Industrial and Commercial Bank of China (ICBC)
ICBC says it rolled DeepSeek out across the bank in 200+ scenarios
ICBC, one of the world's largest banks, says it deployed DeepSeek's open-source reasoning model on its own infrastructure and connected it to its in-house large-model platform. It reports use in 200+ scenarios across 20+ business areas, inquiry-trading efficiency up by 3 times (the Chinese wording 提升3倍, with no baseline given) and about 10% shorter calls in key remote-banking scenarios.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
提升3倍 (improved by 3 times)
Efficiency of inquiry trading between head office and branches with the ChatDealing product, per ICBC1
压降约10% (reduced by about 10%)
Call duration in key scenarios with the remote-banking agent assistant, per ICBC1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Large banks want generative AI across many jobs, from trading desks to call centres and credit approval. ICBC says it chose to bring DeepSeek's latest open-source model into its own systems and use it across the bank.
What they built
ICBC says it completed a private deployment of DeepSeek's latest open-source model, meaning the model runs on the bank's own systems rather than a public service. It is connected to the bank's large-model platform, called 工银智涌 (Gongyin Zhiyong). ICBC says the models now support 20+ business areas and 200+ scenarios. Examples it names are ChatDealing, a conversational trading product for financial markets, a call-centre assistant for remote-banking agents, and a risk-control assistant for corporate credit approval.
How they did it
An industry researcher quoted in the article attributes the speed to ICBC's depth in financial technology and its earlier investment. ICBC says it built an in-house computing cloud of a thousand-card scale, and says the DeepSeek model sits next to ten-plus large models and more than 2,000 traditional models already in use. It does not describe rollout steps, costs or staffing.
Results
ICBC says ChatDealing improved inquiry-trading efficiency across head office and branches by 3 times (提升3倍), a wording that gives no baseline. It says the remote-banking assistant cut call duration in key scenarios by about 10%. It reports no results for the credit-approval assistant. All numbers are the bank's own, relayed by the news outlet without independent verification.
Lessons for your project
Run a model on your own infrastructure when data sensitivity requires it. ICBC chose a private deployment of an open model rather than a public service.
Connect new models to an existing platform. ICBC plugged DeepSeek into its in-house model platform instead of starting separate projects.
Pick two or three measurable use cases to report. ICBC cites trading efficiency and call length, which are easier to track than a bank-wide claim. Ask for the baseline behind any multiplier.
Plan for the people it takes. A researcher quoted in the article estimates ICBC's financial-technology staff at more than 10,000 and links that to how fast it moved.
Caveats
- All figures are statements by an unnamed ICBC representative to Cailian Press. The article does not report any independent check of the results.
- The article presents the figures as examples of applications on ICBC's large-model platform that now includes DeepSeek. It does not say how much of each result comes from DeepSeek rather than ICBC's other models.
- The article says ICBC deployed the latest DeepSeek open-source model without naming a version.
- The Chinese wording for the trading result (提升3倍) literally means 'improved by 3 times'. Lumiary reports it as the source words it and does not convert it to a multiple of the base; the article gives no baseline.
- The article gives no cost, team or timeline for the rollout. The deployed date shown is the article month.
- The 10,000 figure for ICBC's technology staff is an industry estimate in the article, not an ICBC statement.
Sources
- 一线|“宇宙行”放大招，工商银行宣布率先完成DeepSeeK全行大规模应用，会有更多银行跟进吗？
Cailian Press (财联社) · 2025-03-08 · News · cls.cn
Summarized in our own words from the sources above. Spotted an error? See corrections.