Thought Leaders in Artificial Intelligence: Charlie Delingpole, CEO of ComplyAdvantage (Part 2)

Sramana Mitra: Am I understanding it correctly if I say that you are doing natural language processing on publicly-available news coverage from around the world to identify names of people who are involved in questionable activities?

Charlie Delingpole: That is correct. 

Sramana Mitra: What kind of coverage does that get you? This will identify Osama Bin Laden and the more prominent people, but you quoted that 15 million people are identified as problem makers as well. Is that a sufficiently rich data source to identify 15 million people? 

Charlie Delingpole: The key challenge is the industrial structure of the process. It’s a question about false positives and false negatives. Some FinTech companies that you read about will have 10 million clients. The challenge is when they onboard a million clients in a week.

In the past, you have seen the whole thing with Robin Hood, Wallstreet Bets, and Gamestop. Suddenly, all these companies have to onboard a million new clients. If you are onboarding a million clients, that might mean that you have to assess each person being boarded.

That means that you have to spend 2 hours per match or 10 million clients. That is going to involve a lot of people. You’d rather find the 5,000 you want to reject. These people could belong to the Mexican drug cartels or any other shady group.

It’s a question of industrial efficiency. You don’t want to miss the one person who is high risk. You don’t have a million people with the same name come up as irrelevant matches. The critical thing is to be as perfect as possible because not being perfect is very expensive.  

Sramana Mitra: I’ll come back to the questions that I was asking. What are the data sources? The publicly-available news coverage is one data source. What other signals are you taking into account in determining who is a problem person?

Charlie Delingpole: There are three types of data and there are four categories of data. The three types are sanctions, political exposure, adverse media, and warnings. The three sources of data are structured, unstructured, and semi-structured. You then have other data like corporate data, registries, and the internet.

What we are trying to do is build up profiles of all seven billion people and companies. The unstructured data would be on the DOJ or FCC website where there is a notice talking about how someone is being banned. It could be an article in the press.

Semi-structured data can be on the journal or German regional website where you have a list of people who have been elected and therefore they are politically exposed. The last one is structured data. Every company in the US is warned of a list of roughly 7,000 people who are sanctioned by the US Treasury. They will be fined or put in jail for that.

The trick is to resolve all these entities from all these types of sources into a single person profile. For example, Donald Trump exists from every paper ever released. You would want to understand to the extent of linkages and if the facts of each paper are true and correct. You also want to know if there is another person called Donald Trump in Scotland who isn’t the real person. On top of that, you need to understand their relative and known associates. If I want to launder money, then I am not going to do it under my name because that would be sanctioned.

Therefore, I will try to do that via a relative. If Donald Trump Jr tries to launder money through a friend and we know that this friend is already linked to him, then that person would be banned as well.

What you see is an arms race. If one industry has cracked down upon money laundering and terrorist financing, then it migrates to a new one. Given the technological innovation in terms of online gaming or Bitcoin, it’s a challenging environment for regulators to ensure that bad guys do not get money. 

This segment is part 2 in the series : Thought Leaders in Artificial Intelligence: Charlie Delingpole, CEO of ComplyAdvantage
1 2 3 4 5 6