Everyone is integrating DeepSeek and WorkBuddy, so what will ultimately set brokerage AI apart?

Title context: Everyone is integrating DeepSeek and WorkBuddy, so what will ultimately set brokerage AI apart? Text: In the past two months, the most intensive move by brokerages has been "moving in." On September 2, Tencent's WorkBuddy open platform was officially launched, and GF SEC became the first brokerage institution to move into the ecosystem zone, with the two sides jointly releasing the "GF SEC" Buddy application zone, debuting 12 self-developed Skills and 9 expert capabilities; on September 7, Alibaba's Qwen open platform launched more than ten financial intelligent agents, four of which were brokerage agents: Guotai Haitong's "Lingxi," Industrial's intelligent investment assistant, Soochow's "Xiucai," and CICC Wealth; in late September, six major skills of "Dragonfly Skill," self-developed by China Securities Co.,Ltd., were also launched on the WorkBuddy skill plaza. From integrating DeepSeek, to launching Skills, and now to brokerages intensively moving into WorkBuddy, the securities industry's embrace of external large models and the Agent ecosystem is still accelerating. But a new question has emerged: as more and more brokerages use the same or similar general-purpose large models, what will ultimately enable brokerage AI to stand out? It is not a relationship of substitution; the differences at three levels are the key Is brokerage AI versus general-purpose large models a binary opposition? In the view of the brokerages interviewed, the answer is no. From the perspective of implementation paths, most brokerage AI applications currently fine-tune or post-train on top of general-purpose large models, and then combine them with brokerages' unique application scenarios to serve employees and clients. The real differences are reflected at three levels. The first is the knowledge foundation. The corpus of general-purpose large models mainly comes from the public internet, while the knowledge of brokerage AI is built on proprietary-domain data such as market quotes, trading, positions, investment research, and regulatory standards, superimposed with the research frameworks and business experience accumulated by institutions over the long term. For the same question, what is obtained from a general-purpose model is mostly a "summary of public information," while brokerage AI can connect it to capital flows, positioning in the industrial chain, and similar historical market structures. The second is the deliverable. General-purpose large models mostly exist in the form of standalone Q&A, delivering information and content; brokerage AI must be embedded in real business processes and operate together with permission systems, risk control rules, compliance reviews, and accountability chains, delivering executable, traceable, and accountable business results. This also explains why brokerage AI ultimately often points to order placement, market monitoring, and post-market review, rather than stopping at a chat box. The third is source credibility and boundaries. The professionalism of the financial industry determines that compliance is the bottom line for innovation, and "reliable sources and controllable output" are two gates that vertical models cannot avoid. Sample one: GF SEC, putting AI into a mature App rather than starting over In July 2026, GF SEC released the AI-native version of Easy Taojin, "AI Lens," built on the self-developed "Tianji Zhirong" large model service platform, integrating RAG retrieval-augmented generation and Agent orchestration technology to form a closed loop of "planning-tools-memory-evolution"; on the enterprise side, it built the Skill Hub platform, accumulating 30 types of standardized intelligent agent plug-ins, and laid out an MCP open workbench, launching more than 110 MCP services including account management, market information, and data tools. GF SEC's particularity lies in the fact that it did not make AI into a single entry point, but into a network. For the client side, it put AI into a mature App. In July 2026, GF SEC released the AI-native version of Easy Taojin, "AI Lens," centered on four types of capabilities: "focus, magnify, see through, and look far," corresponding to the four high-frequency scenarios of stock selection, market monitoring, trading, and Q&A. In the trading scenario, it can recognize users' vague expressions and assist in completing orders after confirmation, enabling AI to move from "assisted inquiry" to "assisted execution." Unlike some peers that build separate independent AI Apps, GF SEC chose "in-place upgrading," so tens of millions of existing users do not need to migrate. For internal employees, the other end of AI is connected to investment advisers and researchers. The self-developed "Tianji Zhirong" investment adviser cockpit integrates nearly 50 professional intelligent agents, covering more than 20 core functions such as individual stock commentary, valuation analysis, and research report summarization, serving more than 300 business departments and thousands of frontline investment advisers; the investment banking vertical large model "Investment Banking Wenquxing" covers 10 types of business scenarios and has implemented 56 subdivided AI functions; on the research side, the "Smart Data" system has accumulated 88,000 selected research reports, more than 90,000 core indicators across the industry, and more than 800 thematic industrial chain maps. With the help of AI chain-of-thought investment research intelligent agents, the time required to generate regular industry reports has been compressed from 4 hours to 15 minutes. The content chain is a link that is underestimated by the outside world. According to the company's introduction to reporters, GF SEC uses the "Jinchuangyi" all-media content operations center to connect the hot spot radar, intelligent creation, intelligent review, distribution center, and data cockpit; on the Easy Taojin App side, it has formed an all-weather companion matrix in which "'Taojin Morning Express' refines clues before the market opens, interprets unusual movements and rotation during trading, and generates personalized reviews after the market closes based on watchlist positions." Sample two: CICC Wealth, a vertical large model betting on buy-side investment advisory If GF SEC focuses on path selection, CICC Wealth focuses on redoing services. According to the company's introduction to reporters, its financial vertical large model is not a simple Q&A Siasun Robot&Automation, but extracts the mature methodologies of CICC's 300 senior researchers and nearly 3,000 frontline investment advisers, and is built on internal investment research data and real-time market quotes, aiming to solve the industry problems of "fragmented output" and "point-based distribution across multiple scenarios" in traditional AI services. On the product side, the intelligent assistant "Xiao Jinn" focuses on moving from "Q&A" to "proactive service": it pushes exclusive AI morning briefings in the morning, tracks unusual movements in individual stocks during trading and proactively reminds users, and when hot events occur, sorts out the complete timeline from policy introduction and industry reaction to capital flows; the "AI Smart Investment 50" series revolves around clients' trading trajectories, covering algorithms, stock diagnosis, timing, stock selection, and position diagnosis. CICC Wealth said the company will soon launch the V13.0 AIAPP, with the main theme of "relevant to me, proactive service," focusing on building an online buy-side investment advisory service system. The ammunition comes from the business itself. As of June 2026, CICC Wealth's buy-side investment advisory assets under custody had exceeded 170 billion yuan, and product assets under custody had exceeded 530 billion yuan. On the external ecosystem side, CICC Wealth's intelligent agent also appeared on the Qwen open platform in September. External model access and self-built capabilities are not contradictory; brokerage AI has entered "two-tier competition" Looking back at the path of brokerage AI development over the past year, technological iteration has in fact undergone several major shifts. In early 2025, brokerages' focus was still on the speed of access to and local deployment capability of foundational models such as DeepSeek. At that time, Huafu, Guojin, Guoyuan, Industrial, Everbright, Guotai Haitong, and many other brokerages successively disclosed completion of DeepSeek deployment, and GF SEC, CICC Wealth, and others also quickly followed; at that time, the main applications were still concentrated in areas such as knowledge Q&A, investment research, customer service, and R&D assistance. By this year, competition began to enter a second stage. After the rise of Skills and Agents, brokerages began to export financial data, research methods, and business tools to more AI entry points; the emergence of WorkBuddy further extended competition to the external ecosystem. But judging from the latest disclosed practices of GF SEC and CICC Wealth, "accessing external models" and "building internal vertical capabilities" are in fact not two mutually substitutable routes. External general-purpose large models solve foundational capability issues such as language understanding, reasoning, and intelligent agent orchestration; brokerages' internal models and financial capability layers are responsible for solving issues of data credibility, professional depth, business execution, and compliance boundaries. The former determines whether AI is "smart enough," while the latter increasingly determines whether it "can truly be used in securities business." Especially in financial service scenarios, compliance remains a threshold that cannot be bypassed. CICC Wealth said its AI output is based on internal investment research data and reliable information sources, and controls model output through financial security mechanisms; GF SEC emphasized that brokerage AI needs to operate together with permissions, risk control, compliance review, and accountability systems. From this perspective, the brokerage AI competition is forming an increasingly obvious "two-tier structure": the bottom layer is the general-purpose large model and Agent ecosystem, becoming more open and standardized; the upper layer is the data assets, research capabilities, business processes, client systems, and compliance capabilities that each brokerage has formed over many years. As foundational models gradually become an industry "standard configuration," what is truly difficult to quickly replicate may be precisely the latter. This article is reprinted from Cailian Press, edited by GMTEight: Chen Wenfang.