Credit risk analysis is one of the areas where generative AI has found its first concrete applications in finance. Three major European banks agreed to document their experiences for this investigation.
HSBC: reduced analysis time, not human risk
HSBC deployed an SME loan file analysis system in 2023. The model reads balance sheets, account statements and sector reports to produce a structured summary in under two minutes.
Measured result: average file processing time dropped from 4.2 hours to 1.8 hours. The bank stresses, however, that the final decision remains systematically human. The model does not produce a score; it synthesises.
BNP Paribas: augmented scoring for large corporates
BNP Paribas took a different approach, integrating an LLM into its scoring process for listed companies. The model analyses annual reports, investor call transcripts and economic press coverage to produce a qualitative signal that complements classical financial indicators.
Internal teams describe improved early detection of financial stress signals, with no published figures at this stage.
ING: caution and restricted scope
ING started with a pilot limited to three markets (Netherlands, Belgium, Spain) on an SME portfolio. After 18 months, the findings were more nuanced: the model performs well on sectors it knows well (retail, manufacturing) but produces insufficient analysis on sectors less documented in its training data.
The bank decided to maintain the pilot but not deploy it in full production until it rebuilds its training dataset.
Key takeaway
All three experiences converge on one point: AI is useful as a synthesis and signal-detection tool, not yet as an autonomous decision tool in credit. Training data quality remains the primary limiting factor.