Companies like Hummingbird and Babel Street are successfully applying new technologies to improve the efficiency of long-cumbersome aspects of compliance. By doing so, they have created a potent human/technology mix that improves both performance and staff satisfaction.
Hummingbird recently released Automations, a new tool for boosting compliance productivity, reducing risk, and lowering costs. Automations removes manual tasks so companies can deploy staff to higher-value duties. It also improves case monitoring and enforces internal policies.
Activity is centralized on Hummingbird’s financial crimes investigations platform, where customers see company data, workflows, and policies become the components of automation. Practitioners can use pre-built solutions or create their own. Automations offers recipes for KYC, KYB, quality assurance, case preparation, monitoring and management, and activity digests.
Hummingbird founder and CEO Joe Robinson is a fintech veteran who served as a senior product manager at Square and VP of risk and data science at Circle. He said he founded Hummingbird to address the problems he saw with safely introducing more efficiency and automation into investigative work.
Robinson said it is essential to separate fraud from compliance and money laundering. In many fraud cases, the victims are notified by their card issuer, thanks to the institution’s ability to detect pattern anomalies. Money launderers will often avoid committing obvious fraud, as they do not want to draw attention to their actions.
Keeping the human in compliance
When designing a compliance program, Robinson said it’s crucial to prioritize the human element. People have rights to financial services; if misdesigned, automation can infringe upon them. Humans should be kept in the loop to avoid bias and ensure legitimate customers are served.
Many compliance issues are complex, and that means labor-intensive. Robinson said data gathering takes time; data fragmentation in many institutions makes the process more challenging. Checks could include reviewing 12 months of transactions and searches for articles and news about people related to the business from open-source intelligence, social media, and other sources.
“All of that takes time, and it takes data gathering,” Robinson said. “There’s a lot of power to automate the more mundane and tedious parts of that work and let the humans apply what they’re so good at, which is interpreting the results and understanding what’s happened.”
Providing choice and explainability
With Automations, compliance teams can choose which activities the system completes, like data gathering and preparation, reminders, and procedures. They can be rules-based or tap AI models to summarize information. That gives customers the final decision on which algorithms and models to use.
Explainability is an essential aspect of any compliance system. Robinson said any automated system must be auditable, down to the technology used and decisions made. With the largest firms, that explainability must extend across thousands of investigators conducting many more thousands of investigations each week.
How Babel Street strengthened its compliance capability
Director of name screening Greg Pinn said Babel Street’s origins were in using information to mitigate border and homeland security risk. Much like compliance, it involved summarizing reams of data in easily understood formats.
Babel Street widened its scope in late 2022 when it acquired Rosette, a text analytics platform that employs machine learning and deep neural nets to extract meaningful information from unstructured data. That helped with name matching and screening, allowing Babel Street to address unique aspects of names from different languages and cultures. For example, it could make sense of documents for an American citizen travelling on a Chinese passport with an airline ticket from a German flier.
In January 2024, Babel Street added Vertical Knowledge, a data products, global insights and intelligence company specializing in helping customers navigate complex business challenges with a library of contextualized data assets. Pinn said that improves Babel Street’s advanced name screening ability.
Looking beyond AI hype to deliver real value
Amid the AI fervor, Pinn said it’s important to focus on what new problems it can solve. For Pinn, that begins with extracting data from unstructured data and intelligence. In the AML world, that’s an elusive problem.
Screeners face several challenges. When considering unstructured news, such as website articles, it’s been a manual process that doesn’t scale. Structured databases take human capital to update.
“Then you started looking at being able to combine those two things, of creating AI technology and natural language processing to extract information, user identifiable details, and risk information to create a live database of constantly updated risk,” Pinn said. “So you understand who is still at risk. That is a huge leap forward in understanding the riskiness of people worldwide.
“The statistics of people being caught today… are horrible. We don’t do a very good job. So, to me, this is one of the key ways we can improve.”
There have been some concerns about opening the compliance door to technologies like AI, As Robinson stressed, there needs to be a significant human element in the loop.
Where LLMs work and where they don’t
Pinn said around 2018, several regulators united to urge innovators to use technology to improve processes. While LLMs are the shiny new toy, companies shouldn’t necessarily start there. Pinn said tools like Chat GPT aren’t suitable for repetitive compliance tasks, as they are weak at summarizing relevant information.
“Several companies are using these large language models to summarize more articles, but that doesn’t solve the problem,” Pinn said. “It just uses a new technology because you wanted to use it.
“The fundamental problem that AI consultants should be solving is how do you make humans do less work that humans are bad at?”
One example is the high cost of staff screening for false positives. It’s repetitive, with high turnover. That’s ripe for change.
Pinn said there is a place for AI to make better decisions on who and what to screen. Trained models need to accurately assess sentiment while filtering out noise.
Looking ahead, a challenge will be in getting access to data from important companies. Pinn said they create obstructive pricing structures that impact both law enforcement’s and private industry’s ability to use that data to detect new and relevant patterns.
Innovation in UBOs, entity resolution
Pinn said entity resolution is another important area for innovation. New technologies can derive value from unstructured data. AI can help investigators comprehensively view a financial institution’s health. That gives them a more accurate base from which to check for fraud.
AI can also help investigators understand ultimate beneficial owner (UBO) relationships, especially as some governments mandate UBO databases.
The intelligence/compliance blend
Robinson said technologies can help business and regulators cope with a rapidly changing regulatory environment. Criminals are using AI, too, allowing them to quickly pivot when the law catches on to their methods.
One consideration when using technology is to ensure customers get the best intelligence while remaining compliant.
“These models are powerful at looking at broad data sets and summarizing important information,” Robinson said. “We’re trying to develop tool sets that bring the right intelligence and information to them at the right times.”
Robinson said he’s excited about the potential for LLMs to summarize large volumes of information. He said they are good at extracting and summarizing relevant bits of information.
Many in the industry have expressed concern about finding suitably large databases to train LLMs free of noise and false information. Robinson said Hummingbird can help financial institutions with another issue – keeping their models free of personally identifiable information (PII) and ensuring those models don’t leak it either.
Also read:
