Riding the Artificial Intelligence Wave Requires Balancing Benefits with Risks

The AI wave is coming. Whether we believe it or maybe are more skeptical about it, we are exposed to AI messaging & initiatives popping up all around us and the pressure is only getting stronger.

“It is not a matter of if, but when - there is the need for us to be fast while the world is changing fast

According to a number of surveys, Artificial Intelligence comes with substantial benefits and the expectations are very high. But do these come for free?

For those institutions that are prepared, riding the AI wave will bring them a lot of advantages. However, a prerequisite here is they have a proper balance of the right skills, methodology, approach, tools, and proper training.

There is a number of institutions out there who are still maybe a bit hesitating, not yet ready, and jumping on the wave without proper preparation might turn out to be quite dangerous. Also waiting and doing nothing is not a very good option, because then we can get overrun by this wave and things can get out of our control quickly.

AI definition & motivation

For purpose of this article, let's use the definition of AI and motivation of why AI matters that was put together by the high-level expert group on AI set up by the European Commission (above). The significant increase of expected European spending on AI over the next years is a concrete manifestation of the AI wave approaching.

I would like to point out that the AI definition put up here is broader than just advanced analytics (machine learning, deep learning...), which some more analytical people might find a bit surprising. 

As a matter of fact, in a recent industry generic AI global executives survey, 60% percent of the executives consider analytics to play only a moderate, minor or no role at all in AI. Only 16% consider analytics to have a central role in AI and this would include mostly the executives responsible for AI programs. Thus there is quite a misaligned perception out there about what AI is.

But what are companies actually doing with AI now? In the same survey, we also inventoried the use cases where the companies are deploying & leveraging AI (see below).

AI deployment functional areas - Industry Generic

Source: The AI global executives survey, SAS & Accenture & Intel & Forbes

As this was an industry generic survey, we can see the classical operational & back-office areas which are present in every industry dominating the AI landscape. The more industry-specific we go the more difficult it is to deploy AI as the right mixture of industry-specific know-how & technology and the AI expertise & skills is required.

Let's have a closer look at some more concrete AI use cases.

AI deployment functional areas - Banking & Risk Management

In a joined survey with the Global Association of Risk Professional, we asked bankers around the world on their experiences with AI in risk management. Below on the left side, we can see the overview of risk use cases, ranked by the current AI adoption rates, with process automation being number 1, where already more than 50% of participants are deploying & leveraging AI. This is closely followed by tradition analytical domain of Credit Scoring.

Source: AI in Banking & Risk Management, SAS & GARP

Overview on the right shows the same use cases but now ranked according to the "expected the adoption rate within 3 years. This is the AI wave coming, as we can see a significant increase of the AI adoption rates, starting from the top with a shift from the 50% area to 70% areas. Even more impressive are the shifts at the bottom, where AI adoption rate more than doubled in the number of the currently "more conservative" AI use case areas.

AI Benefits & Challenges

80% of the current banking adopters of AI already clearly see the benefits. However, the AI benefits do not come for free, and there is a number of challenges & potential pitfalls identified in the survey that banks have to resolve first in order to be successful.

Source: AI in Banking & Risk Management, SAS & GARP

These challenges, represent the risks introduced or emphasized by AI, the "AI risks".

Within banks, there are particular units that have become increasingly vocal about these challenges and are now calling for a more cautious approach towards AI. These are the Model Risk Management & Model Governance professionals whose job is to protect the bank from any losses caused by the usage of improper & inaccurate models.

AI and Model Risk Management

Model Risk Management (MRM) has been around in various forms for quite some time, but over the last few years, as the numbers of models are increasing so does the attention to MRM from both the regulatory & business side. Areas such as bad modeling datasets, manual handshakes between modeling and validation, validation and deployment, organizational silos and lack of governance, all these challenges they have to address already now, even without AI (see our earlier article on Enterprise Model Risk Management for more information).

From their MRM perspective, AI will just add more fuel to the existing burning points within the modeling ecosystem and make them manifest much more dominantly.

Source: SAS Model Risk Management Customer Connection, survey

According to 60% of MRM professional that attended our MRM customer connection event, AI models will make their work more difficult. And this is very understandable because the MRM teams will be the first ones to blame if something goes wrong with the AI models.

Important is to note here that we do not talk only about regulatory models here, 80% of the MRM teams already have either all models or at least all risk models in their inventory which they need to manage and supervise. And within 1-2 years, 90% of the MRM teams expect to have all models in, thus including any AI models deployed within the bank.

This is important to realize here, all AI models, not just the ones used for regulatory purposes, will end up in bank's model inventory and will need to follow its model governance and thus will end up on the table of the MRM teams.

Stepping out of the banking industry for a moment, the same direction can be observed also in the guidelines for ethical & trustworthy AI developed the EC's AI expert group.

The number of principles and compliance checklists (see example below) proposed in this document resemble the scrutiny applied by the banking regulators to the banking regulatory models. But there is a big difference here, these guidelines below apply across all industries irrespective of the functional use case area and are not restricted only to the regulated areas.

Source: European Commission's Ethics Guidelines for Trustworthy AI

Potential Cures for dealing with the AI Risks

The AI risks & challenges can be clustered into the following 5 main categories each having a different impact and requiring a different cure.

Source: AI in Banking & Risk Management, SAS & GARP

Below we provide a few high-level ideas & tips that can be considered to address the AI challenges above. More details can be found by following the links at the end of this article.

White Boxing the AI

How do we improve our ability to explain the complex Neural Network and other machine learning models?

• Introducing a Visual interpretation of modeling logic 
• Establishing a central model management framework which will prohibit the creation of multiple & parallel AI islands within the organization, each doing their own thing. 
• Leveraging the high-performance capabilities to automatically perform and reperform variable importance and sensitivity analysis (LIME & ICE) and "stress test" the inputs of AI models. 
• Similarly, we should also perform frequent benchmarking and champion & challenger model comparison.

Healthy Feeding the AI

AI technology is data hungry and how do we ensure that we feed it with healthy data?

• The preparation and management of modeling data should form an integral part of the end-to-end modeling landscape 
• We should leverage the latest technology trends such as automation, distributed data storage and support all the formats & structures needed by the AI models

Jumpstarting the AI

How do we decrease the time and costs needed to reach the AI benefits?

• Always start with a business problem that you want to solve before thinking about methodology/technology 
• Starts small - but do that with an eye towards the scale so once you adopt it at a small scale so you can roll it out enterprise-wide 
• Reuse existing resources, both human and technical ones. Enable existing risk talent rather than solely relying on hiring dedicated data scientists. 
• Establish one way of working enterprise-wide - both on process & technology level

Humanizing the AI

How do we attract & retain talent and make people understand what we are doing with AI? To be successful, we need the right people to design & run the AI, and we need to make the rest of the people & aware of what we are doing and what the AI is doing to them.

On one side, we might be struggling with getting the right resources in, and here concepts like citizen data scientist, open source embracement, standardization, and automation can help to ensure that we have resources available and interested to work for us and that they focus more on value-adding activities rather than "boring" but necessary activities.

On the other hand, promoting transparency, building awareness about AI and about Model Risk Management can help us to build trust and give us the support that we need to push & establish AI successfully within our organization.

Operationalizing AI

How do we move the AI ideas from our drawing boards and the heads of data scientists into the daily operations? Here technology will play a crucial role.

Here we need to ask ourselves how does my AI technology platform help me to:

• Automate the existing labor-intensive manual processes within the end-to-end modeling lifecycle 
• Ensure the proper performance to leverage the latest trends (GPUs, containers, EDGE / Cloud deployment, LIME & ICE model interpretation frameworks..) 
• Centralize model governance and apply to across all models (not just AI) across the entire modeling lifecycle 
• Integrate with the existing systems and how can it be reused for other purposes?

What is next?

AI Wave is coming, and it is coming fast. I think it is only fair to assume, that we will also see much more attention given to the management of the model risk arising from the usage of AI models. Irrespective whether it will come from internal and/or external stakeholders, institutions should already proactively establish their MRM processes around the AI models. This will ensure that:

• they can proceed with their AI initiatives in an efficient & governed manner and do not get burned along the way, 
• have all the answers ready when the questions about their AI models will arrive.

If you are interested in more details about the above please feel free to:

1. Watch our dedicated PRMIA webinar Balancing the Two Sides of AI: Benefits vs. Risks
2. Read our article on The emergence of Enterprise Model Risk Management
3. Sign up for our PRMIA Model Risk Management Virtual Training course. 
4. approach me :-D

This article by Peter Plochan, FRM

SAS North EMEA Principal Risk Management Advisor & PRMIA Risk Trainer you can follow him in linkedin