A survey of 503 developers with machine learning experience suggests a large percentage of artificial intelligence (AI) projects are being abandoned before completion.

Conducted by Civo, a provider of cloud computing services, the survey finds more than half (53%) of respondents have abandoned between 1-25% of projects, with an additional 24% having left between 26-50% of projects. Only 11% of developers said they have never abandoned a project, the survey finds.

Given the total expense of building AI models, that rate of abandonment shows there is a lot of costly experimentation, notes Josh Mesout, chief innovation officer at Civo.

The survey also notes that nearly half (48%) of the developers surveyed believe ML projects are excessively time-consuming. For example, the survey found that 34% of those surveyed personally spend 0-10 hours configuring or setting up ML projects each month, with a further 24% spending 11-20 hours.

In fact, organizations that have a strong pedigree using DevOps practices to deploy applications are typically more efficient, adds Mesout. Rather than having walled gardens between data science and application development teams, organizations need to find ways to meld machine learning operations (MLOps) with DevOps workflows, he notes. “Organizations that have walled gardens are less agile,” says Mesout.

On the plus side, the survey finds that organizations that use open source tools are more efficient. Nearly three quarters of respondents (72%) credited open source software with reducing the time from implementation to insight. Nearly half (46%) saved 0-10 hours, while a quarter (25% saved 21 to 30 hours, the survey finds.

Given the level of urgency being assigned to AI initiatives, it’s only a matter of time before the productivity of the teams assigned to these projects becomes a more pressing issue. Many of the data scientists assigned to these projects, unfortunately, don’t have a lot of business expertise, so it’s not surprising that many AI projects are abandoned.

At the same time, there is often a cultural divide between data scientists and DevOps teams that build and deploy applications. Most of those applications are updated frequently. In contrast, the cadence for building and deploying AI models is understandably much slower.

In the short term, most organizations are still determining how best to operationalize AI in a way that drives a return on investment (ROI). However, as more organizations realize the potential AI has to transform operations, the pressure to deliver on the promise of AI is clearly rising. Business executives are clearly concerned their organization might be left behind should rivals make faster headway.

One way or another, it’s only a matter of time before AI models are embedded in nearly every application. The issue, as always, is to determine which use cases to focus on first, given the level of resources and expertise available. After all, applying AI to the same processes as everyone else is doesn’t create a sustainable competitive edge. The challenge is figuring out what use cases are simply the next set of table stakes to compete, versus something that fundamentally changes the game.