Crédit Agricole CIB uses AI to accelerate data and compliance tasks

Sébastien Piednoir

A priority for every bank is to ensure that it is not doing business with companies engaged in illicit activity, such as bribery, money laundering and fraud. Yet compliance and risk teams are facing two mounting challenges: increasing regulatory pressure and exponential data growth. As such, these teams need to deploy new tools to be more efficient in their tasks, such as know your customer, according to Sébastien Piednoir, recently appointed chief compliance officer at CACEIS, the asset servicing group of Crédit Agricole.

In his previous role as chief transformation officer, global compliance, at Crédit Agricole Corporate and Investment Bank (CACIB), Mr Piednoir worked on four projects using artificial intelligence (AI) to accelerate the ability to read information, such as annual reports, and monitor news, as well as transforming unstructured data into structured data. Across these projects, his main aims were to alleviate team pressure and boost productivity, while increasing the security level.

“With the increase in regulatory requirements and reports, our employees are facing an ever-growing amount of data to analyse, so we need to provide tools to support our workforce delivering these tasks,” he says. “We want to help our people make relevant decisions, quicker and more efficiently, with good data at a high accuracy level. They shouldn’t be spending 90% of their time obtaining the data, which is a low added value process.”

Revealing connections

The first project, developed with IBM’s Watson technology, focused on picking out three data points – company, activity and country – in English language financial reports. CACIB wanted to find the relationship between the three points, moving beyond key word searches. Mr Piednoir says: “The tool can, for example, identify if an entity is doing business in a specific country, but also if it is in a joint venture with a partner. Understanding the relation is the real revelation.”

In just 14 weeks, the project’s AI machine could read a 300-page financial report document in five minutes, “with a quality level far higher than what is achievable by a human in that time”, says Mr Piednoir. It also provides traceability. “Even when we find nothing, we have the ability to prove that the financial reports have been read completely,” he adds.

One lesson learned during development is that AI is more of a human project than a technical one. “The key is to have a multi-disciplinary team that understands the technical aspects but, more importantly, understands what the machine needs to learn. There is a role emerging, something I call a ‘machine teacher’, but is technically known as an annotator,” he says.

Mr Piednoir reports initial resistance from some users because the AI machine was less than 100% accurate. “They rejected the tool because it made mistakes. But the quality level that the machine achieves reading 300 pages in minutes is much higher than what a human is able to do, and it is continuously improving,” he says. “Plus, the machine’s interpretation homogeneity might be better in some fields because each person interprets the information differently.” Since delivering the project in April 2019, CACIB has been rolling it out to about 500 users globally.

Quantity and quality

The second project, also developed with IBM, revolved around reading press releases to identify new developments that would link a client with financial crime issues such as money laundering or the financing of terrorism. The project went live in January 2020 and already the compliance team has saved 25% in the time spent providing opinions to the front-office team. “It allows them to focus on the key information in the specific article, which means that they can spend more time analysing the impact of the information,” he says.

Working with voice recognition software Bertin IT, CACIB’s third project developed the ability to transcribe voice recordings in dealing rooms. Mr Piednoir says: “We are recording 500 people and transcribing these every day, so we can search for specific words. It can help us in spotting potential market abuse.”

In conjunction, the fourth project provides the ability to read and interpret the transcripts. It has been developed in-house using open-source technology and is currently in the proof-of-value stage. “Importantly, none of the tools that we are putting into production today allows the machine to make the decision by itself. Everything we are doing [revolves around] ways to help the human user be more efficient, quicker and more accurate in analysing the information,” says Mr Piednoir.

He adds that at this stage there is no plan to reduce the number of compliance staff. “We are trying to deal with the increase in regulations and data,” he says. “If we didn’t use AI technology, we would need to double the number in the compliance team every two to three years, which is not sustainable. Plus, it is not just about quantity but also quality. Just adding more people will not be sufficient to reach our goal.”