Cobo: How do we use AI for transformation? 】 Author: Alexzuo4, Investment&Custody VP @ Cobo. Starting from the end of 2024, Cobo has been exploring the combination of AI and blockchain in addition to its core cryptocurrency custody and stablecoin payment businesses. The first thing we saw was the standardized skill potential brought by MCP. In theory, if skills are standardized enough, AI can call upon capabilities like plugins, and blockchain will become the most natural financial infrastructure for AI. So we incubated an MCP app store internally. But it was quickly falsified. At that time, the threshold for AI was still high enough that only mature engineers could proficiently call it, and MCP was not standardized enough. Each docking was time-consuming, labor-intensive, costly, and slow to promote, and the landing effect was far less than imagined. But the AI team has finally built it up. Very expensive, difficult to recruit, and impossible to easily remove. So we decided to change direction. Since we can't transform the customer world yet, let's first transform ourselves. The first question: As an asset custody company, Cobo is extremely sensitive to both data and internal technical process frameworks. There are also strict data hierarchies internally. But without data and real business input, it is impossible to develop the company's own agent. Our earliest thought was local model deployment. But the reality is that the intelligence level of the local model does not meet the requirements. Can run, but not very useful; Can answer, but not smart enough. In the end, Claude and Gemini were chosen as the main candidates (you can apply for ZDR - zero data retention clause to achieve the highest level of isolation). But the big model is just the underlying 'brain' of the business. The truly complex ones are data and permissions. We later developed a complete set of internal knowledge base and agent framework. The internal knowledge base+cobo self-developed agent system knowledge base is responsible for the company's internal data layering. Assign readable ranges based on employee permissions. When calling the knowledge base, the agent also inherits employee permissions, rather than having a 'god's perspective'. The details here include: how to isolate the network environment, how to restrict cross layer data flow, how to control log retention and auditability, how to avoid sensitive information leakage. These are not sexy, but they determine whether this matter can continue in the long run. AI cannot become a security vulnerability. The problem after the architecture is built: no one uses it. Even today, the company still faces a real problem: many front-end businesses disdain AI. If we only encourage the use, AI will not change the workflow. We later realized that we had to start with company management. The first breakthrough point: OKR Agent. Our first strongly promoted scenario was not customer service or coding. It's OKR. We use AI to break down company strategies, assist in setting OKRs, track progress, and review checkpoints. That is to say, the management of the company should gradually shift from human management to silicon carbon co governance. This process is extremely uncomfortable for employees. Previously, the goals could be written more beautifully and the process could be explained more reasonably. Now the weekly data is there, with fewer and fewer excuses. From that moment on, the goal was no longer just a discussion in the meeting, but a continuous record in the system. Strategy OKR supervises business progress every week, but also starts with performance, so that every talent truly becomes familiar with AI. Because you don't participate, it will directly affect your salary. From performance to business: comprehensive agency. After OKR started running, we began to promote internal service agency. We use a combination of evaluation and bonuses to force each department to establish agents related to their own business. Customer service acts as a customer service agent. Legal contract assistance agent. Sales work as a CRM agent. Searching for the most bizarre customer agent, over 100 agents were ultimately launched. We cannot accurately quantify the results of 'silicon carbon co treatment'. But at least one change is clear: when encountering problems in the past, the first reaction was "should we hire one more person. The first reaction now is, 'Can we get the system involved first?'. This is actually what we understand as silicon carbon co treatment. Not an AI replacement for humans. But it's people getting used to working with systems. The road we have traveled this year, there are several very practical experiences. Firstly, there is a healthy cash flow. If the company's cash flow is unhealthy, this transformation will not reach its end. AI is not a money saving tool, it is a long-term structural upgrade with upfront investment. Thanks to Cobo's main business and healthy cash flow. Secondly, top-down promotion is necessary. Organizations will not change spontaneously. If the management does not push it forcefully, this matter will naturally fail. As is well known, the founders of Cobo are all heavy AI players. CTO Dr. Jiang started some AI research as a postdoctoral fellow at CMU in the 2000s. Thirdly, it must be mandatory to use. If it's just encouragement, AI will always stay at writing emails. The real change that enters the process must be somewhat 'mandatory'. Fourth, first solve your own business. Many companies talk about AI+Web3. But if you haven't completed AI transformation internally, all you talk about externally are concepts. Looking back, we cannot fully quantify this transformation. The company is gradually shifting from a "human driven process" to a "goal driven system". If 'intelligent organizations' really emerge in the future, they must not have evolved naturally. It was pushed out by discomfort round by round. Thanks to the participation of all employees, the company can better understand the real needs in the AI era. This is also a byproduct of our internal transformation. We recently launched the Cobo Waas Skill. Cobo WaaS Skill is an integrated and operational capability layer designed specifically for AI Coding Agents. Through structured knowledge, executable examples, and scenario orchestration, it enables Agents to accurately call WaaS APIs. We are upgrading the wallet API to a financial capability module that can be directly called by AI agents. The development cycle has been shortened from weekly level to dialogue level. This is not the result of any product inspiration. But it is the result of the natural overflow of our capabilities after this round of silicon carbon co treatment internally. We are still exploring. But at least today's Cobo is no longer the company of 2024.