Artificial Intelligence Initiative: Bank of Portugal

Image, audio and visual analysis are being added to a suite of artificial intelligence (AI) functions, which includes suptech and market sentiment warnings

In the past year, the Bank of Portugal has deployed a pioneering AI platform to automate and streamline different areas of operation, from banking supervision to answering complaints.  The platform, called Alya, provides instant analysis on documents and data, using natural language, audio, image and speech processing. 

Alya was officially launched in 2023, but the idea was seeded as early as 2019, says Filipa Lima, deputy director of information systems and technology at the central bank. Core functions of Alya include document summarisation, email response suggestions and market sentiment analysis.

In the past, supervisors at the Bank of Portugal needed to review hundreds of pages in the annual reports published by the banks to meet the Basel Committee on Banking Supervision’s Pillar 3 disclosure requirements. And, since it is impossible to go through every single report, the staff had resorted to sampling.

With Alya, staff can now automate some of the supervisory process using its document analysis functions, enabling more frequent and structured inspections, and enhancing the comprehensiveness of banking supervision.

“So far, it has not been possible to automate the entire rules, but [Alya] has been a major accelerator for certain processes,” Lima tells Central Banking. “But, over time, [staff] would be able to discontinue all they have done manually over all these years.”

Through the market sentiment analysis function, Alya has helped the central bank flag market trends based on information pooled from news and research. Ricardo Carvalho, AI project manager at the central bank, says that this feature had a proven test during last year’s banking crisis.

“Alya could identify that Credit Suisse was at the top of news and research. It can be used to warn of critical situations and centralise the relevant information,” he tells Central Banking.

In terms of information classification, for example, when the central bank receives queries from banking clients about interest rates on savings accounts, Alya would categorise the message and suggest a reply to staff. The model has 70% accuracy for request classification and responses, Lima says.

Alya has also started to classify banking clients’ complaints as well as information requests this year. The development process of this function, according to Lima, was “extremely fast” as the “technology was already there”. The central bank hopes to introduce the classification technology to other departments in the future.

“The big thing for 2025,” Lima says, “is to have a foot in the generative AI world.” To achieve this, the Bank of Portugal has combined Alya with Tia, the Bundesbank’s AI-powered text analysis tool, to launch ‘Tilya’, a prototype that allows users to interact with documents and validate banking supervision Pillar 3 requirements.

Alya would be responsible for preprocessing documents before they are sent to Tia for analysis, according to Tilya’s architecture blueprint shared with Central Banking. This would provide “a structure for the unstructured data” and make results much more accurate, Lima says.

Another new use case being developed is in the validation of financial services’ advertising campaigns to ensure banking clients are exposed to correct and necessary information about products and services. The main objectives are to increase the number of institutions covered, automate the content validation of various advertising formats including video, audio and image, and extend validation to social media to detect non-compliance.