Coexisting in a crowded (clouded) landscape

September 9, 2022

Recently, I had the pleasure of attending the Gartner Data and Analytics conference. The sessions featured representatives from McDonald's, Bank of America, Samsung, Pepsi, and many others, who shared their methods for using AI/ML in the enterprise, some of their pain points, and their ongoing journeys to become AI-first companies.

Two stats stood out above the rest of the show. First, Gartner is tracking more than 300 MLOps companies. Next, in a recent study, Gartner found that only 54 percent of models make it into production, practically unchanged from a similar study they did in 2020.

Yet, the company representatives I spoke to would regularly cite even lower numbers. Some put the figure of models that make it from experiment to production in the single digits. MLOps platforms are, ostensibly, designed to raise this statistic.

I sense a paradox. While there's obviously a problem within enterprises to bring models into production, few of the platforms today appear to solve that problem.

We've found this is for a few reasons:

  • Business users must be included. Often, departments outside of the data science group assess models for various factors, such as risk or bias. Timely review and understanding of the model and what it will do, as well as some analysis for unintended consequences, are crucial for trusting humans to put AI into production.
  • Not-another-platform syndrome. Enterprises know that there's no panacea for putting models into production. And large enterprises tend to have too many tools in use today. Okta put the average number of applications per enterprise above 80. This is in addition to the number of data-science-specific tools within a company!
  • True coexistence is a lot harder to find than companies claim. Several of the largest players in the market have an incentive not to make it easy to work with their competitors. Others have transitioned to MLOps after starting off as data-centric companies, seeking to increase revenues however they can. Neither of them is natively designed to coexist in a complex landscape.

Here at Vianai, we benefit from senior leaders who understand how fragmented and complex an enterprise's IT landscape can become. This knowledge and further research went directly into the design of our applications and our platform.

We have sought to solve these problems in many ways. One is by showing models in production, with functionality that non-technical folks can use. Another is through interoperability built into our platform the start - we are completely modular and operate in any cloud provider. Finally, we have developed proprietary techniques for understanding risk in a model and in the underlying data.

All of this sits alongside additional functionality, such as optimization, which also helps shrink the size and cost of a model. You can find out more on our website: