Micro Focus is now part of OpenText. Learn more >

You are here

You are here

Why you should prioritize governance of ML and AI

Diego Oppenheimer Founder and CEO, Algorithmia

This is a truly momentous time for machine learning in the enterprise, with investments soaring and a growing number of use cases that can create tangible business value.

Organizations are still struggling with important phases of the AI/ML lifecycle. One particular challenge stands out: governance. A lack of robust governance doesn't just limit the potential success of your AI/ML initiative; it could put your entire business in peril as well.

That was one of our major findings in Algorithmia's "2021 Enterprise Trends in Machine Learning" report. Some 58% of organizations said they struggle with governance, security, and auditability issues. That makes governance by far the top challenge that organizations currently face as they scale up their AI/ML initiatives and put more ML models than ever into production.

This is a significant issue that affects most organizations: 67% of enterprises surveyed for the report said their AI/ML initiatives must comply with multiple regulations, and just 8% indicated that they did not have to comply with any regulatory rules at all.

This will be a crucial year for you to take meaningful action to improve AI/ML governance. Here's why.

Why you need more AI/ML governance

Regulatory compliance on its own is a critical requirement, but governance is about far more than this single issue. It's an area with enormous implications for your brand and your bottom line. Failing to prioritize governance doesn't just put your AI/ML initiatives at risk; it could jeopardize your entire business.

Our independently conducted report—now in its third year—includes input from more than 400 enterprise leaders and practitioners who are involved with their organization's AI/ML strategy. They span a diverse range of roles that reflect the increasingly cross-disciplinary nature of enterprise ML, from CIOs to IT infrastructure and operations (I&O) professionals, to data scientists and DevOps and nontechnical senior leaders.

This group spoke loud and clear: A lack of good governance is a major problem. And it needs to be addressed now, since so many companies moved AI/ML up on their priority lists in 2020 and expect their investments to begin paying off in 2021. Not emphasizing AI/ML governance introduces harmful risks to the overall health of the business, including a lack of customer trust and lasting damage to your brand, among other huge problems.

What is AI/ML governance?

We define AI/ML governance as the overall process for how an organization controls access, implements policy, and tracks activity for ML models. This certainly includes regulatory compliance and audit risk—and can ensure the former while minimizing the latter—but it’s about a lot more than that. Good governance essentially forms the bedrock for minimizing risk while maximizing ROI.

There are many variables that can affect ML model results, for example. Organizations with strong governance in place understand those variables deeply and ensure they have the 360-degree visibility and granular control needed to effectively monitor and operate their production models. In effect, they can integrate AI/ML governance policies with their broader IT policies for greater efficiencies and reduced risks.

Companies with strong governance in place can document and version models, tracking both the inputs and outputs of those models—and as a result can act fast if the variables that can affect model results begin to flash warning signs of potential problems.

The problems with ML-based automation

Introducing automation without clear visibility and oversight, on the other hand, is dangerous. Operating ML models without good governance in place allows flawed processes to produce unwanted results—often quickly and repeatedly.

Common problems include issues such as model drift or degrading application performance. Left unmonitored and unmitigated, such issues can not only negatively impact ML results, but can ultimately erode customer trust as well and the overall health of the business. A lack of governance means you might not even realize there are major problems until the damage is already done.

For this reason, an AI/ML initiative with poor governance in place could be as damaging to an organization as no AI/ML strategy at all. Both roads lead to negative impacts to the top and bottom lines and the business's ability to compete in a constantly changing market. An ML initiative of any size requires governance. Yet too many organizations are not recognizing this.

Why organizations struggle with governance 

There are several fundamental reasons why so many organizations struggle with governance.

First, they're unclear on best practices. This is still early days for ML governance, and many companies simply lack a road map for effectively implementing it in their organizations. They need a reliable source of expertise and prescriptive advice.

Second, it's also early days for the AI/ML regulatory landscape. The rules are evolving and are often difficult to understand in an actionable manner—which means that companies must invest significant resources in compliance unless they have help. Those that don't have ample internal resources can't keep up, and risk losing their competitive edge in the process.

Third, existing governance solutions are often heavily manual and incomplete. Even organizations that have taken steps to implement governance in their AI/ML initiatives tend to do so in ad hoc fashion with a mix of disparate tools and manual processes. This is exacerbated when done piecemeal by siloed teams, each forging its own mini-solution that addresses issues that only that team cares about. This approach becomes a burden to maintain and often creates critical gaps or blind spots in coverage.

Fourth, ML on the whole does not easily integrate with existing enterprise IT policies. This becomes a particular problem when ML is treated as a one-off project, and when there is a lack of collaboration across IT and business units. This makes it extremely challenging to integrate ML solutions into standardized IT processes and business units.

What to do about it

You must tackle each of these issues head on if you expect to ensure strong governance that unleashes the potential ROI of your AI/ML initiatives, all the while protecting the business from the dangerous risks created by a lack of visibility and control over operational concerns like model performance.

You must get up to speed on best practices and stay current on the regulatory landscape, for example, and enlist outside expertise as needed. You need a complete solution for ML operations (MLOps) that replaces patchwork approaches to governance with an organization-wide approach. And you need to ensure that all stakeholders have a seat at the AI/ML table and are effectively communicating and collaborating.

In fact, a lack of organizational alignment is a root cause of governance challenges (or a lack of governance altogether). This is one reason it was so important for us to include a mix of different roles in our report survey: They represent the cross-disciplinary nature of AI/ML.

The different stakeholders in AI/ML strategy and operations care about different issues. For example, data science teams care about getting models into production as quickly as possible, while I&O and DevOps teams care about integrating new technologies with existing systems without introducing security risks or operational burdens. Risk managers care directly about controls and governance, and business teams care about getting revenue-generating products (that rely on your ML models on the back end) to market to ensure top-line growth.

All of these are valid concerns, and they need to be properly aligned. You can have speed and security, for instance, but only when all stakeholders are collaborating and communicating effectively. This is critical to good governance: Without proper alignment, you risk significant gaps and conflicts between the concerns of different disciplines because people only care about the issues most directly important to them. This has to change.

Make governance a top priority

There's a good chance that your organization is already struggling with AI/ML governance and doesn't yet know it. And even if you've begun to implement governance tools and processes, you might have coverage gaps that could pose major risks. Moreover, you’re likely overestimating your organizational alignment and maturity.

This year, make governance a part of your AI/ML strategy that's every bit as important as budget or hiring. Your brand and bottom line depend on it.

Keep learning

Read more articles about: Enterprise ITIT Ops