AI could cut time for money laundering checks by 99%

A leading cryptocurrency exchange has built a large language model (LLM) that its risk managers believe will cut the time for basic anti-money laundering (AML) checks by as much as 99%.

Speaking at a webinar on June 28, Ian Rooney, the head of enterprise compliance at Coinbase, said the company planned to roll out the LLM tool for enhanced due diligence (EDD) checks within the next month. He explained that it could reduce tasks “that we estimate, on average, would take an analyst about an hour-and-a-half of work down to, on average, under a minute to get to the same starting point for the substantive analysis”.

What we’ve done is land on a formula or a model where we have a high degree—95%-plus—of confidence in the output that we’re getting back

Ian Rooney, Coinbase

Artificial intelligence (AI) is proving particularly useful for combatting financial crime by automating previously labor-intensive tasks, and reducing the number of transactions and clients incorrectly flagged as suspect. This allows staff to focus on a smaller number of genuinely suspicious activities.

Dutch digital lender bunq fought a court case in October 2022 to justify its use of the technology for AML, and HSBC has recently indicated that it was intensifying its adoption of AI in this area.

In the first instance, Coinbase intends its LLM to work alongside analysts who will ensure the models are performing as expected.

“We are at a point where we’re confident and comfortable taking it from just an internal pilot and assessing outputs to actually using it in a production environment,” said Rooney.

Compare and contrast

The EDD process is a regulatory requirement. Financial services firms are expected to look into a customer’s credentials more deeply if they are identified as a politically exposed person (PEP), linked to high-risk countries or at risk of involvement in money laundering activities.

Although this is not the whole AML journey, it is a large and important part of staying compliant. Rooney said Coinbase started out by writing prompts for a pre-existing vendor model, such as those developed by OpenAI. The base model already existed, and from there, the analysts began to engineer prompts in a closed environment.

“What we’ve done is land on a formula or a model where we have a high degree—95%-plus—of confidence in the output that we’re getting back,” he said.

Getting an accurate and repeatable output is one of the major challenges with LLMs, since it can be difficult to explain what causes a specific response. Regulators have stressed this lack of explainability is one of the significant worries around letting AI play a more active role in financial services.

Coinbase’s model works by inputting a customer identity, and then letting the LLM run in the background, carrying out the checks and analysis usually left to a human. Rooney explained that, in an ideal state, they would not even need a human to input the customer ID.

In the roll-out, staff would undertake side-by-side comparisons of the model’s work compared with checks by EDD analysts on the same customer. This would allow a comparison of the traditional method against the new AI models. Given positive results, Coinbase aims to ramp up the building of its model.

Next steps

The crypto exchange does not plan on limiting its AI development to EDD checks. For example, if the program is successful, Rooney mentioned combining the pattern detection of conventional AI with the insights provided by LLMs. His idea would be for LLMs to explain the patterns spotted by AI in a way that human analysts can understand.

“How can you assist or co-pilot an analyst, not just on the EDD historical transaction information, but providing some insights about where the transaction behavior may be concerning, and where the analyst should start their analysis or investigation?” said Rooney.

Coinbase also hopes that the extra information available to them through blockchain technology will mean a higher-performing AI. When it comes to creating AI models and LLMs, extra data quality means better training sets. Throughout the process, Rooney said Coinbase had been co-operating with regulators, for example, through meeting certain key performance indicators.

As AI sweeps the financial services world, some may be worried about where this leaves employees. Rooney did not rule out the possibility that many AML roles might be replaced by full automation, but doubted this would happen in the near term.

“We’re not at a point where we’re thinking about letting any outcome be handled solely by a machine learning or AI solution, [but] I think that is not an unrealistic possibility in some distant future,” he said.

Editing by Philip Alexander