As large models become increasingly powerful, what kind of AI are enterprises willing to pay for?
The capability of a model alone is perhaps no longer sufficient to justify a company's adoption of AI. The market's focus has shifted from the model itself to how AI can generate tangible and quantifiable output.
The semi-annual report recently disclosed by Bairong Intelligence presents just such a trajectory of change.
In the first half of 2026, the company's total revenue was affected by external factors and experienced fluctuations; on the other hand, its AICC business saw a year-over-year increase of 52%, with revenue from new scenarios targeting more industries growing by 195%.
AICC stands for AI Contact Center (intelligent contact). In simple terms, it means integrating AI into enterprise contact centers, enabling agents to directly handle specific tasks such as inquiries, marketing, service, and operations.
Bairong disclosed that the company has honed its mature AICC and enterprise-level Agent capabilities in financial scenarios and is now replicating them in more scenarios, including logistics and securities firms.
The daily processing volume of its logistics project has increased from less than 1,000 deliveries at the initial launch to over 15,000 deliveries. A leading securities firm has also expanded its deployment scale from around 30 seats at the beginning of the year to around 210 seats.
Moreover, this approach is continuing to expand outward. The RaaS model adopted by Bairong, which charges based on results, allows processing volume and deployment scale to be more directly converted into revenue.
Once a benchmark scenario is successfully implemented, the same set of capabilities and knowledge can be replicated to more positions and more industries.
As the deployment scale expands, a large amount of interactive feedback generated by actual operation can continue to be fed back into model training and strategy optimization, forming a data flywheel.
This chain continues to expand outward, and capital and industrial resources are also starting to play new roles.
Bairong introduced the concept of "technology reconstruction-type AI Roll-up", which essentially involves acquiring more real business entry points through investment, cooperation or industry integration, and then embedding the verified Agent capabilities into them.
Thus, Baiying plans to start from a specific contact scenario and build a growth mechanism that gradually expands:
Start with a single use case, then scale deployment, gradually replicate across industries, and keep feeding real-world data back into the model. From there, leverage industry resources to unlock more workflow entry points.
In the latest earnings call, Bairong CEO Zhang Shaofeng defined this round of adjustments as a strategic upgrade, with the starting point being AICC and a batch of silicon-based employees who have truly begun to take on specific tasks.
Bairong Intelligence's financial report reveals a strategic shift
Looking at the latest semi-annual report, a notable feature is that Baiying's different businesses have already shown differentiation.
This earnings report also reclassifies revenue according to business attributes, dividing operations into three main segments: AI decision-making, AICC, and other businesses. This reclassification itself makes Bairong's ongoing business transformation more intuitive.
In the first half of 2026, Bairong achieved revenue of 914 million yuan, a year-over-year decrease of 43%; gross profit fell from 1.182 billion yuan to 488 million yuan, with the gross margin declining from 73% to 53%.
On the profit side, the company turned from a profit of 201 million yuan in the same period last year to a loss of 322 million yuan.
Earnings reports explain that this round of fluctuations has a relatively clear external background —
Regulatory documents governing internet loan assistance businesses, released in 2025 and formally implemented that October, combined with a series of rules rolled out in the first half of this year, have led some financial institutions to suspend operations of certain products, in turn affecting Bairong's corresponding mature business lines.
In the first half of this year, Bairong AI's decision-making business revenue was 372 million yuan, down 26% year-over-year; revenue from other businesses was 377 million yuan, down 62% year-over-year, with the latter's decline mainly due to a contraction in intelligent marketing and operations revenue from some credit products.
However, the fundamentals of AI decision-making have not seen corresponding fluctuations in operations.
In the first half of the year, the business reached 141 core institutions, with their revenue share rising from 76% to 84% while retention held at 96%. Core revenue from banks and internet platform institutions grew 1% year over year.
Therefore, the management of Baiying still positions AI decision-making as an important foundation for profits and cash flow, to support the next stage of AI investment.
On the other hand, Bairong's AICC has deviated from its growth curve in the opposite direction.
In the first half of this year, the business achieved revenue of 166 million yuan, up 52% year-on-year; of which, credit scenario revenue was 155 million yuan, up 47% year-on-year, and new scenario revenue was 11.26 million yuan, up 195% year-on-year.
New scenarios currently account for around 7% of AICC's revenue, and the scale remains small, but the growth rate has started to emerge in the financial data. AICC's overall proportion of Bairong's revenue has also risen from around 6.8% in the first half of 2025 to approximately 18.1% in the same period this year.
This increase is due to both AICC's own growth and the contraction of total revenue, but the diversification of the business structure has become relatively clear.
Research and development investment has further reinforced this trend.
Bairong's research and development expenses increased from 302 million yuan to 394 million yuan in the first half of the year, up 31% year-over-year, with the proportion of revenue rising to 43%; as of the end of June, 901 of the company's 1,477 employees were research and development personnel, accounting for 61%.
Choosing to continue increasing R&D investment during a period of overall revenue pressure also means that Bairong's AI strategic adjustment has begun to take shape in terms of resource allocation.
On the conference call, Zhang Shaofeng further clarified that AICC's new scenarios have become a key growth business for the group, undertaking the task of replicating across industries and finding new increments, and new resources will also be prioritized towards AICC's new scenarios and other silicon-based positions.
Zhang Shaofeng, Founder and CEO of Bairong Intelligent
The strategic upgrade announced by Bai Rong this time is first reflected in the expansion of industry scope.
Bairong's previous AICC capabilities were mainly accumulated in the financial field, and are now being extended to service-intensive industries such as logistics, aviation, securities, banking, telecommunications operators, and insurance.
Along with the evolution of the industry, the complexity of tasks undertaken by AI has also changed.
The previous generation of AICC was more commonly applied in outbound scenarios, relying on keyword matching and preset dialogue, and is suitable for standardized tasks such as notification reminders and fixed-process marketing.
In contrast, complex inbound calls are distinctly different. After users initiate a call, their problems, modes of expression, and emotions are all difficult to anticipate in advance, and AI must be able to understand open-ended questions, maintain long-term multi-round conversations, and invoke knowledge bases and internal business systems to continue completing subsequent tasks.
Zhang Shaofeng revealed on a conference call that traditional response robots, when faced with such complex scenarios, have a transfer-to-human rate of as high as 70% to 80%. The new generation of AICC aims to bridge the gap between understanding problems and completing service processes.
Bairong refers to this type of enterprise-level Agent that has entered the actual service process as "silicon-based customer service".
In terms of cross-industry replication, Bairong has also made some tangible progress.
In July, Bairong's silicon-based customer service started handling complex inbound tasks such as package inquiry, delivery exceptions, and complaint feedback in logistics scenarios, with daily processing volume increasing from less than 1,000 calls at the initial launch to over 15,000 calls by the end of August.
In the brokerage sector, as of July, Bairong had signed contracts with or was in the contract process with 13 institutions. Among them, deployment at one leading brokerage expanded from roughly 30 seats at the start of the year to about 210 seats.
These advancements indicate that the AICC capabilities, which were previously mainly focused on financial scenarios, have begun to enter new production environments.
The industries that Bairong has currently chosen to prioritize also have relatively clear commonalities - they have continuous and massive C-end interaction demands, and the industry concentration is relatively high.
Logistics companies constantly generate inquiries about checked items and abnormal items every day, while airlines have long-term service demands such as ticket refunds and rescheduling; leading enterprises are also relatively concentrated, and once a benchmark project is successfully implemented, the same capabilities can be more easily replicated internally to more scenarios and to other companies in the same industry.
The call also revealed another aspect, where the time-consuming steps in promoting across industries are currently focused on the procurement verification cycle of large enterprises, as well as system integration and production cutting.
Before large-scale enterprises formally adopt a system, they typically need to go through a proof of concept (POC) to verify the problem resolution rate, proportion of human intervention, response efficiency, and stability. After entering production, business volume will only be gradually released after safety, compliance, and service boundaries have been confirmed.
So the difficulties of scaling up are no longer just about the model.
A new question arises: since enterprise-level Agents can be applied to many scenarios, why does Bairong choose AICC as the most important breakthrough point for now?
Why Do Enterprise Agents Land on AICC First?
Enterprise-level Agents have increasingly entered repetitive workflows, but to transition from technical feasibility to large-scale implementation, the first challenge to be addressed is how to clearly measure business value.
Ultimately, a company's adoption of AI boils down to the return on investment. However, in many scenarios, it is difficult to directly quantify in the short term how much AI improves decision-making quality and releases latent efficiency.
In contrast, AICC is relatively special.
Contact centers have long been responsible for handling a large volume of inquiries, marketing, and service tasks, with clear boundaries, processes, and evaluation standards. After the integration of AI, the processing volume, transfer-to-human ratio, and unit service cost can all be quickly reflected in the operational results.
Top venture capital firm a16z had also previously listed programming, enterprise search, and customer support as several scenarios where enterprise-level AI is likely to be applied in production earlier on.
The customer service field has clear business objectives, with most processes already documented, while also retaining a manual backup path, making it naturally suitable for Agents to gradually take over standardized work.
Of course, Baiying's choice of AICC is also due to its own accumulation.
Zhang Shaofeng revealed that the company's customer service and marketing silicon-based employee project has been in place since 2017 and has undergone seven or eight years of continuous iteration. The financial industry has high requirements for accuracy, stability, and compliance, resulting in capabilities that highly overlap with AICC's needs for voice interaction, intelligent contact, and complex process execution.
Bairong has thus built a relatively complete technical system around real production environments.
First and foremost is the core model capability.
Bairong currently possesses the BR-Voice multi-modal voice model, the BR-LLM-Proactive proactive dialogue model, and the BR-Vision-Doc document understanding model.
All three belong to Bairong's self-developed vertical basic models, of which the first two are precisely the important foundation for constituting the AICC vertical model.
The technical route of this model system is relatively clear - it does not pursue being "comprehensive", instead placing more emphasis on optimization around specific tasks.
Zhang Shaofeng revealed that the company's models pre-trained for phone scenarios have parameter scales mainly controlled between 0.4B and 30B, with the core goal of achieving a balance between accuracy, stability, response latency, and inference cost that is more suitable for production environments.
On the conference call, Zhang Shaofeng further explained the differences between this route and those of some overseas AICC companies.
Based on his observations, companies such as Sierra and Decagon largely build their applications and harness engineering on top of third-party foundation models from providers like OpenAI and Anthropic. Bairong, by contrast, has integrated the underlying vertical foundation models into its own technology stack, taking full control of key stages including pre-training, post-training, and inference optimization.
This allows for continued judgment of whether the issue lies in the application engineering layer or the foundational model layer after actual operation problems arise, and directly enters the next round of training and optimization, shortening the entire iteration chain.
This also reflects a difference between enterprise-level AI and general models. General models emphasize the upper limit of capabilities, but once they enter a production environment, they must also contend with constraints such as accuracy, stability, response speed, and cost.
Public testing results have provided some external validation for Bairong's vertical industry approach.
In the multilingual conversational speech understanding task at the INTERSPEECH 2026 MLC-SLM Challenge, Bairong achieved an accuracy rate of 94.84%, ranking third globally and first among domestic industry players.
However, model capabilities are just the starting point for production-level implementation, and real-world phone calls are not static question banks. User expressions, business needs, and abnormal situations are constantly changing, so even if the model achieves a high accuracy rate initially, it still needs to be continuously corrected during production runs.
The second major capability that Bairong has been focusing on accumulating has thus fallen to Data Flywheel.
Under the premise of legitimacy and compliance, the company identifies positive and negative feedback from actual production, such as whether a service ultimately completes its task, which expressions lead to deviations in understanding, and which links ultimately enter manual processing, and then converts this information into reinforcement learning signals to continue adjusting the model, service strategy, and conversation approach.
The image was generated by AI.
Real-world operations become part of the model's continuous training, forming a cycle of business operations, feedback, model iteration, and return to business.
However, after the model can continue to iterate, there is still a threshold to truly enter the contact center, which is the need to run stably in the telephone network.
Traditional phones typically use 8kHz narrowband voice, which retains significantly less acoustic information than the 16kHz wideband audio commonly used on the internet, posing higher requirements for speech recognition and real-time interaction.
The real contact center also needs to handle communication softswitch, high concurrency scheduling, system docking, and recording quality inspection.
Bairong has long been focused on building a telecommunications engineering system for phone networks, while also adapting to aspects such as narrowband identification, noise reduction, user interruption, and conversation turn judgment.
However, once the technology is truly integrated into the business, it must also confront another layer of issues that are highly dependent on the scene: delivery.
Different companies have distinct knowledge systems, operational rules, and system environments, and after AI completes a conversation, it may continue to query information, call internal systems, execute tasks, and write back the results.
Bairong has set up a team of FDE frontline deployment engineers for this purpose. After FDE enters the business site, it will sort out service standards, access conditions, knowledge bases, and system interfaces, break down the original manual workflows, and then reorganize the links suitable for AI to take over into Agent workflows.
In this way, the multiple layers of capabilities formed by Bairong around AICC are perfectly interconnected: vertical model provides task capability, real business feedback drives model iteration, communication engineering ensures AI can run in a telephone environment, and FDE embeds the technology into specific workflows.
After the Agent enters production, it continuously generates new feedback, which is then used to re-enter the next round of model training.
Bairong's subsequent AI investment is also aimed at a more specific goal: making AI a long-term, sustainable production capability.
And once Agents start taking over work originally done by humans, what changes next is no longer just about technology.
The way companies purchase AI is likely to change as well.
AI is starting to enter production systems
The software industry has developed to date, and enterprise procurement of technology has formed a relatively mature business model.
In the early days, software was primarily sold through licenses, but with the rise of SaaS, subscriptions and seats became the mainstream; the era of large models has also introduced token-based billing.
Although these models have different pricing units, the core of what enterprises buy is still a kind of technical capability.
The rise of Agents has enabled AI to further penetrate the task execution phase, directly undertaking specific work.
In this process, the object of enterprise payments gradually shifts from technology usage to actual output, and the value range it can reach extends from traditional IT budgets to operational costs and labor costs.
Bairong is currently adopting a RaaS (Results as a Service) model, which can be seen as a real-world manifestation of this change.
According to the company's disclosure, RaaS can charge based on successfully completed and automated effective calls or work orders, or based on the actual deployment scale of Agents, and in some scenarios, it will also charge based on the business scale it helps to promote.
On the conference call, management revealed that taking the already implemented logistics business as an example, after charging for effective calls that successfully resolve issues, the single-processing cost can currently be reduced by more than 50% compared to traditional manual methods. The income of the technical service provider is thus directly linked to the actual work completed.
This is also the biggest difference in billing between RaaS and traditional software.
After the software is delivered, subsequent operations are mainly undertaken by the user; pay-for-results further pushes technology companies into the internal business process, where model performance, stability, cost, and task completion effects all need to be incorporated into the same operational logic.
As AI begins to penetrate operations and human resources, a bigger issue arises: how can this newly formed AI capacity be integrated into more real-world workflows and further replicated on a large scale?
The AI Roll-up trend, which has surged in popularity from Silicon Valley in recent years, has drawn attention against this backdrop.
Roll-up was originally a form of industrial integration, typically involving the consecutive acquisition of decentralized enterprises within the same industry, followed by scale integration and back-office sharing to achieve synergies.
Once AI is integrated, the key shift is that after securing a real business entry point, it continues to restructure workflows, service delivery, and cost structures through vertical models, agents, and engineering systems.
Therefore, the generalized AI Roll-up is no longer limited to acquisitions. Capital acquisitions, strategic investments, industrial cooperation, and even direct deployment of Agents can all become ways to enter real workflows; what truly matters is whether AI has entered daily operations and continues to change efficiency, capacity, and unit economics.
Currently, a relatively typical route has taken shape in North America: capital first acquires traditional service companies and workflow entrances, and then AI engineering teams carry out renovations.
For example, the integration of accounting firm Current and the acquisition of PartnerHero by Crescendo AI are examples of this practice of capital control combined with AI engineering transformation.
Domestically, two paths are emerging:
One category is the industrial aggregation-type AI Roll-up, which first enters a decentralized service network through capital, joint ventures, or industrial integration, obtaining business systems and real workflows, and then carries out AI upgrades.
Another category is the technology restructuring-type AI Roll-up, which can be directly applied to enterprise workflows, using vertical models and agents to undertake specific tasks, and then gradually expanding the deployment scope, while also extending outward through strategic investments and industry cooperation.
Bairong is one of the representatives of technology-driven AI roll-up.
The two paths have different sequences: the industrial aggregation type emphasizes aggregating industrial entrances first, then carrying out AI transformation, while the technological reconstruction type first uses AI to transform workflows, then expands industrial coverage.
What they have in common is to further push AI from the periphery of enterprise software systems to the interior of production operations.
From this perspective, RaaS and AI Roll-up actually correspond to different levels of enterprise-level AI commercialization: the former solves how AI is billed after completing work, while the latter further addresses how this production capability is organized and enters more real businesses.
AI Roll-up is responsible for changing the way production capacity is organized, while enterprise-level Agents actually implement specific tasks.
The software has digitized a large number of business processes, and large models have further provided understanding and reasoning capabilities. Now, Agents are beginning to enter the processes themselves, directly undertaking specific tasks within them.
After entering real-world industries, the boundaries of competition have also expanded. Enterprise-level AI needs to connect models, engineering, and business processes into a long-term running system, and continuously undergo testing in real production environments.
The industrial significance of Agent has also begun to exceed that of a single large model product, and is gradually becoming an entry point for machine intelligence to enter real economic activities.
As a result, machine intelligence began its transition from a technological product into a factor of production in economic activity.
Historically, every productivity revolution has often started with a small number of specific jobs, and then gradually changed the cost and organizational structure of the entire industry.
This time, enterprise services are becoming the deepest territory for AI, and Baiying has stepped into this field through AICC.