Micro Focus is now part of OpenText. Learn more >

You are here

You are here

Why most machine learning projects stumble

Neal Gilmore IT Transformation Director, Insights & Data, Capgemini Canada

Despite widespread interest in machine learning (ML), relatively few projects leave the proof-of-concept phase and enter production. In fact, in a 2020 report, Capgemini found that roughly 85% of all ML projects grind to a halt across Capgemini's client organizations—despite successful preliminary models and ample support from executive leaders.

Further, the study found, only half of the world’s leading AI-powered enterprises successfully roll out artificial intelligence projects, including ML models, and this number drops substantially among organizations without dedicated ML teams. 

In recent years, AI solutions have attracted the interest of executive leadership across industries. Machine-learning models, perhaps the leading subset of AI, have particularly interested enterprises racing to digitize in the modern market because of their ability to automatically "learn" and update.

These self-sufficient ML models have powered countless products and services, from performing fraud identification to creating personalized playlists, and promise to allow organizations to operate more efficiently, reduce costs, and enhance consumer engagement, among other benefits.

The disconnect between ML proof-of-concept (POC) projects and production has left enterprises to question why the majority of these projects never launch—and, more importantly, how organizations can get ML models up and running at scale.

Here are three leading causes of failed ML projects, followed by how your team can quickly address them.

Why most ML projects fail

1. Internal miscommunication and misalignment

Although executive leaders have a vested interest in deploying ML models, their excitement and support is often overshadowed by miscommunication or misalignment across the various teams. Traditionally, management has misunderstood a number of key aspects throughout ML development and deployment. They are not only unaware of applicable use cases, but also struggle to understand that creating preliminary POCs does not guarantee a finished product.

Moreover, leaders also do not typically understand the complex data governance, security, and access constraints required throughout development and deployment. Similarly, design teams are equally unaware of leaders' assumptions and thus fail to manage expectations and vocalize the time and resources they will need to launch successful ML models at scale.

2. Lack of proper hiring

Lack of adequate and appropriate hiring is another central cause of issues in scaling ML programs. While most enterprises have hired teams of data scientists to build preliminary ML models, seemingly few organizations have the data engineering and ML operations specialists needed to advance POCs throughout development and deployment.

Moreover, many enterprises do not include these team members in the initial design phase, causing friction.

3. Workflow friction

As with any large-scale project with many stakeholders, there is often significant workflow interferences throughout ML production. Many ML teams struggle with issues such as source code management, continuous integration, and data collection and management.

Often, enterprises do not establish an overarching strategy to address these concerns and thus do not use the appropriate ML platforms, frameworks, tools, or teams needed to see their projects to completion. Without this clear road map, design teams struggle to develop ML models, leading enterprises to waste time, money, and resources.

3 ways to stop the false ML starts

1. Establish a solid data and ML strategy

It's essential to form a coherent strategy to leverage ML and align executive leadership with their ML teams. By establishing a clear road map for ML projects that focus on strategic objectives, expected outcomes and insights, and data management, ML teams will be able to successfully design, build, and deploy ML models at scale.

This strategy will also help ML teams overcome communication and alignment issues with executive leadership, providing leaders with clear action items and tangible use cases that offer business value.

2. Hire the right skills

To put these ML strategies into motion, enterprises must engage multidisciplinary teams that span engineering, data science, and AIOps, among other technical areas. However, assembling this ML dream team may admittedly be easier said than done.

As of last year, 82% of the world's leading AI-powered enterprises noted a high demand for employees with ML skills, and almost 90% of organizations faced an inadequate supply of ML-skilled employees. As this tech talent crunch is not expected to abate anytime soon, enterprises should rapidly develop hiring and upskilling strategies to overcome the competitive job market and enlist the right roster of skilled team members.

3. Identify and use proper platforms, frameworks, and integration tools

With numerous steps between initial development and final deployment, it is essential to use the right ML platform, framework, and data integration tools throughout the design process. By identifying these platforms and tools early on, ML teams can fast-track the design process and use features such as prebuilt algorithms, model monitoring, and hardware and software stacks.

Moreover, by establishing a portfolio of platforms and tools for all team members to use, ML teams can reduce workflow friction and align their various specialists throughout the production cycle.

The future is bright, but ML models remain in the dark

Although enterprises have increasingly turned to AI solutions, the majority continue to face considerable hurdles developing and deploying ML models at scale. With a mere 15% success rate for ML projects, ML teams across enterprises are grappling with misalignment, miscommunication, staffing, and technological challenges.

By holistically addressing these issues and revamping ML production, enterprises can finally deploy ML models at scale and claim a stake in an increasingly competitive market.

Keep learning

Read more articles about: Enterprise ITData Management