Defining the sandbox: Governing artificial intelligence
August 4, 2020
At ATB, we hold the belief that AI may be the biggest and most disruptive technology advancement we see in our lifetimes. The power and reach of machine learning is rapidly expanding into many areas of ATB, and into our everyday lives. The long-term sustainability of any enterprise is predicated on trust and for any data-driven organization, the respectful and ethical treatment of data is core to building reliable AI solutions. Furthermore, an organization-wide governance process needs to be established, to ensure the AI solutions work as intended and the right controls for risks are in place. That starts with our AI Governance.
We know that in the early days of pioneering any new scientific or technical field, there are bound to be failures and learnings – even when we humans have the best intentions to start. Especially when the data we collect may have an inherent bias and can be further amplified by machine learning models. Look no further than news stories about AI tools which, when released into the wild, have been found out to have inherent biases. Now imagine the complexity and potential consequences of relying on AI (or rather on the data sets it uses and how it interprets them) when it comes to finances.
Responsible AI development and deployment is key to ensure we have the best product to serve our customers. But who defines what’s “responsible?” At ATB, we’re fully aware of the gravity of this task. Therefore as we move forward into building solutions using AI, we have also made the decision to create frameworks to serve as ‘guardrails’ to steer our work. As such, working with a dedicated and focused group of ATB team members, from our AI Guild, transformation risk, and guidance from legal for ethical matters around artificial intelligence, we’ve built an overarching AI Governance Framework.
ATB’s AI Governance Framework defines a set of rules that we as an organization must adhere to that can be repeatedly applied to different processes - guided by reason and logic.
AI governance:
Enables and ensures responsible use of AI without inhibiting the power of AI
Creates a robust business and operational framework to be utilized during the entire AI life cycle
Clearly defines accountability, responsibility, and oversight.
Ensures that AI activities and AI-related decisions and actions are consistent with the ATB values, and ethical, social, and legal responsibilities.
As artificial intelligence continues to evolve and mature, we undertook the challenge of creating a comprehensive framework to serve as a rulebook from scratch, this itself was a massive undertaking given the lack of industry expertise, best practices, and standards.
Here’s how we started building it:
First, we had to create a baseline of definitions and truths internally to move forward. We even had to define “AI” because how can we expect to govern something if we can’t have an agreement upon what “it” is? We took into consideration the aspects of our business that AI would touch: from data security, cybersecurity, fraud management, risk management, to model accuracy and explainability--the application of AI related technology is limitless.
Throughout the framework development process we dove into significant research across different aspects of AI and the impact these technologies can have, both positive and negative, across our operations. We tried to cover as much ground as we could, knowing that this will be one of the most important and frequently used frameworks across ATB.
Then, we needed to ensure that the framework protected everyone: us as an institution, and importantly, our customers. A rock solid foundation of principles for governing AI development would help define the sandbox our team can and will play in. This protects our organization and customers while allowing innovation to flourish at the same time.
How the Governance Framework will work in action:
As each team at ATB explores building AI solutions, the concepts will be submitted to a committee of practice, opening them up to ethical assessments, reviews of existing solutions to try and determine if the solution would be greenlit, end-user communications, and feedback loop management. This committee will be composed of leaders and various subject matter experts (SMEs) across the organization to give their input, filter requests, and determine the approach. The development of AI solution will then follow our operational life cycle, including design, build, test and deploy. Lots of independent validations and security checkups will be undertaken--these are necessary and critical steps to develop exceptional AI driven services. How we got here …
Last year, the Enterprise Data Science team at ATB, developed a Model Management Framework to guide ATB data scientists’ on their work developing data science models to help avoid potential issues and biases, and enabled independent validation of our machine learning models and helped standardize our development process.
It was through the creation of this that the team found the need for a top-tier governance methodology to act as an umbrella under which other frameworks would live (such as our Ethical AI Framework and Model Management Framework). So the Data Science team and the office of the Chief Transformation Officer joined forces to lead the initiative and create this overarching governance model. For now, much of the framework is based on assumptions and presumptions and is pulled heavily from our existing policies and frameworks as we learn more about the direction that we would need moving forward.
This is an ever-evolving science, and as such, the framework is a living document that will change with the times. The goal is to ultimately create AI solutions that our customers will trust. We want to make sure that we are developing AI technology responsibly, especially as a financial institution, and we’re working at all times to ensure that data is used for the right reasons, in the safest manner.
To learn more about our work in AI subscribe to the ATB Innovation Lab below.