IDC: Small and medium financial institutions find industry large model products more cost-effective

On April 7, IDC’s latest report pointed out that most financial institutions (especially mid- and small-sized banks, insurance companies, etc.) use optimized financial-industry large model products, which offer greater advantages in terms of cost, compliance, and efficiency than building from scratch. IDC’s research shows that building a large model from scratch requires heavy investment in technology R&D, data accumulation, and talent reserves, and also involves multiple risks such as compliance review and technical adaptation; it has a long cycle and high barriers, making it less feasible for mid- and small-sized financial institutions with limited resources. At present, companies including Alibaba Cloud, Baidu Intelligent Cloud, Borui Yunchuang, Ant Digital Technology, Qifu Technology, Zhongguancun Kejin, Zhongke WenGe, and others have all launched financial-industry large model products. These can directly connect to business needs, significantly reduce R&D costs and shorten time to go-live, while also avoiding compliance and technical risks during the self-building process, thereby providing various financial institutions with efficient and secure large-model capability support.

According to IDC’s research, for strong (enterprise-level) agent development tools targeted at the financial industry, they should have capability features such as scalability and customizability, large-scale operation and flexible deployment, multi-agent orchestration, compliance and security, continuous monitoring, continuous improvement, a rich plugin/tool system, and integrability. In this field, companies including Alibaba Cloud, Baidu Intelligent Cloud, Borui Yunchuang, Kemairuidai, KELAN Software, Ant Digital Technology, Runhe Software, Sinochem Information, Shidai Yintong, Smaite Software, Yicheng Interaction, Zhongguancun Kejin, and others also provide enterprise-level agent development tools for financial institutions.

Embedding agents within applications is the mainstream form for agent deployment in the financial industry today, mainly due to business system characteristics and compliance requirements. Financial-industry business systems are complex. Core business processes (such as credit approval and risk-control review) have, after long-term iteration, formed a fixed framework, making it difficult to empower business by separating from existing applications and rebuilding them. Embedded agents can empower business without changing the original processes, thereby reducing deployment difficulty. At the same time, compliance and audit requirements in the financial industry are strict, and intelligent capabilities need to achieve “embedding into processes, with controllability and traceability.” In an embedded form, the agent’s entire operational trail and decision logic can be recorded end to end, ensuring that business processes are auditable and risks are controllable—thus both meeting compliance requirements and efficiently enabling the auxiliary role of agents, which is why it has become the current mainstream deployment approach.

There is no unified standard for splitting applications of financial-industry agents. It can be flexibly split in combination with business processes, job responsibilities, or task types. IDC’s observations show that, because different financial institutions differ in business models and organizational structures, the splitting of agent applications does not need to follow fixed rules and should align with each institution’s actual needs. Splitting by business process can achieve end-to-end intelligent enablement—for example, in credit businesses, splitting can separate agents for pre-loan review, in-loan monitoring, and post-loan management. Splitting by job responsibilities can adapt to the needs of different roles—such as agents exclusively for account managers and risk-control specialists. Splitting by task type can focus on the reusability of capabilities—for example, in credit business scenarios, the report-writing agent, information-extraction agent, and mathematical computation agent can be reused across multiple tasks during pre-loan, in-loan, and post-loan stages. Flexible splitting methods can enhance the relevance and efficiency of agent deployment.

Large models are driving finance in various fields to shift toward an RaaS (Results-as-a-Service, RaaS) model, but challenges such as quantifying outcomes and compliance arise during deployment. The multimodal data processing and predictive analytics capabilities of large models can precisely match business needs in areas such as credit, risk control, and marketing, enabling an RaaS model where customers pay based on results—i.e., financial institutions pay based on the actual effectiveness of intelligent services, reducing risks from upfront investment. However, there are clear difficulties in rolling out this model: there is no unified standard for outcome quantification, making it hard to precisely measure the actual value of intelligent services; business outcomes are affected by multiple factors, leading to attribution challenges; responsibility delineation is unclear—if risks occur, it is difficult to distinguish responsibility between the model provider and the financial institution; at the same time, issues such as data security and compliance review also constrain the large-scale deployment of the RaaS model.