomes, we have deeply integrated our self-developed BRLM Bairong large model and our Baigong enterprise-grade intelligent agent operating system (Agent OS) to help our clients build their silicon-based workforce systems.
We charge through two models: one is job outsourcing, where we dispatch one, two, three, or four similar positions to our clients and charge based on the position’s salary. The other is full business process outsourcing, where the entire business process is handed over to us, akin to outsourcing an entire department to Bairong, and we charge based on the overall end-to-end results of that department.
Our end-to-end delivery focuses on end-to-end results or the outcome of a specific position. We do not charge for delivering or licensing tool software. Since its inception, Bairong has been a company focused on delivering results.
As of today, we serve more than 8,000 institutional clients. Our core client retention rate has reached 98%, maintaining a consistently high level. Our representative clients span multiple industries such as e-commerce, internet, education, telecommunications, finance, healthcare, and so on. Our solutions are widely applied in scenarios like marketing recommendations, customer operations, intelligent customer service, contract and ticket parsing, credit authorization, anti-fraud, claims processing, archival management, recruitment, training, and more.
The second part provides an overview of our performance. As everyone knows, 2025 still presents numerous macroeconomic challenges. However, Bairong continues to deepen our AI strategic upgrade and achieve steady growth by relying on our solid product technology, deep understanding of the business, and years of strong customer relationships.
Firstly, regarding our revenue structure, our cornerstone MaaS (Model as a Service) business remains robust, with year-over-year growth recovering to 9%, surpassing last year—or rather, 2024—when it was at 5%. Everyone understands the environmental factors from last year, including policy impacts and Document No. 9. Despite these significant challenges, achieving 9% growth fully demonstrates the exceptional competitiveness of our business and products.
Our revenue exceeded 1 billion RMB while realizing dual growth in both our customer count and average revenue per customer. Additionally, the retention rate of our MaaS core clients further increased to 98%, setting a historical record. This retention level is absolutely among the most leading in the B2B industry, reflecting our clients' high recognition of and continued reliance on our AI capabilities.
In terms of our operations and services, including the BaaS business, as we communicated multiple times during last year's interim report, policies like Document No. 9 on compliance in reporting have indeed impacted revenue for financial industry clouds and insurance industry clouds to varying degrees. However, it is worth noting that under such a macro environment, the total premium transactions facilitated by our company still grew by 20% last year.
Moreover, the most important metric, first-year premiums, increased by 25%. The persistency rate of life insurance premiums remained stable at over 90%, placing us at a very leading position across the entire industry. Our revenue pressure is mainly due to regulatory impacts, specifically the reduction in commission service fee rates.
What do these figures indicate? They suggest that although short-term regulatory pressures exist, our own business competitiveness continues to rise, meaning our relative market position is improving. Once the macro environment recovers, we should experience even higher growth rates because such an environment will cause many competitors to gradually fade away. Our ultimate goal remains the continuous increase of market share.
I understand that everyone is most concerned about our group’s overall revenue and profit performance last year, especially given the significant pressure on profits. Let me provide a detailed explanation. At the beginning of last year, we also explained that our focus for 2025 was to significantly enhance R&D and business investment, even at the cost of aggressively capturing market share. This was also stated verbatim during our annual report meeting last year.
If we were to cut our AI-related investments to zero last year, our mature businesses would actually demonstrate very strong profitability. These upfront AI-related investments include several areas which I will outline below.
The first area is the reserve of AI talent. Last year, we invested substantial resources in recruiting AI professionals, reinforcing our foundational AI algorithms, engineering teams, and product development teams to build a more robust AI-native organizational capability. If you follow our annual report, you can see that our R&D staff increased by 383 colleagues year-over-year, bringing the total number of R&D colleagues to over 1,100, accounting for 64% of the company's total workforce. This is a very high proportion and lays a solid talent foundation for our long-term AI product R&D.
The second area is spending on purchasing exclusive private domain data to enhance our competitive edge in vertical-specific proprietary data, strengthen our model training and corpus, and consolidate our data moat. The third type of expenditure is on computing power infrastructure, including expanding our server rooms, purchasing GPUs, adapting our domestic graphics card hardware and software to train our own platforms, and reducing our long-term computing costs.
The fourth category of spending relates to operational capabilities associated with AI, including expansions of our office spaces and personnel increases that have led to related peripheral costs. The aim is to ensure that our organization maintains efficient collaboration, particularly in administrative functions, during this rapid expansion phase.
By comparison, our upfront investments in AI have been relatively cost-effective. Despite macro pressures last year, our gross margin remained above 72%, a high level. We maintained profitability even after these heavy investments, placing us in a top-tier position among current AI or intelligent agent companies. Moreover, these investments are crucial and urgent for a company committed to becoming a great AI enterprise in the long run.
Some of you may have noticed that before the Spring Festival, a US research institute called Century Research published an article titled '2028 Global Intelligence Crisis,' a lengthy piece of about ten thousand words. It analyzed that within two years, intelligent agents, or silicon-based employees, will become widely prevalent. The dawn is near, and time waits for no one. We must make early investments to gain a first-mover advantage. If we hadn't invested last year, this year we would only be spectators, especially before the Spring Festival, watching others discuss Open Core while we ourselves would not be able to participate.
The third aspect is a report on the progress of our business. We have high hopes for some of our businesses, and we believe that some of our friends have great expectations for our new ventures, which I fully understand. As friends and investors, there is an urgent desire to see our investments quickly translate into financial revenue and profit.
Here, I will offer a perspective on this approach. In the early stages of AI agent development, or intelligent agents in the B2B market, there are many opportunities to generate substantial short-term revenue. What kind of short-term revenue? Essentially, it involves project-based development for enterprises, especially large companies, similar to what previous-generation software companies did by renting out their manpower—sending their employees or developers to embed within companies to help them with custom projects. Each contract might seem large.
This method can quickly boost short-term revenue. However, we already know that China’s traditional software industry ultimately ended up with a very poor business model. If we were to take on such demands, we could break through hundreds of millions in revenue just from this segment alone this year, making the revenue look impressive in the short term. But that's not a sustainable business model. What we aim to build is a long-term, recurring revenue model with healthy gross margins—a Results as a Service (RaaS) business model.
In other words, we will not sacrifice our long-term competitiveness and ruin our business model for the sake of pursuing short-term project-based revenue. From the very beginning, our company has not followed such a model. If we had pursued project-based work in the early days, we wouldn't have achieved the sustained and steady revenue and profits we enjoy today.
Under this model, we are unwilling to undertake customized projects for enterprises. Compared to project-based firms, our revenue may grow relatively slower. Some listed companies on the market essentially operate on a project basis, and while our short-term revenue growth may not match theirs, our revenue is of high quality. It is sustainable, not one-off, and our gross margins are relatively healthy. Once we get through the difficult early stages, we will see the light at the end of the tunnel. We believe that both our revenue and profits will experience long-term, healthy growth.
Therefore, instead of focusing on short-term revenue, we will try to provide more trackable leading indicators. Jensen Huang often talks about this too—he says he doesn't focus on revenue but rather on internal indicators, forward-looking metrics. For example, how much efficiency has our product helped partner companies improve? How much have we reduced their labor costs? Have we improved customer satisfaction? Have we lowered employee turnover rates? How many high-quality clients have we acquired? How many silicon-based employees have we continuously deployed? Such leading indicators are much healthier than short-term project-based mistakes. We won’t distort our behavior for short-term gains.
On the product side last year, we established two main categories of silicon-based employee systems. The first is Customer Experience (CX), and the second is Employee Experience (EX). These are internationally recognized classifications: one directly helps companies enhance their interactions with customers, improving customer experience, and the other helps companies boost the productivity of their internal employees.
For CX-oriented silicon-based employees, such as our silicon-based marketing specialists called Baiying, they have been integrated into over a thousand different scenarios, including telecoms, retail, and wealth management, helping our partner companies save up to 90% of their costs. Silicon-based employees have no turnover rate, whereas carbon-based employees, like us humans, often experience turnover rates as high as 70%. Silicon-based employees never leave unless you unplug their power.
Consultation conversion rates—for example, inquiries leading to sales revenue purchases—have increased by around 200%. Customer satisfaction among our enterprise clients' customers has risen from 16% to 55%, thanks to our silicon-based customer service specialists.
Another category of our silicon-based employees is called deep analysts. Many professionals involved in investment or strategic design might be testing our Baizhi silicon-based analyst. By connecting multiple types of sub-agents, it assists professionals in quickly completing the entire process loop of listening, recording, writing, and generating comprehensive reports. Previously, this task might take two analysts ten days or one analyst twenty days, but now it can be compressed to just one day.
The employee experience, or EX, is supported by an expert service platform in legal, tax, and financial matters called Baijian. It handles standardized tasks such as data collection, case retrieval, document drafting, and process management. Relying on a private knowledge base of up to 100,000 entries, we reduce the annual cost of high-end consulting services for clients—such as legal advice and overseas location consulting—from five million to one million.
Our silicon-based recruitment specialist, Baicai, helps companies automate hiring through AI-driven interviews and resume screening. This reduces the recruitment cycle from 28 days to just two days and increases job-matching rates from 60% to 90%. A single HR professional who previously recruited five people per month can now recruit twenty people monthly.
Additionally, within Bairong Company, our silicon-based workforce's HR system—called the Silicon Employee Home—is built on our Agent OS, which operates as an intelligent body operating system. This Silicon Employee Home manages over 200 types of silicon-based positions across various administrative functions like finance and legal affairs. Each silicon-based role corresponds to real-world carbon-based employee positions, significantly reducing workload while increasing efficiency. These silicon employees are utilized across multiple departments, including legal, finance, HR, customer operations, and more.
We have also started directly serving our corporate clients, including new ventures, using silicon-based employees. In fact, our silicon-based staff currently serve over 2,500 enterprise clients throughout the entire process—including client onboarding, contract reviews, renewals, upselling, and cross-selling different products—all handled end-to-end.
This has allowed our customer operations department to reduce its carbon-based workforce servicing small clients from 50 employees to just five human employees supplemented by 18 categories of silicon-based employees serving 2,500 corporate clients.
On the client-facing side, our silicon-based employees have achieved commercial breakthroughs across various industries, validating the versatility and adaptability of our AI agents. While we initially focused heavily on banking credit sectors, today, leading securities firms like CITIC JianTou are adopting silicon-based employees provided by Bairong. Some mid-sized brokers have begun utilizing our silicon-based financial advisors and client onboarding specialists.
Banks have also started procuring our silicon-based call center agents. For about 5,000 yuan per month, these virtual agents help banks drastically cut labor costs. Additionally, some well-known trust companies have adopted our silicon-based agents for customer service, handling complaints and resolving issues efficiently.
Some renowned international investment banks have begun employing Bairong’s silicon-based employees for phone services and WeChat Enterprise interactions. These AI-powered agents interact with clients naturally, recommend products, and successfully guide them through transactions. Notably, several large state-owned banks have signed artificial intelligence expert service agreements with Bairong, marking the first major professional service contracts with top-tier institutions.
In the telecommunications sector, we are deeply involved in the digital transformation of the three major operators, systematically verifying and implementing our silicon-based workforce in provinces like Guangdong, Shanghai, and Sichuan. In Guangdong alone, our AI systems facilitate thousands of telecom package sales daily for certain carriers without human intervention.
In healthcare, we partnered with a prestigious medical university to provide an AI-driven mental health counseling agent serving their students and faculty. Even those studying medicine face psychological challenges that require support. Additionally, we collaborated with a leading medical technology company to develop an AI solution for chronic disease management and genetic testing reports, representing another breakthrough in this vertical field.
In the life services sector, our platform, BaiCai Recruitment, provides AI-driven recruitment solutions to multiple large-scale platforms employing tens of thousands or even hundreds of thousands of workers. This solution focuses on fully automated AI-powered recruitment for blue-collar workers. In high-end manufacturing and new retail scenarios, we partner with emerging electric vehicle companies like NIO, XPeng, and Li Auto, assisting in lead generation, sales follow-ups, and more.
We have also implemented a contract review specialist named 'BaiCheck' in a premium postpartum care center in South China, marking the first deployment in the maternal and infant industry. This system helps them draft, modify, and review contracts. Additionally, we collaborate with a provincial environmental policy research institute to create silicon-based employees capable of monitoring, reviewing, and approving dual-carbon projects. This validates that BaiRong's AI capabilities can generalize from financial insurance to broader and more complex industries, as the foundational abilities remain consistent.
Our B2B services cater to high-end clients with stringent requirements, especially in the financial sector, which has exceptionally high standards. We find that once we enter other industries, about 99% of those sectors have lower requirements than the financial industry. These leading clients share one commonality: they are willing to pay for silicon-based employees whose ROI (return on investment) can be calculated. This willingness to pay far exceeds that for traditional tool-based software.
The fourth part I’d like to introduce is our R&D progress, which we prioritize highly. Over the past two years, BaiRong has invested heavily in attracting top-tier AI talent while significantly increasing our investments in data, computing power, and algorithms to strengthen our infrastructure. Our technology and accumulation have reached a critical mass, and we believe a breakthrough will occur starting this year.
To date, the company has accumulated 573 patents and software copyrights. In terms of industry-academia collaboration and cutting-edge technological reserves, we officially established an AI joint laboratory with the Gaoling School of Artificial Intelligence at Renmin University last year. Through three donation funds and six joint research initiatives, we have explored a seamless collaborative mechanism from scientific research to talent development and成果转化 (results transformation), promoting the deep engineering implementation of agent technologies in real-world scenarios.
Our primary focus with Renmin University’s AI school revolves around core advancements in agentic RAG, MOE (Mixture of Experts), and post-training techniques for hybrid expert models. The main goal is to tackle some of the most globally valued challenges, such as long-term, complex task automation. We are also collaborating with other universities and research institutions to address these difficult problems.
On the application side, we’ve made significant strides in transitioning agents from usable to user-friendly. For instance, our language and speech models have evolved from the traditional three-step process—converting audio signals to text, processing the text through a large model, and converting it back to audio—to an end-to-end large model. This means audio signals go directly into the model, and the output remains in audio form without intermediate text conversion.
This end-to-end model significantly enhances user experience in voice-based scenarios while reducing costs for our company. It currently supports over fifty different voice tones, including dialects such as Cantonese, Sichuanese, and Shanghainese, as well as foreign languages. Response times have been compressed from several seconds to under 200 milliseconds, sometimes as low as 40 milliseconds. Our accuracy rate is as high as 99%, with emotion recognition accuracy exceeding 95%. Recognition of interruptions by real consumers reaches 100%.
In addition to providing silicon-based employees, we manage their entire lifecycle, including onboarding, performance evaluations, retraining if needed, and redeployment. Our human resource management systems for silicon-based employees, akin to managing carbon-based employees, are becoming increasingly mature. The development cycle for silicon-based employees, which previously took two months, has now been reduced to an average of two weeks.
Silicon-based employees can not only perform self-assessments before deployment but also collect real-time feedback on their job performance after deployment. They can then autonomously improve and optimize themselves, enabling self-iteration. This represents a highly significant—and potentially disruptive—technological breakthrough.
In terms of foundational models, we have built our proprietary large-scale models tailored for various industries and specific subfields. Over the years, these models have undergone approximately six iterations, covering scenarios such as telecom package recommendations, e-commerce confinement packages, wealth product recommendation explanations, marketing, resume screening, credit card analysis, and dozens of other distinct use cases. Additionally, our recently discussed multimodal visual model has made breakthroughs from zero to one in areas like health insurance and complex document structures, which are known for their intricacy, paving a new path for commercialization.
In terms of cost efficiency, we restructured the cost that determines income ceilings. In the AI field, costs often dictate outcomes, as they are significantly higher than in traditional industries. We redefined this reality by developing our own training and inference platform, reducing latency on mainstream GPUs by 33%, cutting resource usage by 50%, and doubling throughput.
At the same time, our primary models have been adapted for domestic GPUs. By implementing a hybrid elastic architecture, we've ensured stable operations across multiple regions. Leveraging the BR vertex inference engine, which integrates computational intelligence scheduling with multi-level caching, our computing power utilization efficiency increased by over 30%, while our mixed throughput rose by 300%.
Fifthly, I'll briefly report on some accolades we received last year. Our AI capabilities are transitioning from industry recognition to defining industry standards. Internationally, we were successfully included in Morgan Stanley's top sixty AI firms in China and received comprehensive coverage from IDC across six core dimensions, including industry-specific large models, agent development platforms, marketing, HR, legal, finance, and other deep application scenarios.
Additionally, we have ranked in KPMG’s top fifty AI companies for two consecutive years, a significant endorsement that establishes our global leadership in AI-native applications. Nationally, both our intelligent agent development platform and our LLM large model passed the National Internet Information Office's registration process almost simultaneously. Furthermore, under the Intelligent Agent Development Platform of the China Academy of Information and Communications Technology (CAICT), we became the leader of the first group of intelligent agent ecosystem construction units. Last year, we also obtained our third-level certification, solidifying our full-stack AI R&D capabilities and establishing a compliant moat.
Within the industry, from being named an outstanding provider of intelligent agent services in China for 2025, to winning IDC's P2S Game award for best practices in banking and insurance intelligent agents, to being selected as a trusted data space resource partner, becoming an industry benchmark for credit risk prevention, winning the Best Financial Award from Xiamen International Bank, and earning the Annual AI Interaction Innovation Award from a financial media outlet, all of these achievements underscore our silicon-based workforce's strong commercial penetration across different vertical and horizontal markets.
Sixthly, let me report on our progress in ESG. As a leading example of new productive forces, we earned an A rating in the Wind ESG evaluation. Last year, our intelligent agent operating system empowered disabled employment by creating an end-to-end self-service solution for silicon-based employees for Chongqing’s Disabled Employment Guidance Center.
Inside the Chongqing Disabled Employment Guidance Center’s official account, named 'Job with Love,' we integrated our large silicon-based employee model to create H5-based intelligent Q&A, helping disabled individuals efficiently and conveniently seek job information and undergo career retraining. This addressed previous inefficiencies and difficulties faced by the disabled community in finding work, serving around 600,000 disabled people in Chongqing and building an intelligent and efficient employment channel.
Moreover, we deployed our silicon-based workforce into other employment support fields. For instance, we introduced our silicon-based employees to an employment consultation center in a western city to provide frequent services like employment advice, policy interpretation, and subsidy application assistance. We have established a standardized process encompassing intelligent outreach, employment factor collection, policy interpretation, and post-employment follow-ups. This project is already operational and serves as a replicable model that can be expanded to other cities.
The introduction of silicon-based employees transformed traditional, extensive employment services into refined, precision-based ones, advancing grassroots governance from labor-intensive tactics to intelligent scheduling. Each government service has become more efficient and smoother. Our system handles close to a thousand follow-up tasks daily, with an 80% connection rate, far surpassing traditional human-based service levels. Data collection is now more complete and standardized, providing a solid foundation for subsequent employment services.
We firmly believe that the best technology is not just about gazing at the stars, but also about being tangible, trustworthy, and understandable, truly serving every aspect of people's lives and helping them solve work-related pain points—a new technology that genuinely delivers results.
The seventh part aims to help everyone look ahead to the future development of Bairong. We will provide a simple outlook. The year 2026 marks a pivotal turning point for Bairong’s AI strategy, transitioning from the capability-building phase to the value-realization phase. With the increasing maturity of underlying AI technologies and the growing anticipation and acceleration of industry applications, we will achieve scaled profitability through mature scenarios while leveraging innovative scenarios for forward-looking strategic positioning. This dual approach ensures robust profit growth while laying a solid foundation for our development over the next three to five years, comprehensively meeting the intelligent transformation needs across various industries.
We have experience serving more than 8,500 large-scale rigorous institutions. In the coming years, we aim to broaden, deepen, and expand the expertise and data accumulated over the past decade in serving large institutions and vertical industries, converting value across different sectors, and establishing the company’s leadership position in the agent era.
In the new year, we will focus on several key areas, providing a brief report. First, regarding the industry—many might be concerned about the impact of Document No. 9 issued by the Financial Supervisory Commission last year. We can say that with the issuance of Document No. 9, most of the potential risks in the financial insurance sector have been addressed. In the foreseeable future, the likelihood of significant adjustments in the financial insurance industry is relatively low.
At the same time, within the financial insurance sector, there are numerous niche scenarios with untapped potential. For example, marketing specialists for low-interest credit loans—positions that have long been lacking among ordinary people—silicon-based wealth management consultants, and specialists handling non-performing assets who have seen countercyclical growth in recent years. Demand for these roles is rapidly increasing.
Additionally, starting from the second half of last year, we began expanding the experience accumulated over the past few years into various industries, laying a strong foundation for the future. Let me share some highly promising scenarios. What was the first large-scale application of AI globally to achieve product-market fit (PMF)? It was AI coding, where silicon-based employees replace carbon-based programmers to write code on a massive scale.
What is the second scenario? It is globally recognized as the contact center, using AI to handle customer service, conduct marketing, and confirm information via phone calls, emails, or messaging apps like WeChat. These various niche scenarios fall under what is collectively referred to as the contact center, or CC for short.
Contact centers were traditionally reliant on human agents. How many human agents are there in China? You can ask any large model, and you'll find approximately 30 million human agents, including bank customer service representatives numbering in the millions, debt collection agents in the low millions, identity verification personnel, and over 20 million specializing in various types of marketing, totaling around 30 million agents. What is the average cost per agent, including social security? The average annual cost per human agent is roughly RMB 100,000.
Even capturing 10% of such a vast market represents a trillion-dollar opportunity. Our focus is not on software fees but on the labor wage market.
Secondly, beyond contact centers, we will aggressively expand into professional services. What are professional services? A typical example would be hiring lawyers for legal consultations, engaging top-tier consulting firms like McKinsey, Bain, Boston Consulting Group, or Roland Berger, where a single consulting project could cost three to five million. There are also financial advisory services, auditor consultations, accounting and tax advisory services, all falling under the category of professional services.
These service sectors are particularly prone to being empowered by AI, because what is their output? Mostly individual documents, and today AI can write personal documents very efficiently. We are building a specialized AI-native legal, commercial, tax, and financial advisory platform for global enterprises aspiring to expand overseas. If you're looking to provide outbound services, you may need to seek foreign experts, which involves highly complex tax planning, logistics consulting, and more. Typically, a single consultation can only be done by large multinational consulting firms or accounting firms, with costs generally exceeding four to five million.
Globalization is a significant shared demand that all Chinese enterprises must face in the future, representing the main theme of the next decade. Bai Rong needs to seize this opportunity to help Chinese companies go global. We hope to realize a new type of legal, commercial, tax, and financial advisory service through AI, aiming to replace 70% of such work with AI-powered silicon-based employees. This platform is called Baijian, providing architectural design, site selection guidance, financial and tax services, and professional planning specifically for small and medium-sized enterprises going global. Currently, we have not found any competitors in the market with equivalent competitiveness dedicated to this direction.
Let me share some updates on our investments in technology and ecosystem strategies. First, we will open up and empower industries such as travel, home appliances, telecommunications, and aviation with our strongest voice interaction and intelligent decision-making capabilities. We firmly believe that natural language interfaces will replace graphical user interfaces to become the most popular interaction method of the next generation. Over the past few years, computers and mobile phones have relied on graphical screens, either using a mouse click or finger touch. The next generation will definitely not follow this model; it will involve speaking directly or typing at most, eliminating the need to search, click, or switch channels.
We are also developing our own intelligent distributors, independent agent developers, and system integrators as ecosystem partners to jointly drive this industry forward while sharing benefits. At the same time, we will leverage capital to establish an AI industry incubation fund to nurture AI-native enterprises and build a vibrant agentic AI industrial ecosystem.
We collaborate with top-tier capital shareholders such as Sequoia Capital, Hillhouse Capital, and Gaocheng Capital to co-invest and incubate upstream and downstream AI enterprises. You may soon hear exciting announcements about us working with our top shareholders to create new ventures. Why? We want to declare to the industry that Bai Rong is willing to share profits with ecosystem enterprises. Not only do we provide intelligent infrastructure, but we also offer capital support to attract traditional software companies onto our platform, allowing them to build intelligent agents using Bai Rong's silicon-based employee operating system. We then share revenues, sometimes taking 30%, other times 50%.
Let’s look at two of the most prominent AI-native enterprises in the US: OpenAI and Anthropic. Both companies operate on two fronts: organic business expansion through product development and sales, and collaboration with capital to invest in other enterprises. After investing, they hope these enterprises will utilize their large models and overall builders. Bai Rong has invested in, incubated, and acquired several companies, including BPO contact center enterprises, hoping to assist them in upgrading their carbon-to-silicon ratio, replacing carbon-based employees with silicon-based ones, reducing costs initially and scaling up subsequently.
We have also invested in and incubated outstanding enterprises in the wealth management sector, hoping to help them achieve AI upgrades for mutual benefit. Of course, there is also the possibility of full consolidation someday. In the future, Bai Rong will focus on its own technology R&D, product expansion, and capital synergy for growth, adopting a dual-pronged strategy akin to Tencent's approach over a decade ago.
Together with our ecosystem enterprises, we aim to serve through collaboration, indirectly helping the hundreds of industries and consumers they serve, thereby forming a robust ecosystem centered around Bai Rong. This ecosystem is crucial for Bai Rong's long-term development. Admittedly, this process is neither easy nor quick, but we are patient and committed to this path over three, five, or ten years, much like how NVIDIA spent nearly 20-25 years developing CUDA before seeing results. Today, CUDA represents a deeper moat than NVIDIA's hardware chips.
Finally, I would like to share our views on long-term trends with our investor shareholders and analyst friends. As everyone knows, the US has two leading AI-native ventures: the most prominent being OpenAI, followed by Anthropic. OpenAI focuses on the consumer side, aiming to replace Google, while Anthropic targets the enterprise side, seeking to enhance productivity for businesses.
Historically, OpenAI has been far ahead of Anthropic in terms of visibility and valuation. However, in the past six months, especially after the launch of Open Cloud, Anthropic has achieved tremendous success in the US enterprise agent market, showing clear signs of surpassing OpenAI in both revenue and valuation. Just before the Lunar New Year, several top US analysis firms began considering the birth of Open Cloud as a turning point for Anthropic to overtake OpenAI. Note, this is an extremely important inflection point.
What is the reason? What signals does this reflect? The industry generally believes that the technological innovations of the previous generation of the internet and mobile internet were more focused on the consumption side and the circulation side, rather than the supply side or the production side. This AI transformation, however, brings more advancements to the production and supply side. Therefore, the industry has started to recognize Anthropic's growth potential because it assists the supply side and the manufacturing side, meaning its growth potential might be greater than OpenAI’s.
If we look back at several major industrial revolutions in history—the steam engine revolution, the electrical revolution, and the information technology revolution—they all began with changes on the production and supply side, and only after a decade or two did they converge on the consumption and circulation side. This is why the current AGI transformation represents a breakthrough comparable in scale to the previous three generations of revolutions, whereas the transformations brought by the mobile internet and the internet could not reach the heights of those earlier revolutions—because they primarily targeted the consumption side. Every major revolution first begins with changes on the supply side.
Anthropic has already proven in the US that enterprise-grade intelligent agents have crossed the critical turning point from concept to implementation, from tools to productivity. Bai Rong Company aspires to be China’s Anthropic, and we have the capital, the DNA, and the resources. Since our company's inception, we have delivered results to enterprises, rather than just delivering tools.
Although for a long time our perspectives and models may not have become mainstream in the industry and haven’t been widely recognized, we have steadfastly held our ground for over a decade. We are no strangers to the enterprise-level market; we didn’t just jump on the bandwagon. Instead, we’ve been deeply involved for years, especially in some of the most challenging fields, earning recognition from them and from large-scale institutions. Moreover, we have accumulated replicable experiences and results across multiple industries.
Of course, we clearly see that China’s enterprise-level market is still somewhat weaker than that of the US. It has its own unique pace, longer decision-making chains, longer validation cycles, and companies aren’t as accustomed to paying for services. However, if your value is measurable, we have already proven that many companies are absolutely willing to pay. It’s just that our shareholders and analyst friends might need to give us some time to validate our assertions.
Regardless, this doesn’t change our long-term assessment since our inception in 2012: Chinese enterprises are reluctant to pay for tools but are willing to pay for quantifiable results. From day one, twelve years ago, we never engaged in traditional software businesses. All our operations have been result-driven and priced accordingly. Delivering results, rather than tools, is embedded in our DNA from the very beginning.
In 2025, I mentioned last year during our annual report that 2025 would be the starting year for intelligent agents. In 2025, we announced to the world for the first time the RaaS (Results as a Service) model. After our announcement, it sparked an overwhelming global acknowledgment of this transformative shift. On Twitter, Bai Rong’s product launch event on December 18, 2025, gained attention and was shared by four hundred thousand people, who recognized that Bai Rong is leading the transformation of enterprise-level business models.
The difference between delivering results and delivering tools may not be noticeable in the short term, but in three years, a massive transformative change will unfold, widening the competitive gap.
Another point worth noting is the significant difference between the tech ecosystems of China and the US. In China, the focus has traditionally been on B2C. Over the past decade, the largest companies have been consumer-facing. Meanwhile, in the US, there haven’t been any major B2C players since Facebook. In contrast, the US excels in B2B, while China’s B2B sector struggles significantly. How much do revenues from Chinese B2B enterprise software companies account for compared to their US counterparts? Only about 4%, which isn’t even on the same scale. They are neither large nor strong, facing numerous operational difficulties, with listed software companies losing billions.
However, this time, please note what we are witnessing: silicon-based employees and intelligent agents. In this era, the products, services, and pricing models offered by tech enterprises in both China and the US may converge historically. This is an epic convergence, presenting a monumental opportunity for China’s B2B tech firms to overcome their historical setbacks.
As for Bairong, although our overall scale is not very large, if we set aside the industries we serve, including those from the past—whether it’s finance, credit, e-commerce, or telecommunications—and delve deeper into the essence of Bairong's product offerings, what exactly is the nature of our products? It’s about end-to-end delivery of results and our pricing model, which involves either charging based on position wages or result-based transaction sharing. This helps us understand that what Bairong does, even in the US, is at the very forefront this year and is a pure form of Agentic economy, also known as the intelligent agent economy.
From day one of the company's founding, our traditional tool-based software revenue has been almost zero, accounting for nearly zero in our overall revenue mix. So, what is the essence of the product we offer? It is to amplify the cognitive labor capacity within the human carbon-based labor market—or as we like to put it, 'cognitive productivity.' This is the most advanced form of cognitive productivity.
Recently, Joe Tsai, chairman of Alibaba, made an announcement at Siemens’ RXT conference stating that the global white-collar labor market amounts to how much in wages annually? Fifty trillion in wages, and he claims this value will certainly be restructured by intelligent agents. Therefore, you can see that two weeks ago, when Alibaba launched their 'Wukong' project, they heavily discussed the concept of silicon-based employees. This concept was first introduced globally by Bairong Cloud Creation, or Bairong AI Company.
This strongly validates Bairong's strategic path: instead of charging for tool-based software, we directly tap into the labor market by collecting labor compensation—a market that is a hundred times larger than software, reaching up to trillions, a truly massive and formidable space. In terms of business essence, we are no different from Anthropic—a purebred AI-native enterprise engaged in designing, producing, training, and dispatching silicon-based employees. We are committed to becoming pioneers and leaders in this earth-shattering transformation, setting an example for China’s B2B technology enterprises.
What kind of silicon-based employees have the brightest future? The first type I mentioned is in programming, the second is in contact centers—for instance, marketing and customer service—and another category handles unstructured data such as documents, images, audio, video, report writing, consulting, etc., like professional services. The third type manages complex and flexible workflow orchestration, even dynamically generating processes. Why did Open Cloud recently catch so much attention? It’s because it’s the first time a system could automatically handle complex processes and run for days without human intervention.
We believe these scenarios represent not only the most promising ones in China but globally as well. However, there is one caveat: personal open-core solutions cannot be used in corporate settings. Meta has already encountered issues because the personal version of open core has many security and permission management problems. On the other hand, Bairong has arguably created the first large-scale, high-quality use case of silicon-based employees validated by top-tier corporate clients in China. We have real-world application scenarios.
According to forecasts by EqualOcean, the market size for AI intelligent agents in China will grow from over 50 billion yuan in 2023 to 3.3 trillion yuan by 2028, underscoring the immense scale of this market. It represents a workforce—a powerful tool for enhancing productivity. We are at a historic juncture with an enormous market potential, marking the beginning of a major wave in the broader AI era. While this window of opportunity emerged quickly, determining who will emerge as the champion may not take three years; it could become clear within just one or two years.
Just as over a decade ago, Zhang Yiming allowed ByteDance to seize the opportunity presented by personalized information distribution in the mobile internet era, propelling ByteDance to become the dominant player among Chinese tech companies, far outpacing competitors, Bairong is determined to fully capitalize on this critical leap forward. Within the wave of the silicon-based economy, we aim to solidify our leadership position and create long-term sustainable value for our shareholders.
The core management team and I are filled with unwavering determination to succeed. Since the company’s IPO, as the founder, I have never sold a single share. There have been rumors online about me selling stocks, but those transactions were automatic sales of shares granted through equity incentives to cover tax payments. In reality, I haven’t sold a single share. This reflects my confidence, along with that of other management colleagues, in the company's business model, technological roadmap, and societal vision.
That concludes my brief summary of Bairong Cloud Creation and Bairong AI Company’s 2025 annual performance and future outlook. Thank you once again for attending our 2025 annual meeting to discuss our financial results. Thank you all.
Sandy
Thank you, Chairman Zhang Shaofeng, for your enriching, profound, and inspiring presentation. Now, let's invite Kelson Chen, Executive Director and Chief Strategy Officer, to present the company’s performance highlights and financial key points for the fiscal year 2025.
Kelson Chen
Thank you, Sandy, shareholders, investors, analysts, and friends from the media. Allow me to present our performance for 2025. In 2025, the company strategically positioned itself ahead of the agentic AI opportunity, increasing strategic investments while maintaining stable operational performance and financial results. Annual revenue reached RMB 2.92 billion, remaining flat compared to the same period last year. Gross profit reached RMB 2.102 billion, with a stable gross margin of 72%, primarily benefiting from the scalable RaaS business model, further demonstrating the advantages of scaling.
Operating profit was RMB 60.38 million, net profit was RMB 73.87 million, with operating profit margin and net profit margin at 2% and 3%, respectively. Under non-IFRS standards, net profit reached RMB 179 million, with a net profit margin of 6%. EBITDA was RMB 244 million, with an EBITDA margin of 8%. Despite a challenging environment this year, Bai Rong successfully stabilized its revenue and maintained continuous profitability, showcasing resilience and market competitiveness.
In 2025, BaaS business revenue declined by 5% year-over-year to RMB 1.901 billion, accounting for 65% of total revenue. Within BaaS, financial industry cloud revenue slightly decreased by 3% year-over-year to RMB 1.371 billion, representing 72% of BaaS revenue and 47% of total revenue. The insurance industry cloud continued to navigate regulatory changes, with revenue declining by 10% year-over-year to RMB 530 million, accounting for 28% of BaaS and 18% of total revenue. Meanwhile, our cornerstone MaaS business revenue grew by 9% year-over-year to RMB 1.019 billion, accounting for 35% of total revenue. This revenue structure demonstrates a pattern where BaaS remains dominant, supported by steady growth in MaaS.
Our MaaS business assists institutions in intelligent decision-making through silicon-based employees across various roles. Of the RMB 1.019 billion in MaaS business revenue in 2025, core clients—those contributing annual revenue exceeding RMB 300,000—accounted for over 78% of revenue. The number of core users increased by 12 year-over-year to 223, with average revenue contribution per core user growing by 6% year-over-year to RMB 3.59 million. The MaaS business maintained steady growth, continuing to provide the company with sustained cash flow.
Our BaaS business is based on generative AI technology, assisting institutions in intelligent marketing and operations through proprietary silicon-based marketing specialists. This significantly improves operational efficiency in wealth management, insurance, and internet technology industries. The BaaS financial industry cloud mainly adopts a business process outsourcing (BPO) collaboration model, charging based on achieved business results and scale. We do not charge any fees before assisting institutional clients in generating revenue.
As the leading player in AIGC application implementation in 2025, our BaaS financial industry cloud revenue slightly decreased by 3% year-over-year to RMB 1.371 billion, primarily due to policy environment impacts. Our insurance industry cloud provides comprehensive customer insights through decision-based AI, accurately recommending insurance products, and utilizing a combination of silicon and carbon methods for high-value policy user operations.
In 2025, our BaaS insurance industry cloud revenue was RMB 530 million. Let's look at their revenue composition. New policies contributed RMB 450 million, down 8% from RMB 487 million last year. The life insurance premium continuation rate exceeded 90%, ranking among the top in the industry. Renewal contributions were RMB 80 million, down 19% from RMB 99 million last year. However, the total premium reached RMB 6.52 billion, increasing by 20% from RMB 5.442 billion last year, reflecting our resilience in value delivery and momentum for business recovery.
In 2025, our R&D expenditure reached 637 million, a 25% increase from 509 million in the same period last year, accounting for 22% of total revenue, up five percentage points year-over-year. The increased investment primarily went into building top-tier AI-native capabilities, including talent expansion, large-scale procurement of data assets, self-developed infrastructure, and operational expenses. Sales and marketing expenses amounted to 1.142 billion, representing 39% of total revenue, consistent with last year. General administrative expenses were 328 million, or 11% of total revenue, also flat compared to last year.
We will continue to invest in core technologies and computing power to ensure Berrong's long-term competitiveness. At the same time, we are focused on improving marketing efficiency and controlling management costs to ensure enhanced profitability and capital returns while pursuing growth.
In terms of cash and assets, as of 2025, our cash, cash equivalents, and similar financial assets totaled 3.377 billion, an 8% decrease from 3.657 billion at the end of last year, due to Berrong's improved capital efficiency. Of the 3.377 billion, 726 million is cash and cash equivalents, 1.776 billion is in large-denomination certificates of deposit, and the remainder consists of liquid financial assets.
The company is highly confident about the future and places great importance on shareholder returns. In 2025, the company repurchased a total of 9.6 million Class B shares from the open market, amounting to 90 million Hong Kong dollars. As one of the few domestic companies that has found true application scenarios for AI, particularly agentic AI, and remains profitable, Berrong maintained stable revenue and continuous profitability in 2025. Moving forward, Berrong will continue to seize strategic opportunities, focus on advancing AI, and aim to empower various industries. Thank you all.
Sandy
Thank you, Kelvin, and thank you to the management team for the excellent presentation. We will now move to the Q&A session. All participants are welcome to ask questions. Investors and analysts joining via telephone who wish to ask questions should press star followed by one on your keypad. For those attending online, you can submit your questions in text form through the Q&A section on the Roadshow platform, and I will read them out later. Please state your name and organization before asking your question. My assistant will now begin the Q&A session.
Jiang Dan
Hello, Mr. Shao Feng, and hello to all members of the management team. This is Jiang Dan from Zhongtai Computer. First of all, thank you very much for giving me the opportunity to ask a question. My question mainly relates to the strong development of Berrong’s intelligent agent platform, which seems to align with what Mr. Shao Feng mentioned earlier—similar to Anthropic and Open Cloud—in its ability to answer questions and complete tasks with minimal user intervention, making decisions and taking actions accordingly.
Overall, we believe that the technological threshold for building AI agents will continue to decrease in the future, with deployment becoming even more convenient. Against this backdrop, I would like to ask: does the company believe it can achieve its vision of commercializing silicon-based employees at scale as planned? Additionally, what aspects represent the company’s overall differentiated competitive advantages? Thank you.
Zhang Shaofeng
Thanks to Jiang Dan from Zhongtai Computer for this question, which will be answered by CEO Zhang Shaofeng. Thank you, Dan, for this excellent and crucial question. As you mentioned, whether it's open-source or closed-source models, they are increasingly integrating stronger agentic AI capabilities, and awareness of agentic AI among enterprises and individuals has significantly grown. This year's Spring Festival with Open Cloud feels very much like last year's DeepSeek moment, which is quite exciting.
Recently, I've received inquiries from many friends, including entrepreneurs, asking if they can just download Open Cloud themselves. I'm not sure if everyone remembers that last year I said that the release of DeepSeek would make many people think they could simply download it themselves. However, realistically speaking, how many companies managed to create a high-end enterprise-grade agent by merely downloading the free open-source version of DeepSeek? By enterprise-grade, I mean agents that have much higher requirements than personal ones, such as permission management, response speed, while also balancing costs.
Open Cloud, at its core, is an open-source framework for personal agents. Many might not realize that last week, Meta faced significant safety incidents due to aggressively adopting an open framework, resulting in information leaks. Thus, we realized that building an enterprise-grade Open Cloud isn't as simple as directly copying Open Cloud.
This involves many inherent enterprise-level security environments—whether your business is understood and well-managed. For example, our company has its silicon-based employee management system, developed long before Open Cloud existed, starting from early 2023.
The emergence of Open Cloud has only accelerated industry recognition, which I believe benefits Bai Rong, without diminishing its competitiveness. It merely validates our correct judgment on the matter. Bai Rong has long been focused on enterprise-grade solutions. Before Open Cloud, we had already created numerous enterprise-grade versions of Open Cloud, promoting them across industries and even enhancing individual awareness. So, first, Dan, I see this as beneficial for Bai Rong because our past two years of promoting ideas to enterprises are now recognized by mainstream audiences.
Secondly, generally speaking, big tech companies provide general-purpose large models capable of handling diverse tasks—from astronomy to geography, poetry composition to painting. For instance, Tencent and Alibaba have launched their own versions of 'Crayfish,' such as WeChat Crayfish and DingTalk Crayfish. Many ask if this impacts us negatively, but I disagree. Upon closer examination, whether it’s Tencent, DingTalk, or Microsoft back then, their focus has always been the B2B general office market, right?
Overall, whether it’s DingTalk or Feishu, these platforms target the office market, unrelated to specialized sectors like healthcare, telecommunications, finance, or retail. Meanwhile, Bai Rong has consistently targeted areas beyond general office markets. Though both serve the B2B sector, one could argue they also serve B2C since individual users aren’t deeply tied to specific industries, whereas Bai Rong dives deep into embedding within enterprise processes with cognitive and permission-managed systems.
So how do I view this? I see it benefiting Bai Rong, as it allows our enterprise-grade silicon-based employees to connect more seamlessly with large companies' general-purpose office agents. We collaborate; for instance, Bai Rong’s legal employee agent 'Baicha' can integrate more conveniently with corporate software like DingTalk, Feishu, or Enterprise WeChat, connecting directly to their office ‘Crayfish.’ Our ‘Crayfish’ serves a different purpose.
Additionally, building an enterprise-grade agent requires stringent knowledge information security and permission management, far beyond merely downloading a large model or Open Cloud, or using generalized office software from major companies. If you’ve used them, you’ll find they’re essentially upgraded versions of Office, right?
Thus, I believe Bai Rong doesn’t compete with open or closed-source large models or 'Crayfish.' We never intended to enter their market. Will they enter Bai Rong’s market? My view has two points: First, about two weeks ago, Anthropic’s founder mentioned in an interview that the largest future AI market lies in applications. When asked why he doesn’t pursue it himself, he responded that he couldn’t possibly cover all industries or accumulate private domain data across every niche, so he focuses on his layer.
Therefore, this judgment has not changed due to whether Open Cloud was born or not. We made this judgment two years ago, and Bai Rong still needs to firmly focus on which layer? First, between general large models, general lobsters, and enterprise-level agents, we will build our Bai Gong. Bai Gong is enterprise-focused; you can think of it as an Open Cloud builder for enterprises. Additionally, we will develop some lobster agents in areas where we have industry knowledge. At the same time, we will open up areas that we are not very familiar with, allowing others, who might now be called IAV (independent agent developers), to build their agents based on Bai Gong, and we will share the revenue.
Combined with our capital empowerment, our future model will resemble Tencent's approach: Tencent empowers through traffic and capital, whereas we empower through enterprise-level infrastructure plus capital, forming a dual-driving model. This year, an increasing number of traditional ISVs have started using Bai Rong’s Bai Gong to develop their scenarios. These intelligent agents are good. That’s a brief response. Thank you, Mr. Dan.
Yang Lin
Alright, thank you very much for the management's time. I am Yang Lin, an analyst from Guotai Haitong in the computer industry. I have several questions. First, after listening to the chairman’s industry insights, we strongly agree. My questions are divided into three parts. The first question pertains to the progress of the company’s silicon-based employee strategy, including customer numbers, future industry share, and revenue expectations regarding silicon-based employees.
The second question relates to what was mentioned earlier about the core advantages of our silicon-based employees compared to various lobsters in the market. As more participants enter the lobster space, how will it affect our business? The third question, which has been widely discussed recently, concerns AI cloud price increases. Since this affects our costs—how will it impact the cost and revenue of our silicon-based employee platform going forward? Thank you for addressing these three questions.
Zhang Shaofeng
Alright, Mr. Yang Lin, let me restate your question. The first one is about understanding the progress and revenue situation of silicon-based employees. Let me ask Mr. Shaofeng to provide a brief answer. Mr. Yang, here’s the situation: actually, as I mentioned earlier during the overall introduction, Bai Rong is somewhat unique. It wasn’t only recently that we started discussing the concept of silicon-based employees. From the company’s inception, we’ve been delivering end-to-end results, even though back then there was no concept of agents or silicon-based employees, right? But if you break down Bai Rong’s business model clearly...
Whether it’s our MaaS (Model as a Service), which we call our H1 business—the relatively mature business—or BaaS (Business as a Service), which we refer to as our H2 business—our so-called growth-phase business—or even our newly incubated businesses, which we call H3, high-growth businesses, none of our businesses are tool-based software. From this perspective, our company has always been focused on silicon-based employees since its inception. Our first business involved creating many silicon-based risk controllers and silicon-based approval specialists to assist financial institutions in reviewing various financial applications from consumers and enterprises.
What kind of scale are we talking about? Every day, we handle tens of thousands of applications, which used to be manually processed by humans sitting at desks. In the early days, paper forms were filled out, and each application form was literally a piece of paper. Bai Rong was the first company to revolutionize this industry. We calculated that if converted into carbon-based specialists—who can typically review dozens of cases per day—Bai Rong has roughly hundreds of thousands of silicon-based specialists, specifically risk control specialists.
That’s the first category within our MaaS business. The second is our BaaS business. What exactly does it entail? Think about it this way: consider traditional banks issuing credit cards. They outsource a lot of card issuance work to BPO companies. How do those companies operate? They go door-to-door in office buildings, handing out flyers and signing people up for credit cards. For each successful sign-up, they earn two to three hundred yuan. It doesn’t matter how they achieve it; they get paid based on results. That’s outsourcing.
At its peak, Bai Rong had nearly two hundred thousand silicon-based specialists making AI calls and chatting on WeChat to introduce financial products. Setting aside the impact of Document No. 9 for now, we’ve seen a reduction since then. Besides these scenarios, we also mentioned our H3 industries. For example, in certain provinces, operators use AI to help customers renew broadband services, upgrade packages, add family members like my dad, and offer roaming packages when traveling to another province. These operations have already reached hundreds of thousands in volume, though still far from our goals, but we’re seeing rapid growth.
So if you ask me about the total number of silicon-based employees, we've already reached hundreds of thousands today. Some fields grow slower, while some niche areas are growing extremely fast, with growth rates potentially dozens of times higher. There was also the example of post-natal care centers and how to provide childcare solutions after leaving these centers — there are many such cases.
Your second question seems to be asking about the competitive advantages of Bailong's silicon-based employees compared to various lobsters on the market. President Yang, correct, correct, correct. Actually, I briefly touched on this when answering Zhongtai's Jiang Dan earlier. The situation is like this: large companies generally don't focus much on enterprise-level applications, right? You must also be very clear that what they have started launching now is akin to upgraded versions of DingTalk, Feishu, and WeCom from back in the day, correct? These are still mainly for office scenarios. We don't compare ourselves with them, nor do we aspire to operate in those scenarios.
But if you're talking about truly enterprise-level operations, we at Bailong haven't really seen any company with a massive scale of one or two hundred thousand silicon-based employees generating billions in revenue annually. Almost none. Companies reaching that revenue scale are mostly traditional software firms. Traditional software companies are tool-oriented, charging for so-called one-time project fees or licensing fees, unlike us. We genuinely charge based on job positions, transaction amounts, or completed workloads under a piece-rate model.
So how can they be comparable? In terms of current scale, we are definitely the largest. Second, because we have long been committed to building silicon-based employees for enterprise-level software, our understanding runs deep and broad. We know there are numerous pitfalls to navigate. If a traditional software company attempts this transition, they would need to encounter all the challenges that Bailong has faced. For instance, many enterprises tell me, 'Hey, I use DouBao for voice interactions, and it feels great. But why can't I achieve the same results as Bailong when using DouBao’s output?'
That's because DouBao has strong internal infrastructure surrounding its AI models. You lack that infrastructure. For example, in communication softswitches, when we were developing our voice intelligence system, we found that concurrency couldn’t handle fifty to a hundred thousand users. There's another infrastructure component involved here — not AI-related, but rather a communications infrastructure known as a softswitch. It's unrelated to AI itself, but your enterprise lacks it. They’ve had it for a long time but never intended to share it externally.
Therefore, big tech companies haven't walked this path; they don't understand the pain points of enterprises. Meanwhile, small and medium-sized traditional software companies lack the capability since they aren't originally AI-focused. Unlike Bailong, which was born into AI — the first generation being decision-based AI and the second being generative AI. We started working on generative AI in 2017. So we possess both the technology and an understanding of the pain points. Previously, Bailong didn't externalize this ability as it was entirely internalized. Now that we are beginning to offer this capability, we find that most companies aren’t even aware of these issues needing resolution. Was that your second question?
As for the third point, the price increase for Bailong's AI cloud platform does not mean a general cloud platform price hike. What impact does this have on the cost and revenue of Bailong's silicon-based employee platform? First, hardware like memory has indeed become more expensive, right? That certainly impacts us. Additionally, regarding the cloud platform price hike, I believe there are pros and cons. The advantage? Many of Bailong's silicon-based employees and underlying large models are self-developed. This gives us an edge over other silicon-based employee providers who may rely on major generic platforms.
Second, what are the risks? A small portion of our use cases — say out of a hundred questions you ask, we might only need to use three to five — will call upon a general large model. This happens for very unusual questions, which we refer to as corner cases or edge problems. This doesn’t significantly affect us. Overall, I think this is beneficial for us because we have self-research capabilities. Like last year, everyone suddenly became bullish on Google, why? Because Google has its own TPU and isn’t reliant on NVIDIA. Okay, President Yang, I hope I’ve clearly answered your three questions?
Yang Lin
Very clear, thank you, Chairman. Those were my three questions. Thank you.
Li Chengru
Thank you, management. My question is as follows: first, regarding the performance, because there was some growth in the first half of last year, but our performance in the second half experienced a cliff-like decline. What is the reason for this? The first issue is about the two factors mentioned in our company's performance forecast: AI investment and the impact of Document No. 9.
Has the company internally conducted any case testing on its business model or R&D window? Why did the external policy impact quickly escalate into a halving of overall profits within half a year? What is the specific impact or expected duration of Document No. 9 on the company’s revenue? And what are the revenue and profit guidance levels for 2026? That's all, thank you.
Duan Ying
Thank you, Mr. Cheng Ru, for your question. Regarding the cliff-like decline in performance you just mentioned, the core reason is due to the impact of the 2025 policy. As everyone knows, Document No. 9, which officially took effect in October 2025, had a significant impact on the financial industry. The main effect of this impact is that some of our clients had their products removed from shelves due to the requirements of Document No. 9. Because of the removal, they are preparing better products, and it takes time to get those back on the shelves. This transition period happened during the second half of 2025, affecting the scale of operations for Berrong's silicon-based marketing specialists, thereby impacting the revenue of the BaaS business, with the major impact seen in the second half of the year.
During this process, the company's upfront investments in AI, especially in our R&D expenditures, have not decreased compared to the first half of the year. We hope to seize the golden window of China's AI development over the next two years by rapidly accumulating a first-mover advantage and expanding customer scenarios. As previously mentioned by our CEO and CSO, if we add back these upfront AI investments, you can see that our mature businesses—MaaS and BaaS—still maintain robust profitability.
Moreover, long-term investors who have followed Berrong may recall a similar regulatory policy impact in 2020, which led to the first year-on-year decrease in Berrong's revenue, with a 10% year-on-year drop. Back then, the year-on-year decline in the BaaS business was even more pronounced. However, by 2021, our revenue began to rebound strongly, with a year-on-year growth rate as high as 43%.
On one hand, this demonstrates that the industry we operate in has ample cash flow; on the other hand, it also shows that Berrong has very high organizational resilience, enabling us to navigate cycles and support long-term business development. For the BaaS business, due to the implementation of Document No. 9 in 2025, many clients’ high-risk businesses were largely cleared out by the second half of 2025. Therefore, we experienced a short-term period of pain during this year.
Of course, if you focus on the various low-risk businesses we mentioned earlier, such as the silicon-based credit marketing specialists, silicon-based wealth management specialists, and silicon-based non-performing asset disposal specialists mentioned by our chairman, these counter-cyclical growth scenarios could still achieve significant rapid growth. So, we are at a brand-new starting point. If you change your perspective and imagine Berrong as a newly established AI agent venture, backed by over three billion in ample cash reserves, our future potential is enormous. That’s a brief response to Mr. Cheng Ru’s questions. Thank you.
Yang Yiran
Hello, may I ask if you can hear me? Yes, thank you. Hello, management team, hello, Mr. Zhang. I am Yang Yiran, an analyst from Guotou Securities International. I mainly have one question. Our business has certain technical and data barriers, but when we develop new businesses, we still need to use enterprise data for model training.
In this context, we might actually be on the same starting line as some of the major general large model vendors. Do these big companies have any cost or time advantages? If a large number of small businesses compete through these extensive large model platforms, could we end up at a disadvantage? I would like to ask Mr. Zhang to analyze this for us. Thank you.
Zhang Shaofeng
Thank you, Yang Yiran from Guotou International Technology Division, for inviting me, Shaofeng Zhang, to answer this question. Actually, I briefly mentioned earlier something related to your question, Yang. Your question is very important; if not thought through clearly, it could lead to issues. As I just mentioned, even the CEO of a company like Anthropic has stated that the most significant layer is in the application, not in the technical mega-problems layer.
Let me interpret this further, and thank you for giving me this opportunity to do so. A general large model, due to its pursuit of generality, aims for AGI, or even ASI for some companies. Thus, it must be knowledgeable about everything—astronomy, geography, poetry composition, and more. But Yang, consider a business: how many of your clients are looking for such wide-ranging knowledge? You mostly care about specific areas. For example, if you're providing legal consulting services, you wouldn't typically handle questions about ordering meals, right? Nor would you deal with tourist attractions.
Therefore, your problem can essentially be confined to a relatively fixed scope. At this point, if we were to use a general large model, think about its parameters. Nowadays, any company working on general large models typically has around one trillion parameters, right? Typically, anything worth mentioning is at the trillion-parameter scale. Although there is MOE architecture, overall, it still isn’t as efficient as our dedicated models.
At BRY, we basically aim to keep the dedicated models we develop for different industry scenarios under 10 billion parameters. Do you know how small our smallest model is? Just one billion parameters, the smallest model we have. So what are the benefits? First, the cost is definitely low. Earlier, a friend asked about cost. Since the cost is proportional to the square of the parameter count—assuming no chip upgrade—the minimum cost is directly tied to the square of the model’s parameters. Second, response speed—larger parameter sizes mean slower response times.
Third is the issue of hallucinations. The larger the parameters, the higher the hallucination rate. Take DeepSeek for example; last year everyone thought it was great, but while it was good, its hallucination rate was also high. However, enterprise-level applications have much lower tolerance for hallucinations than in B2C scenarios. So from the perspective of large models and their general rules, this is the first point: why we try not to use purely general large models—not that we can't, but if you use dedicated ones, the results will definitely be better.
What is the second point? There is a lot of private domain knowledge that cannot be found on public networks. Let me give you a simple example. Today, if you use DeepSeek, Qwen, DouBao, or OpenAI to perform debt collection tasks—trying to get people who borrowed money but haven't returned it—you'll see how effective they are. Most large models are designed to be agreeable. If someone says, 'I'm unemployed, how am I supposed to pay back the money?' the model might respond, 'You're absolutely right. Relax, find a job first, take a break.' Can it carry out collections effectively?
Why can't it do that? Because it lacks the data, the data specific to that exclusive domain. Debt collection is just an example; can you find such data online? It's not crawlable, nor purchasable—it involves privacy, right? This is called private domain-specific knowledge. Without the relevant scenario, you can't train a suitable model.
Thus, BRY's competitiveness lies in two aspects: first, the model itself is dedicated, smaller in size, faster in response, lower in cost, and has a lower hallucination rate. Second, we have accumulated a lot of scenarios over time, and our private domain data keeps increasing. This makes our dedicated models perform better than general large models, even if our parameter count is the same. OK, you might ask, can small businesses create such models themselves? I touched on this a bit earlier. Didn’t DeepSeek come out last year? How many small businesses succeeded in creating something similar?
Because although you may say that your technology has advanced a lot, the demands on your company have also increased significantly. Everything rises with the tide, right? OK, you say I couldn't achieve DeepSeek last year, but this year, with technological progress, I can do it. But OK, your competitors' requirements are no longer at last year's level, right? Your competitors' requirements are higher than yours, so if you say that achieving DeepSeek's level this year is useful, it actually isn't because your competitors might be using technology far beyond DeepSeek. This forces you to adopt relatively advanced technology due to competitive pressures. For example, as I just mentioned with lobsters, right? Try setting up an enterprise-level lobster business; can you do it? Even Meta ran into big problems and couldn't pull it off. Alright.
Robin
Hello everyone, thank you, thank you management. This is my question, I am Robin from Bloomberg Intelligence. There are substantial investments in AI agents. Can you help us break down the specific areas of development, including R&D expenses, computing power procurement, talent acquisition, and more? Additionally, regarding AI-related investments, what is the ratio between expensed and capitalized costs? Looking ahead, will the AI-related capabilities expenses in 2026 be one-time investments or sustainable? Also, what is the budget for labor costs in 2026, and how is the personnel expansion planned? Thank you.
Kelson Chen
Alright, thank you, Mr. Robin from the Bloomberg Intelligence division. This question will be answered by Mr. Kelson. Thank you for this question. Last year, we indeed made substantial upfront investments in AI. Our total R&D expenditure for the year was approximately 637 million, an increase of about 25%. In the first half of the year, the expenditure was around 302 million, growing by 33%, while in the second half, it was 335 million. So, you can see that our R&D investment was evenly spread throughout the year. However, if you zoom out and consider other related cash expenditures such as servers, IDC data centers, and investments in developing new ecosystems and incubating businesses, the overall business investment growth would be even larger.
The largest portion of our R&D investment goes to talent, as every company is competing for talent. In 2025, the introduction of AI R&D talent alone will increase by over 100 million, growing by 24%. However, our return on investment in talent is very high. As our CEO mentioned earlier, 2026 will be a critical year for the company to transition from strategic capability-building to value realization. Therefore, we will continue to place great emphasis on optimizing and upgrading AI talent.
But our investment in talent is truly worthwhile. We can calculate the efficiency of our employees: our per capita output value was close to two million in 2023 and exceeded two million in 2024. Over the past few years, our average annual salary per employee has been around five to six hundred thousand, resulting in a talent ROI of up to three times. Currently, mainstream AI companies have yet to see profits, but BR Group is genuinely one of the few companies doing real AI work and achieving actual profitability.
We have sustained profitability and are determined to become the Google or Amazon of the agent AI era. Hence, we must have the resolve to make these investments. Some AI companies, including some listed in Hong Kong, which I won’t name, have accumulated losses exceeding 6.2 billion over three years, with some losing 3.6 billion in just three quarters. A single data center costs fifty billion US dollars, and Xiaomi plans to invest two hundred billion over the next five years in their chip AI operating system.
So, you can see that, in comparison, our investment cost-effectiveness is already very high. We have always focused on vertical implementation and result-oriented investments, enabling us to deliver solutions to our clients rapidly. Thank you for your question.
Sandy
Alright, thank you, Mr. Kelvin, and thank you, Mr. Robin. Due to time constraints, our Q&A session is now concluded. Thank you all for your questions and thank you to the management team for their answers. Once again, thank you for participating in BR Group's 2025 annual results briefing. If you have further questions, please visit our investor relations website at ir.brgroup.com or email IR at brgroup.com for inquiries.
We also welcome everyone to scan the code and follow Bairong's digital AR assistant, Bairong Finance, which can interact directly with you and regularly release updates on the company’s business developments. We look forward to meeting and speaking with you again at our next earnings report. Thank you.
More details:BAIRONG-W IR
Note: The above content is generated by an AI language model based on publicly available data and third-party automatic subtitles. The content does not represent any position of Futu, nor does it constitute any investment advice. Futu Group makes no express or implied guarantees or statements regarding the accuracy, timeliness, or completeness of the above content.
Risk Disclaimer: The above content only represents the author's view. It does not represent any position or investment advice of Futu. Futu makes no representation or warranty.Read more
Comment (1)
to post a comment
21
3