Artificial Intelligence – ‘Star Wars’ automation or something more practical?

If you can cut through the hype, and develop the right use cases, AI can maximize savings and process improvements

As the world bids farewell to Darth Vader, it brings to mind the magnitude of change our little planet has seen since we first met this super-villain in 1977.  It’s opportune to look at a topic that gets as much hype as Star Wars: artificial intelligence.  Stuff of theory in 1977, it’s easy to see today that AI can be valuable to financial services businesses, both now and in the longer term.  But how?

There are specific technologies that can be leveraged, and others that, while interesting, do not provide the stability of an ‘AI’ solution. In our opinion, finding the use cases that maximize cost savings and process improvement while minimizing the most troubling and controversial aspects of AI use should be the starting point for firms looking to leverage AI.  AI promises to revolutionize processes large and small – possibly humanity itself.

Every Star Wars fan knows that in a galaxy far, far away, a humanoid robot was once heard to exclaim

“Oh my goodness! Shut me down. Machines building machines. How perverse.” Star Wars: Episode II – Attack of the Clones – 2002

Nearly two decades later, however, the idea that AI will massively replace human labor remains premature.  Equally, the near-term opportunities AI brings to financial services and related industries remain obscure.  This is our advice on where, and how, to start to make AI work for you.

Just what is AI?

Artificial intelligence is the general concept of machines behaving like humans – C3PO being a perfect example.  Within the overall domain of artificial intelligence is also ‘machine learning’, a narrow application of artificial intelligence in practice.  Machine learning solves specific problems that were traditionally thought not to be automatable – processes that require human thought and decision making.  And within the concept of machine learning lies ‘deep learning’: the use of neural networks as the foundation of learning models.  Deep learning allows the machine to learn based on examples, the way a human learns.

What are we not talking about?

Specifically, RPA – Robotics Process Automation. RPA is a technology that has been in the news as a solution for automating formerly manual processes performed by humans, with some going so far as to include it as an ‘AI’ technology.  Our view regarding RPA is that it is a flexible but ultimately unstable technology, requiring ongoing adjustments as unanticipated changes to the underlying conditions of the process inevitably emerge.  AI, and specifically deep learning, should be able to adapt to changes over time, learning as it goes.  We see deep learning as the preferred technology for building stable automation solutions.

How should AI be applied in business?

There are dozens of success stories that tout the benefits of deep learning – from self driving cars to diagnosis of cancer to facial recognition to beating the world’s best Go player.  In financial services, the lion’s share of the attention is on financial crime and compliance.  Deep learning is used to identify behavior patterns in financial transaction activity and breaks in those patterns, to point investigators toward potentially nefarious activity.   The technology holds strong promise, but faces a long period of regulatory scrutiny before mass adoption is possible.  Concerns with explainability – why did the algorithm identify one set of transactions versus another – will slow benefits realization for the use case.

We see opportunities in less controversial processes by focusing on where people spend time doing repetitive things that have traditionally been difficult to automate.  For example:

  • Financial spreading – credit departments review thousands, even hundreds of thousands of annual reports every year to assess credit quality of clients and potential clients.  Spreading software exists, but getting the information from the annual report into the spreading software is manual.  Banks employ armies of staff with the knowledge required to read and extract key data from the annual reports.
  • Contract understanding – there are times when banks need lawyers to review contract terms for specific protections.  At other times, contract review is for a specific purpose.  In bankruptcy review, contracts are reviewed for acceptance or rejection, based on factors that can be taught to a deep learning set of algorithms.  In lending, terms and conditions in contracts defining the relationship of funds to their investors can be mined to better understand investor behavior under various investor scenarios.  LIBOR transition involves review of contracts to identify clauses subject to renegotiation.  These are all specific use cases where deep learning can be applied to automate that which was formerly not automatable.
  • Rulebook ingestion – virtually all regulatory jurisdictions make their regulations available on line via a website.  PDF documents, HTML pages, text feeds – the formats are varied.  Banks need to keep up with changes in regulations and refresh their processes based on these changes.  A complication of this process is understanding the overlap of regulations across regions or even at times within a region.  Deep learning can be taught to understand the meaning behind the regulations and categorize them for ingestion by the GRC platforms that allow banks to evidence compliance with these regulations.

There is a common thread across these use cases. Taking the unstructured data embedded in documents and structuring that data, then taking action based on this data, can transform the way many processes in financial services are structured and executed.  An evolutionary approach will allow firms to improve the quality of the process as well as to expand the scope over time of what is automated.

How to identify a Use Case

Step 1 – look for teams who spend their time reading documents and taking action from the information in those documents.  Maybe input to a bank system.  Maybe classifying the document itself according to its characteristics.

Step 2 – understand the nature of the activity.  There is a concept of ‘prediction’ versus ‘decision’.

  • Prediction – what is the likely status, outcome, next step?
  • Decision – take action based on the information

In general, look to automate processes that are in place to make a prediction rather than those that require decisions.  At least for now, human decision making will be preferred by regulators, so the opportunity to more quickly achieve the benefits of automation are realized by creating a hybrid team of machine and humans, with the humans focused on decisions.

Step 3 – assess the information contained in the documents for two characteristics:

  • Confidentiality / privacy requirements of the data
  • Specialized knowledge required to complete the process under consideration

The first step in a deep learning implementation project involves gathering relevant documents and annotating to facilitate training of the deep learning models.  This process is heavily dependent on the two characteristics above.  Sionic can help to structure the annotation effort working with partners to ensure appropriate combination of skills and the appropriate confidential platform to ensure success.

Is that where the story ends?

This really is just the beginning. AI is a big topic, and a popular buzzword in financial services just like everywhere else.  But while the hype is real, it makes sense for firms to think about specific opportunities and technologies that lend themselves to early and uncomplicated success.  We regularly help clients improve their understanding of the current state of the market. If you would welcome a real world discussion with a human face, please contact us.

About the author

Ken Bigelow

Managing partner

As an experienced global CIO, I know how to navigate change, negotiate stakeholder needs and deliver results.