There is a lot of buzz around artificial intelligence (AI) and machine learning (ML) and their application to financial services; specific examples include use cases applied to chatbots, robo advice, and credit risk scoring. Almost daily, articles are published focusing on specific functional areas, but do not address the potential of machine learning from an end to end perspective. This is the first of two lists whereby I document use cases by process area and potential use of ML algorithms and techniques. This initial list outlines AI and ML application to client on-boarding processes for new clients or for new products.
Client on-boarding is usually quite a cumbersome process as it involves a lot of information gathering, filling up questionnaires, executing various checks on the client identity and background to ensure adherence to KYC and anti-money laundering regulations. A number of processes can be automated (black list checks, white list checks, name matching, digital on-boarding, digital identification), but this is based on standard algorithms that allow for (robotic) process automation. However, there is no learning involved in these processes and the algorithms will stay unchanged even as more data is collected over time. These processes are not really AI/ML and should not be confused as such. Below are some areas where AI/ML can go beyond these automation processes:
There is a number of processes that can use ML techniques to detect fraudulent attempts, usually based on historical attempts. These techniques can detect anomalies or fraudulent patterns in the client identification process through nearest neighbors’ algorithms or other classification algorithms (i.e. support vector machines, decision trees or logistic regression). Using a set of existing customer data, along with non-compliant cases, can help predict and flag negative cases and the model can learn as new cases are added.
Client segmentation and profiling can also be automated using predictive techniques that will use existing client data information classify newly on-boarded customers into matching customer segments or risk profiles. For this process, we can use unsupervised learning techniques to detect customer patterns and segments or supervised learning to classify clients in pre-defined segment or risk profiles based on nearest neighbor’s algorithms or decision trees.
- Client identification can be enhanced using facial or voice recognition and deep learning techniques (for example, convolutional neural network) to compare data with external identity databases and spot fraudulent attempts. This data can then used later for processes such as transaction authorization, a technique already used for payments.
Customer Experience Enhancement
Natural language processing techniques can be used to enhance the customer experience when integrated into chat bots. Bots gather the required information usually done by hand by a relationship manager and can greatly improve the customer experience while reducing administrative burden. Bots are not currently advanced enough to support all processes, but the development of solid models will support more complex bot processes in the future. More can be read for example under http://www.wildml.com/2016/04/deep-learning-for-chatbots-part-1-introduction/
- Finer client on-boarding data can also be enhanced by information collected via social networks to identify customer risks or preferences. However, this pursuit would require financial institutions to partner with external parties like Facebook and Google to have a better view of the incoming customer and be able to offer a much smoother experience. The challenge remains the willingness of potential clients to share their personal information across multiple suppliers.
When clients want to buy specific investment products, advisors must adhere to suitability rules and product risks must match the client risk profile. The process of matching investment products to a client’s investment profile is usually already automated using rules engines or other automated process, but machine learning can help.
Supervised learning (for example, decision trees) or unsupervised learning algorithms (nearest neighbor clustering) can verify if the particular product is well suited for the client based on previous client on-boarding patterns and if similar products have been bought by clients in similar groups. This process can greatly improve the customer experience as it gives the customer an increased level of trust in the recommendation made by the relationship manager (or by the automated sales process).
In the second list (to be released in early January), I address sales and advisory processes. Stay tuned!!!
– Text by Patrick Rotzetter
Patrick Rotzetter is an IT director who has acquired more than 20 years of experience of leading and delivering complex change projects in digital banking, asset and wealth management in Europe, US and Asia. He translates strategy into actions and ensures delivery using strategic and creative thinking.