For Financial Services firms, AI technology is increasingly a must-have for the growing digital economy. Firms of all types are increasing AI’s use to further improve productivity and to increase operating efficiency and revenue-per-customer.

In today’s world, one digital transaction alone provides significantly more insights about a buyer’s preferences when compared to cash-based transaction; additional data may include the buyer’s search history, preferences, and more.  Collectively, this data can enable firms to more quickly identify significant patterns at a sector, company, group, and/or individual level.

Moving to AI

AI, and especially natural language processing (NLP), can significantly improve financial services customer experience, offerings, and services. NLP is one of the fasted growing types of AI and has applications across the financial services value chain such as these:

Helping customers with virtual assistants

The ability of NLP to process and interpret human language is transforming financial services self-service support and investment management. Bank of America offers Erica, an intelligent virtual assistant, to provide financial recommendations to its 45 million-plus customers.

Speeding credit agreement reviews

An average reader can read 200 – 300 words per minute (wpm), understanding approximately 68% of the content¹.   For specialized content e.g. a legal document, comprehension levels and/or wpm may be significantly lower.  Manual review of 12,000 annual commercial credit agreements normally requires approximately 360,000 hours³.  AI/NLP can process the same number in seconds.

Improving customer satisfaction and sentiment

Analyzing social media, chat, email, and online activity enables financial services firms to improve customer satisfaction and to deepen relationships.  Negative public feedback significantly impacts revenue.  Businesses risk losing as many as 22% of potential customers from one negative posting alone, and up to 59% when there are three postings.  [Source: moz.com]

For analyzing social data including social post content, AI offers multiple sentiment analysis solutions that can quickly analyze data and gauge feelings regarding a subject matter, a company, a product, or a person.² Firms can then take the appropriate action.

Increasing fraud-detection accuracy

AI can be especially useful when time-to-action is short and high accuracy is required as in analyzing transactions and communications for real-time fraud detection. Traditional rule-based analytics have proven to be too simplistic in fraud detection and DL/AI offers greater accuracy, speed, and scalability.  This has led some banks and insurers to explore Deep Learning (DL)/AI’s effectiveness in reducing level of false positives generated by rule-based fraud analytics.

Re-architecting the datacenter for DL/AI

Over the past years, financial services firms have been eliminating data silos and building big data lakes on-premises to manage growing big data volumes. Early big data developers chose Hadoop/MapReduce to analyze their big data, as well as run DL. More recent big data developers are using Apache Spark, a proven faster alternative to MapReduce.

A DL/AI system built on generic Hadoop infrastructure that has scalability limitations which reduces economic benefits and makes it ill-suited for cloud-based computing and/or scale-out deployment.

The default Hadoop Distributed File System (HDFS) has a default replication scheme that requires storage space equal to three times the actual size of the data. Adding storage capacity in HDFS increases space, power, administrative cost and compute capacity.  Increasing compute capacity in turn requires purchasing additional Hadoop licenses.

Commercial versions of the generic Hadoop offering include a default YARN workload manager. Spark workloads running on YARN have exhibited a long-tail time-to-completion profile in a multi-tenant environment for those jobs that require high throughput which is likely problematic for DL.

Software defined DL/AI infrastructure

Just as software defined computing – which decoupled workloads from the underlying hardware – enabled large scale, quantitative risk analytics, a software-defined version of the deep learning technology stack can enable large scale DL/AI in Financial Services.

By decoupling the DL/AI technology stack’s layers – services & support, resource management & orchestration, native services, workloads, and Spark/Hadoop – users gain greater granularity of control, enabling cost vs performance flexibility.

Multiple applications and instances of an application can now share compute resources for greater economy and scalability. The end-user can now choose the storage/file system that meets their needs, avoiding HDFS redundancy and copy data requirements. Firms then have the choice of running calculations where it makes sense –  on-premises, on the cloud or a combination of both.

With a software defined approach with support for multi-tenancy, fine-grain resource management, and open source APIs, IT can create a highly scalable, well-behaved, distributed deep learning environment to support digital interactions.

Learn more about IBM software defined solutions for AI and DL here.