Bank of Jiangsu
Bank of Jiangsu uses DeepSeek models to automate contract checks, email sorting and reconciliation
Bank of Jiangsu, a regional Chinese lender, says it fine-tuned DeepSeek-VL2 and DeepSeek-R1 for contract quality inspection, email classification, product matching, transaction entry and valuation reconciliation. Press reports relay the bank's figure of a recognition success rate above 90 percent.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
over 90 percent
Recognition success rate stated in the sentence after the description of DeepSeek-R1 automating email classification, product matching, transaction entry and valuation reconciliation; Yicai does not define it1
thousands
Emails the bank receives each day that DeepSeek helps sort and respond to2
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Banks handle large volumes of contracts, emails and reconciliation statements, and much of this work is routine and manual. Chinese lenders were looking for low-cost ways to use large language models after DeepSeek's open models appeared in early 2025.
What they built
Bank of Jiangsu, based in Nanjing, says it deployed and fine-tuned two DeepSeek models. DeepSeek-VL2 is a multimodal model, which means it reads images as well as text, and the bank uses it for contract quality inspection. DeepSeek-R1 is a reasoning model, and the bank says it automates email classification, product matching, entry of transactions and reconciliation of valuation statements. Global Finance adds that the bank uses DeepSeek to sort and answer thousands of emails received each day and to analyse large amounts of financial data.
How they did it
The sources give few details. Both DeepSeek models are open-weight, meaning their files are published and can be run on a company's own systems. Yicai describes the bank as having fine-tuned them, which means further training on its own material. The articles do not describe the team, the cost or the rollout steps.
Results
According to data from the bank cited by Yicai, the recognition success rate is over 90 percent. The articles do not say what was measured or against what baseline. Global Finance reports that the bank has the models sorting and replying to thousands of emails daily. Yicai notes that, across Chinese banks, such assistants are mostly used to support staff and are rarely used for core account transactions.
Lessons for your project
Start with back-office document work that people check anyway, such as contracts, emails and reconciliations, where an error is caught before it reaches a customer.
Open-weight models can be fine-tuned and run inside the bank, which suits firms that want to keep data local.
Keep the model away from core transaction processing at first. Industry experts quoted in the coverage say output is hard to control and content is still reviewed by hand.
Define the success measure before launch. A figure like a recognition rate needs a clear definition of what is being recognised.
Caveats
- The deployment and figure come from the bank and are relayed by journalists, so the evidence is company-reported. No bank primary source was found, and no independent audit exists. Yicai does not say outright which task the success rate covers.
- The sources do not define what the recognition success rate measures, or how it was tested.
- The bank's deployment date is not stated. The first source, dated 10 February 2025, says it has deployed the models, so deployedDate is the month of that report.
- Neither source gives costs, team size or timelines. The Yicai article says AI assistants are mostly used for content that is then manually reviewed and are rarely used in core banking transactions.
- Both articles list other banks, including ICBC and Postal Savings Bank, that adopted DeepSeek, without figures.
Sources
- China's Financial Sector Hastens Integration of DeepSeek LLMs
Yicai Global · 2025-02-10 · News · www2.yicaiglobal.com
- DeepSeek AI Takes Root in China's Banks
Global Finance · 2025-03-01 · News · gfmag.com
Summarized in our own words from the sources above. Spotted an error? See corrections.