Deloitte Consulting published a report today that suggests a golden age of AI is in the offing, assuming organizations can implement and maintain a consistent approach to machine learning operations (MLOps).
Citing market research conducted by AI-focused Cognilytica, the MLOps: Industrialized AI report from Deloitte notes that the market for MLOps platforms is forecast to generate annual revenues in excess of $4 billion by 2025.
Several startups are already focused on providing these platforms. Less clear, however, is the degree to which MLOps might become an extension of the DevOps platforms many organizations rely on today to build and deploy software.
At the crux of that debate is the way organizations currently build and deploy AI models. The average data science team is lucky if they can build and deploy two AI models a year. In the wake of the COVID-19 pandemic, however, organizations have accelerated their investments in AI as part of an effort to drive digital business transformations, Deloitte AI Institute executive director Beena Ammanath said. “This space is going to heat up in the next 18 months,” Ammanath said.
As the organizations look to operationalize those AI models, managing MLOps at scale becomes a significant issue, she added.
Naturally, venture capital firms and other investors are pouring funds into MLOps startups in the hopes that one might go public or, more likely, be acquired. The most likely candidate for making such an acquisition would be a provider of an existing IT management platform, as MLOps increasingly becomes part of mainstream IT operations.
MLOps, however, is different from artificial intelligence for IT operations (AIOps). The former refers to the process of building and deploying applications that are infused with AI models, while the latter refers to applying AI to automate the management of IT operations. MLOps, in effect, is borrowing many of the principles that DevOps teams pioneered to automate the building and deployment of applications to AI models.
Those MLOps processes not only extend to how AI models are built and deployed but also how they’re governed and eventually retired. One of the major issues with AI models is that results can drift over time as new data sources are made available or business conditions change beyond the scope of the initial model. That requires organizations to either update that AI model or replace it altogether with another AI model. In all cases, IT teams need to continuously test and validate recommendations AI models make to ensure they are consistent, relevant, and operating within ethical guidelines.
Coordinating that level of activity across teams of data scientists, developers, data engineers, quality assurance personnel, and IT staff requires a highly disciplined approach to MLOps, Ammanath said.
At this juncture, there is still a fair amount of healthy skepticism when it comes to applying AI. Well-defined processes tend to lend themselves better to automation using machine learning algorithms. The days when AI might one day replace the need for humans altogether is still far away. Instead, most AI applications merely augment the capabilities of humans.
The challenge organizations face now is that many of their existing processes are being rendered obsolete as organizations embrace digital business transformation. Applying AI models to business processes that are not as widely understood is a lot more challenging than automating a process that has been operating in the same manner for years, Ammanath noted.
Regardless of use case, there’s no putting the AI genie back in the proverbial bottle. Just about every application will be augmented to varying degrees by one or more AI models. The challenge and opportunity now are to provide the platforms that make it possible to not only build and deploy AI models at scale but also when necessary roll them back before any permanent damage is inflicted.
