EN / 中文

How to Deploy Enterprise Agents: Data Integration, Human-Agent Division and Iterative Operation

by aifenxiifenxi·November 16, 2026

"We avoid grand narratives and instead identify entry points within specific business scenarios." Liu Tao, Deputy Director of the Technical Committee Office at Guangdong Radio and Television Station, chose to start with concrete tasks such as segmenting athlete footage when exploring Agent applications. Northeast Securities shares a similar perspective. Facing business needs, their technical team first understands how employees work daily, identifies pain points, and maps out relevant processes and systems before collaborating with business units to define suitable scenarios.

How can enterprises effectively leverage Agents? The common approach emerging from several companies' practices is to start with specific scenarios, integrate data and business workflows, and then realize value through the participation of business personnel and continuous operations. At the Enterprise-Level Agent Industry Practice Session of the Tencent Global Digital Ecosystem Conference, China Resources Trust, Guangdong Radio and Television Station, Yunnan Baiyao, and Northeast Securities shared their respective explorations.

This article synthesizes these cases to outline methods and experiences for deploying Agents in enterprises.

01、Start with Specific Tasks to Identify Suitable Scenarios

The first question enterprises must answer when advancing Agent deployment is: Which task should we start with? Yunnan Baiyao discovered during its exploration that people often harbor the expectation: "I need an assistant that can answer everything about the company." However, further analysis revealed that while these all appear to be Q&A requests, the underlying needs vary significantly. For instance, querying company policies and regulatory requirements demands accurate and consistent answers; whereas querying sales figures in a specific region or proposing subsequent marketing strategies involves dynamic operational data, statistical definitions, and business judgment.

These two types of tasks require different information and have distinct standards for results, so they cannot be treated generically as a single Q&A requirement. Therefore, the first step in scenario selection is breaking down needs into specific tasks: Who will use it? What task needs to be completed? What information is required? And how is the usability of the result judged? Once needs are clarified, priorities must be set. Northeast Securities considers business value, usage frequency, risk levels, and existing system and data infrastructure. They proceed cautiously with high-risk areas involving direct customer interaction or those requiring manual review that is practically difficult to execute; instead, they prioritize mid-to-back-office operational scenarios with high value, broad user coverage, and a certain level of digital foundation.

China Resources Trust selected bond transaction information processing as its entry point. Traders need to monitor a large volume of messages in trading groups, analyze them one by one, extract key transaction elements, and input them into systems. Information processing must be both timely and accurate: omissions or errors can impact subsequent work, and slow processing during market fluctuations may lead to missed opportunities. Centered on this specific task, China Resources Trust built an intelligent bond assistant to aid in transaction information processing. Su Nan, CIO of China Resources Trust, introduced that this assistant operates continuously through three agents, achieving an accuracy rate of 99%.

Guangdong Radio and Television Station's entry point was the segmentation of athlete footage. Previously, staff had to search for materials among vast amounts of video recordings and select shots individually. Now, the station utilizes facial recognition, bib number recognition for athletes, and semantic analysis to assist in locating key athletes, as well as shots featuring goals, cheering, and audience emotions. These capabilities were utilized during the 15th National Games-related work.

Starting from specific tasks like material retrieval, Guangdong TV subsequently built agents for clue collection, material searching, and script proofreading, gradually expanding to other stages of news production. The commonality in these practices is starting with clearly defined work tasks and then determining priorities based on value, risk, and existing conditions.

02、Integrate into Business Processes by Aligning Data, Systems, and Human Roles

After selecting a scenario, making Agents truly part of daily work requires solving three problems: accessing necessary data, connecting with existing systems, and defining which steps require human oversight. In his sharing session, Su Nan, CIO of China Resources Trust, emphasized that agent platforms must be built upon a solid data foundation. If data is not organized, even having a platform makes it difficult to achieve expected outcomes.

In 2021, China Resources Trust initiated data governance, reorganizing tasks and processes around data usage, management, and construction. For trust businesses, the degree of non-standardization varies across different services, leading to differences in data and rules. When building agents, it is essential to clarify where data comes from, what the business definitions are, and how related rules apply. While organizing data, connectivity with existing systems must also be considered. Northeast Securities incorporates structured data from various business systems into a data middle platform, aggregates documents scattered across different systems into a knowledge middle platform, and centrally manages system interfaces available to agents via an AI gateway. Once data interfaces and enterprise knowledge bases are connected to the agent platform, business personnel can call upon these existing capabilities when building applications. However, system integration has boundaries. Xiao Jinwei from Northeast Securities noted that directly connecting agents to core trading systems in the securities industry is currently not realistic. Enterprises need to determine which data and interfaces can be opened and which links require cautious advancement based on system criticality, compliance requirements, and risks. Guangdong Radio and Television Station faces different challenges. Its original news production, non-linear editing, and broadcasting systems are relatively closed, with data and capabilities dispersed across different systems; furthermore, news materials involve sensitive information, imposing high requirements for deployment and security.

To address this, Guangdong TV built an AI middle platform on a private cloud platform, encapsulating capabilities such as speech recognition, video segmentation, and content moderation as foundational services for upper-layer business applications to call, while promoting connection with existing business systems. This allows different applications to utilize common foundational capabilities while addressing data and security issues according to news production workflows. Once integrated into processes, the division of labor between Agents and humans must be clear. In contract reviews, Guangdong TV uses agents to flag clause risks based on preset positions and check for errors/omissions, while legal personnel focus on legal terms and liability judgments; in daily O&M, agents initially filter and analyze problems, while personnel handle further judgment and emergency response. These practices demonstrate that Agent deployment requires alignment with the enterprise's existing data, systems, and job responsibilities. Which information can be used, which actions can be executed, and which results require human confirmation should be clarified in advance for each specific scenario.

03、Engaging the Organization Requires Joint Participation from Business and Technical Personnel

After building the Agent, a practical issue arises: Why would employees want to use it, and can they use it effectively?

Xiao Jinwei from Northeast Securities pointed out that previously, business departments were accustomed to submitting requirements to IT; now, there is a desire for business personnel to participate in construction themselves. If this effort is not directly related to job responsibilities and performance evaluations, why would they invest time? Even if willing to try, business personnel may not be familiar with tools, process design, or effectiveness evaluation. Therefore, Northeast Securities encourages employee participation through management promotion, including large model creativity competitions, agile project creativity competitions, and incentive/performance mechanisms. On the other hand, the IT team provides common capabilities such as platforms, data, models, and plugins to lower the barrier to entry. Training does not aim to cover every platform function from A to Z but focuses on hands-on practice around specific scenarios like knowledge Q&A and document generation, allowing business personnel to start with tasks they are familiar with. Meanwhile, the working style of technical personnel is also changing. Northeast Securities requires application developers to participate more in requirement communication, starting from specific pilots to learn knowledge Q&A, workflow orchestration, and agent construction, thereby gradually enhancing their comprehensive understanding of both business and technology.

Yunnan Baiyao promotes usage from both employee and manager perspectives. For employees, entry points must be direct and convenient, enabling them to find applications when needed. Yunnan Baiyao is planning a unified agent entry point while using internal activities to attract business personnel to build applications themselves. For managers, the key is whether applications solve actual problems and deliver business value. Business departments face assessment pressures; managers are only motivated to promote agents within teams if they see tangible benefits for work. Su Nan, CIO of China Resources Trust, also emphasized that driving AI requires organizational, mechanistic, and resource guarantees, with top leadership support being particularly important. After management sets the direction, business departments and technical teams must still collaborate on specific tasks to ensure applications become part of daily use. Truly engaging the organization requires both employee willingness to participate and access to tools and support that solve real-world problems.

04、Continuous Improvement and Reusability Post-Launch

Launching an Agent is only the first step into business operations. After employees start using it, enterprises must focus on: whether results meet requirements, how to adjust when business rules change, and whether construction experience from one scenario can be applied to others.

In quality inspection applications, the IT team at Northeast Securities collaborates with business departments to sort out rules and formulate test cases, then conducts testing via the platform. Cases include both normal situations and error-prone scenarios, helping technical and business personnel jointly judge whether results meet requirements. This method was first used for WeCom chat quality inspection and then extended to investment advisor marketing material quality inspection using the same approach. Xiao Jinwei introduced that the quality inspection application has gone live, with an overall accuracy rate exceeding 90%.

As application scenarios expand, existing tools and construction methods serve as starting points, but business rules, data permissions, and test cases still need to be reconfirmed based on new tasks. Enterprises also need to continuously update knowledge and rules and adjust application effects based on issues identified during employee usage.