Credit Agricole Bank Polska
Credit Agricole Bank Polska runs Poland's open Bielik model on its own servers for staff tools
Credit Agricole Bank Polska picked Bielik, a Polish open-weight language model, after an internal AI hackathon, and bought GPU servers to run it on its own premises. It now powers a fine-tuned plain-language text corrector, polCA, and an internal chat that every employee can use. The bank says it has no efficiency KPIs yet.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
ponad 50 (more than 50)
Project submissions received at the bank's 2024 AI hackathon1
pięć (five)
Hackathon ideas selected for implementation1
What it took
The problem
Credit Agricole Bank Polska wanted to start working with large generative models, but access to tools such as ChatGPT was restricted, as is normal in banking. Only a few people, including its Innovation Lab, could use them, and under control. The bank also had no infrastructure of its own for running language models. It wanted a way to let more staff use generative AI safely, without sensitive data leaving the bank.
What they built
The bank uses Bielik, a Polish language model that is open-weight, meaning its model files can be downloaded and run on your own servers. Two staff tools run on it. polCA is a text corrector that rewrites texts to follow plain-language rules. The second is an internal chat that any employee can use to talk to the model or to work on documents. Bielik also powers tools that remove or mask sensitive data before text is sent to a cloud model, and tools that help marketing and HR draft content to the bank's internal standards from a short draft.
How they did it
The trigger was a 2024 AI hackathon held with a technology partner, Deviniti. The bank needed an engine to power the participants' ideas. The hackathon drew more than 50 submissions and the bank chose five to implement. polCA, one of them, became the first production use. It was kept deliberately simple, tested against other models, and fine-tuned on the bank's own text corpus, which means the model was trained further on the bank's own data to suit its needs.
The hackathon and polCA led the bank to buy GPU servers and install Bielik on them, after which it opened the internal chat to everyone. New hardware arrived only at the end of 2024, which the bank names as its largest hurdle. It now also runs agent-building platforms (n8n and RAGFlow) powered by Bielik and other models, and ran a second hackathon in 2025 on agents.
Results
The model runs on the bank's own machines, so sensitive data stays on its infrastructure. polCA is in production and every employee can use the Bielik chat. The bank says it measures the number of texts corrected or generated, agent runs and data processed, and that these show the tools let staff do more tasks. It states it does not yet have precise indicators of percentage time saved, and is defining KPIs as part of a wider approach to running AI at scale.
Lessons for your project
Use a hackathon to find the first use cases and the first model at the same time. Credit Agricole's 2024 hackathon produced over 50 ideas and the case for a shared model.
Start with a simple first production tool. The bank chose polCA, a text corrector, because it was easy to prove, then took small steps to the next projects.
Compare models on your own task. The bank tested several, and polCA worked best on Bielik after fine-tuning on the bank's own text corpus.
Give staff a safe internal chat so they learn what language models can and cannot do. The bank says Bielik is weaker than much larger models but good enough for that.
Plan for hardware lead time. The bank names getting and installing the GPU servers as its biggest challenge.
Build measurement in early. The bank counts texts corrected and agent runs but has no efficiency KPIs yet, and says defining them is part of its next phase.
Caveats
- The only source is an interview with the head of the bank's Innovation Lab, published by Poland's state development fund PFR. All claims are the bank's own and no independent source confirms them.
- The bank states that it does not yet have precise indicators of how much work time or productivity has changed. The numbers above are counts of ideas, not business results.
- The article gives no go-live dates for polCA or the internal chat. The deployed date shown is the year of the hackathon and of the GPU server delivery (end of 2024); the article does not say when polCA or the chat reached users, which may have been after 2024. The year comes from the text and no month is given.
- The article gives no cost, no number of servers and no number of chat users.
- The bank also uses other models, including Qwen and gpt-oss, in its agent platforms, and says Bielik gave way to other models for some tasks.
- The interview page blocks automated readers; we read it through an archived copy.
Sources
- Start z polskim AI – od pomysłu do wdrożenia
Polski Fundusz Rozwoju (PFR) · 2026-02-26 · Interview · pfr.pl
Summarized in our own words from the sources above. Spotted an error? See corrections.