← Cases
ChinaFinance and insurance2021 – ongoing

China Great Wall Securities

Privately deployed DeepSeek-R1 and Qwen3 with in-house agents for wealth management, writing and compliance

Great Wall Securities said its investment-banking writing assistant raised drafting efficiency by over 30%

In progressCompany's own claim

Problem

A brokerage handles large volumes of customer calls, compliance checks and document drafting, much of it manual.

Approach

Great Wall Securities set up an AI platform team in 2021, which became its in-house Great Wall Zhiyuan AI Center. The firm runs DeepSeek-R1 671B and Qwen3 on its own servers, on a cluster that includes Huawei Ascend chips. An in-house agent platform supports agents for wealth management, investment-banking drafting, compliance review and call-recording quality checks. A digital-human system handles account opening, and AI generated summaries for 17,000 effective customer calls.

Results

over 30%improvement in document-drafting efficiency from the investment-banking writing assistant, company-reportedCompany's own claim[1]
above 90%compliance review accuracy of AI checks on hundreds of thousands of call recordings, company-reportedCompany's own claim[1]
over 50%share of account openings on the Tonghuashun channel done through the digital-human system, company-reportedCompany's own claim[1]

Takeaway

The firm kept the models on its own hardware and built an agent platform that business teams could use to build their own agents.

Categories

Innovation type
Process innovation
Evidence
Company's own claim

Sources

  1. 以AI为笔,长城证券书写证券数智化新范式
    China Fund News (中国基金报)18 May 2026News article

Source summaries

[1]China Fund News (中国基金报)18 May 2026

A promotional article written by Great Wall Securities staff and carried by China Fund News. It describes the firm's private model deployment and agents, and reports drafting efficiency up over 30%, call-review accuracy above 90% and over 50% of Tonghuashun account openings through a digital human. The figures are the firm's own and the article gives no method.

Full details are in the original sources.

Added 1 Nov 2026 · Updated 1 Nov 2026
Open case page ↗