Charlie Henderson is on a mission – to free businesses from the daily interruption of non-core tasks. Together with business partner Lucas Wurfbain he has spent the last few years figuring out how to structure enterprise activity data in order to identify hidden risks and maximise value using AI. The result? FeedStock, a firm founded in 2015 which utilises natural language processing (NLP) technologies to analyse communication streams, helping complex businesses to track, filter and interpret the data and digital communications that define their relationships.
“We both worked in financial services, and we were constantly having our daily jobs interrupted by multiple manual tasks, such as manually logging interactions for both compliance and commercial reasons,” Henderson explains to RegTech Insight. “So we took a step back and said, “What are the businesses trying to achieve here? How can we automate this?” And then we did.”
The partners created a data capture platform that analyses human interaction and maps workplace relationships in real-time across investment banks and asset managers initially, principally focusing on MiFID II compliance. The goal is to help businesses understand the value of research relationships and the cost of sales – and how those two elements progress and interact.
The firm offers two core products. Cortex uses machine learning to automatically track research interactions and identify research inducements across communication streams, analysing research interactions for asset managers automatically in the background. It plugs into emails, for example, and automatically classifies them using NLP to track conversations – allowing users to monitor and manage their inducement risks for a regulatory perspective, and also to refine and streamline their research relationships in the light of MiFID II.
It also offers the same product essentially in reverse through its Synapse solution, which targets investment banks, automatically classifying their relationship with asset managers so that they can bill them correctly.
The advantage of AI – and particularly NLP – for this type of work task classification is consistency. A frequent issue for both asset managers and investment banks when it comes to manual interactions is the quality of compliance and commerce – the interpretation of when (and what) something should be classified. “What we found is the manual data capture solutions that exist today are unfortunately often very subjective. They have recency bias, and they have huge variation on what is defined as an actual interaction,” explains Henderson. “We see that a good 30-40%, even 50% of interactions, are sometimes completely missed and not even logged. The benefit of NLP in capturing these is obvious – we’ve actually been blown away about how many more interactions a week we capture using NLP based approach.”
This gives an obvious commercial edge – asset managers that have more consistent data about their research relationships are going to benefit compared to those who are still relying on manually logged interactions.
“To give you an idea of the scale of the challenge, if you’re a fund manager and you have 100 research relationships, that’s 100 providers, with perhaps 10 analysts for each. That’s 1,000 people, providing say 20 research interactions a week. That is 20,000 interactions that needs to be logged consistently per week. It is a Herculean task, and arguably never going to happen.”
On the sell-side, it’s the same problem in reverse. How much time is going into servicing each client? What is the cost of service to that client? Are you accurately logging these interactions to determine whether you’re going to make a profit?Can you replicate that same digital footprint with another client to help upsell?
“The data has all the answers and it’s just a new way of looking at it,” thinks Henderson. “And thanks to NLP, you can do it now – whereas four years ago, even three years ago, you couldn’t really get accurate enough NLP models to surpass the historical roads of manual interaction for capturing.”
But the market is changing all the time, and technology needs to keep pace. FeedStock spent almost five years doing R&D for NLP for its specific tasks, revising and updating its models numerous times. “In an NLP world, it’s all about product development. You have to constantly try out new models and new methodologies to get the best results,” says Henderson.
“We’ve moved from this siloed manual data entry system to a living, breathing organism – and that is the way the world is moving. Data can now help you make commercial-based decisions on a real-time basis, across your whole enterprise, irrespective of what time of day it is, what day of the week it is or where you’re sitting at that moment.”
The firm raised £2.5 million in funding towards the end of 2019, and has contracted with some of the world’s biggest financial institutions to implement its process. “We’re seeing continued demand from both sides to really refine additional use cases out of the data that perhaps we hadn’t fully appreciated with such considerable problems,” reveals Henderson. “We’re using that capital to continue with our R&D, to continue to service these considerable corporates and grow our team in line with the requirements.”
Going forward, FeedStock hopes to work with clients on a more holistic view of relationship management.
“We’re already talking to some of the other large banks about operational efficiencies and how we can improve the interactions between their suppliers and their internal actors, as well as their front office. Understanding the value that data can provide around streamlining all operations is just the beginning.”