The Framework for ML Governance

Chapter 4. Putting It All Together

We’ve established that companies are struggling to get value out of their AI and that a comprehensive ML governance implementation is how to achieve that value. While MLOps has become ubiquitous (even if only in name), taking it one step further to manage your MLOps with ML governance is the next stage of maturity for enterprise ML.

ML governance isn’t nearly as much of a buzzword as “AI” or “MLOps,” but it is by far the most important component for delivering value with ML. It is the final step to making ML a standard part of any organization. The first companies to implement standardized ML governance will have a once-in-a-lifetime opportunity to dominate ML in their business vertical.

Getting Value from Your ML with ML Governance

MLOps is the set of best practices and tools that allow you to deliver ML at scale. ML governance is how you manage those practices and tools, democratizing ML across your organization through nonfunctional requirements like observability, auditability, and security. As ML companies mature, neither of these are “features” or “nice-to-haves”—they are hard requirements that are critical to an ML strategy.

The value of a comprehensive governance implementation is inherent. Through effective management and controls, you unlock better, faster, and more secure ML. While some governance features may sound vague or high level, they’re components of software and ML alike. Governance drives:

• More accurate ML through observability and monitoring

• Shorter time to deployment through well-managed pipelines

• Secure ML through controlled processes and security tooling

• Less incurred risk through explainability and visibility

• Rapid improvement cycles

The best part about ML governance isn’t that you’ll deliver more value with ML. It’s that you’ll continue to deliver more value with ML at every step. The most essential part of governance is that it sets you up for long-standing, continuous improvement. Like Agile, governance is a framework for surfacing important information that enables iteration and improvement. Companies that implement governance won’t win the AI race because it helped them deploy a model. They’ll win because “governing” ML enables cyclic and organic growth and is essential in maintaining a competitive advantage in an era of rapid digital transformation.

ML governance is a framework for delivering value in many ways, but it’s primarily a framework for growth. While there’s no underestimating the importance of the ML models themselves, it’s the framework that surrounds them that enables actual value. Without governance, no matter how much an individual model improves, the ML of your organization will not.

How to Set Up an ML Governance Program

ML governance is not an individual responsibility. Part of the reason ML ventures have been so challenging is that data scientists have been somewhat siloed in their efforts. This is not their fault, and data scientists have always continued to do what they do best—train effective models. However, organizations need to recognize ML projects as the massively cross-functional initiatives that they are.

Algorithmia’s “2021 Enterprise Trends in Machine Learning” report found that a wide range of business units and roles are involved in setting ML priorities for organizations—from IT infrastructure and operations to data science to DevOps to product teams. To succeed with ML, organizations must consider all the decision makers who will be involved, regardless of department or role. If your ML model will touch almost every unit of your organization, it makes sense that you would want each of those units represented for alignment and support. Most notably, at 57% of organizations, IT infrastructure operations leaders are responsible for setting the priorities for AI/ML, whereas a CTO or similar head of innovation set the priorities for AI/ML at 39% of organizations. If you want to develop a successful ML strategy in your organization, then it will require alignment across a number of business units and domain experts in your organization. Successful AI absolutely cannot be driven by data science teams alone.

It might come as a shock that the head of data science sets the priorities for AI/ML at only 31% of organizations—but actually, this exemplifies just how important the operations side of a project is. Any organization will likely have very particular guidelines for software that will also affect ML.

To make things harder, a lack of domain knowledge can make it difficult to communicate across business units. IT and infrastructure may not understand AI, and few data scientists have extensive experience delivering compliant software. When these groups come together to build an ML governance strategy, it can be difficult to bridge the gap and find common ground for alignment. Porting existing software policies to ML deployments may not be a one-to-one mapping and will require experience in both areas.

Setting up an ML governance program involves education, coordination, and effective communication. Every group expected to contribute must be incentivized (and remain incentivized) to create a robust governance framework. An organization needs to foster a governance-forward culture with clear roles, responsibilities, and value-driven goals.

Involving business leaders such as the CTO/CIO can provide AI projects with the necessary sponsorship to bridge these gaps and generate a culture necessary for the success of your governance strategy. Bringing in other essential groups like IT is crucial in removing governance roadblocks before they happen. Something holistic like ML governance cannot feasibly be created in a vacuum without the input of multiple business units.

Aside from the obvious data science team, any successful AI project will need key stakeholders involved from across the business.

Infrastructure and Operations

These individuals should be responsible for helping integrate and scale new technology in your organization. They will also be well versed in the governance concerns of software and the requirements for your AI application. Since you will need to meet all of the infrastructure and operations (I&O) requirements (which could be extensive), it’s imperative to involve them as early as possible in the planning process. Meeting security or other governance requirements could have a significant impact on the scope of a project and the time it takes to deliver.

I&O is far less concerned with the performance of your model than they are about the technical implications and risk of introducing it into their production systems. I&O isn’t as concerned with the statistical performance of your model as they are with the operational and compliance implications of introducing it into production infrastructure. Having a time investment and relationship from these individuals is absolutely critical in ironing out the details and requirements they’ll inevitably need to impose on your AI solutions.

If you wait until development is almost done before finding out that your AI deployment will require role-based access controls and SSO integration, your project is likely to experience massive delays and incur unexpectedly high cost or risk. Involving these individuals during the planning phase of a project makes sure that groups are aligned and the effort can be estimated accordingly. Making them part of the process also helps foster the necessary relationships and ML governance culture required to keep individuals responsible, interested, and involved.

CTO and CIO

Although they are unlikely to be involved in the day-to-day, it’s obviously important to have executive sponsorship to ensure continued dedication and support for a project. It should come as no surprise that CTOs are involved with ML decision making at more than a third of organizations. The involvement of C-level executives almost always increases the chance of success for a project.

While it’s unlikely that someone at this level will be naive to the potential value proposition of AI, it might not be clear how important governance is in enabling that value. Getting C-level sponsorship for ML governance initiatives will come down to how effectively you can communicate the value of that governance. Common metrics for ML “success” include time saved or money saved. ML governance is a strategy to dramatically increase the ROI on ML initiatives through fine-grained controls and repeatability, and that is a strong motivator for value-driven businesses.

Head of Data Science

The leader of data science at a company is a must-have for setting the AI vision and keeping delivery on track. The responsibilities of this role may vary widely, but as the chief representative for data science across the organization it’s important they create and manage the unified vision for AI delivery. Lack of unification will often lead to duplication of effort. Without a centralized messaging and approach, organizations can end up with multiple, incomplete ML workflows instead of extensible, integration-first platforms that are potentially reusable across multiple use cases within your company. The head of data science should shoulder this responsibility and play a key role in fostering an ML governance culture to achieve success as an ML-driven organization.

Other Business Team Members

The persona of other stakeholders in an AI project can vary widely from company to company. This usually includes SMEs, product teams, marketing experts, and others. AI initiatives require nontechnical, business-oriented representation to stay focused on the primary goal of delivering value for the company and contextualizing the purpose of AI in the first place. Data science teams need to align with the business on key success metrics for the project to ensure they are not delivering AI for AI’s sake but to achieve a business-level goal.

How to Action on This Framework

The implementation of a framework for ML governance will depend on the state and maturity of your organization. Some organizations have robust, efficient pipelines but struggle to manage them, while others are still working to better automate the deployment stage of the lifecycle. Regardless of where your company is, here are five strategies for being successful in creating, driving, and delivering a holistic ML governance strategy:

Get C-level sponsorship.

The fact is that C-level backing will always increase the chances of success for a project. If directors and above aren’t convinced that they need ML governance (or ML), then it will be difficult to align this initiative to your team goals and get the investment you need from other key groups, like IT and ops. If your company wants to be an ML company, then key decision makers need to understand the long-term value and necessity of a comprehensive governance strategy for ML. The ability to bring ML governance into your organizational goals and all the way down to your objective key results (OKRs) will focus teams across your organization and guarantee commitment.

Bring in domain experts early.

IT, ops, and platform professionals are not waiting around for the next project. Most likely, they have a long backlog themselves and will need ample notification to assist on any initiatives. If you don’t bring domain experts in early, not only are you at risk of not having their time—but you may also completely miss key requirements for your ML deployments. If your company has specific security requirements for software, then you need to be aware of that ahead of time to estimate a project more accurately. Lack of effective notice and communication for key enabling teams is a common and avoidable cause of ML project delay.

Focus on transparency and communication.

It sounds obvious, but miscommunication can lead to large lapses in initiatives. Make sure action items have clear due dates and owners. This is part project management and part effective communication. You can discuss ML governance all you want, but until ownership for each action is clearly defined, progress will likely remain slow. Implementing a governance framework is about as cross-functional as it gets, and coordinating work across a number of teams requires excellent project management and communication skills.

Don’t reinvent the wheel.

Organizations will vary widely in their technical literacy and capacity. So, buy where you can and build where it makes sense. A common mistake data science groups make is doing everything themselves—from security to infrastructure. This is a surefire way to fail. Depending on where a data science team sits within an organization, they may not even have the direct support of software engineers or I&O professionals. If an enterprise solution solves 80% of your requirements and enables you to build the remaining 20%, then you want to heavily consider purchasing it. Top players in the ML governance space like Algorithmia are integration-first platforms. They take care of the boring, difficult pieces like security and monitoring so data science teams can focus on what they need to focus on—data science.

Remember the core tenets of ML governance.

ML governance is MLOps management, a control plane for your MLOps. The largest features it delivers are control and visibility, which in turn generate growth. In some cases, governance is a bit abstract, and it isn’t always as clear-cut as implementing a tangible feature like security. However, if you’re driving initiatives that improve the data science experience and allow you to deliver quickly, then you’re helping foster governance.

Conclusion and Next Steps

The framework for ML governance is a framework for managing your MLOps and getting value out of your ML. The next few years will see a stark contrast in companies that treat ML like a requirement, and those who treat it like a “nice-to-have.” In Gartner’s “Top 10 Data and Analytics Trends for 2021”, the number one trend will be smarter, more responsible, and scalable AI. MLOps has dominated the ML space over the last few years, and ML governance is beginning to build on that momentum and drive companies into the next phase of enterprise ML adoption.

MLOps may be a buzzword, but that’s because of the undeniable value of MLOps to drive real business impact. ML governance, the ability to control and manage MLOps, is the next big phase for companies as “ML” settles in as a permanent fixture of the businesses of 2021 and beyond. The advantage ML governance will give companies in the coming years cannot be overstated, and those who treat it as a “data science” concern rather than an enterprise concern will be left behind.

We need to start treating ML more like software and foster a culture of ML governance that allows for rapid, risk-free model development and deployment. ML is all about iterative cycles of improvement, and ML governance is about dramatically shortening that cycle so that your company can drive value faster than ever before.

We’ve outlined the components of a comprehensive governance strategy, but we also know that change at large companies doesn’t happen overnight. Building and implementing an ML governance strategy requires a lot of work and, more than anything else, cross-functional coordination. Here are four ways you can get started today:

Start the conversation.

You won’t be able to implement ML governance if no one knows about it. Even if it isn’t a sanctioned project, folks within your organization should be aware of the industry’s move to governance and what they need to do to stay relevant. Start conversations with key stakeholders now, so that when the time comes, they’re aware and ready to action on it.

Research end-to-end MLOps platforms.

Whether or not you have a large, capable engineering organization, the fastest way to develop the technical parts of ML governance is to buy them. The enterprise space for MLOps has exploded, but there are fewer integration-first platforms focusing on first-class governance.

Gather statistics.

Whenever you want to make changes within your organization, you’ll have to make a case for it. Start gathering important metrics from recent reports so you can accurately showcase the value of ML governance to key decision makers in your organization.

Implement ML governance in your everyday work.

ML governance has a lot of technical components, but it also has a lot of components that just come from good software development culture. If you’re involved in ML projects on a daily basis, do what you can to encourage the implementation of different ML governance framework components. Making these a standard part of ML projects at your organization is the first step toward a robust and complete framework.

For your organization, ML governance may happen gradually or it may seem to happen overnight. But if you want to generate value with ML, it needs to happen. Otherwise, you won’t really be doing ML. The framework for ML governance is essentially the framework for valuable, reliable, and secure ML.