Don’t try to predict the future when you scale your AI projects

Years ago, most companies that looked to AI to create products, services, or processes would have found a steep learning curve — a raft of new terms and new ways of thinking. But by now, in 2020, many have likely leaped beyond those early struggles and have completed proofs of concept, conducted initial pilots, and may have some projects in production. The next step is getting to scale, said Dataiku chief customer officer Kurt Muehmel during a session at Transform 2020.

Dataiku is a software company that helps other companies — usually enterprises — develop the internal ability to build their own AI products and services, hook those into their own business processes, and scale them. These days, it’s not uncommon for a Dataiku customer to be stuck at the scaling phase after building something with AI. “Maybe they’ve succeeded once, twice, maybe 10 times,” Muehmel said. “What we’re talking about at scale, though, is not 10 use cases deployed into production, but maybe 10,000.”

There are better and worse ways to scale AI, of course. What’s the biggest mistake enterprises make when they’re looking to scale their AI projects? “Sometimes what we see is that they try to predict the future,” he said. “They try to lock into the one future technology that [they think] is going to get them there.” Muehmel pointed to the relative rising and falling in popularity we’ve seen with Hadoop, Spark, and Kubernetes over the last six or so years — which is to say, these things have and will continue to be unpredictable.

“In a sense, that’s good,” he said. “Because it means that there’s innovation that’s going to continue, and new and better technology that’s going to come out.” The key is for organizations to essentially plan for the unpredictable, roll with the reality that there will be changes, and set themselves up to be able to swap those technologies in and out. That’s what Dataiku is designed to help companies do, Muehmel said — it provides an “insulating layer” between the people who are working on a project and the underlying compute layer.

Just as latching onto a given technology is usually a mistake, creating a broad and inclusive organization is the best way to scale, in Dataiku’s view. “The right way to do this at scale is all about bringing more people in. Bringing in not only the data scientists and machine learning engineers, but also the business analysts, the marketers, [and] the shop floor technicians — as consumers of those results, but importantly, as creators, and true participants in the AI development process, as well as its deployment, maintenance, [and] update process,” Muehmel said.

To get to that point — where a plurality of a company’s team members are using its AI tools in whatever way makes the most sense for them, be it via code or a visual interface — companies need to start by unsiloing their data. Ideally, that gives more people the same data to meet business challenges. Muehmel pointed to the example of a global pharmaceutical company that began its AI journey back in 2012. Dataiku worked with that company early on, mapping out which teams needed what data, unsiloing the data, and broadly scaling. “They’re talking about 3,000 different projects that they have running in parallel, hundreds of thousands of data sets that they’re working on, and hundreds and hundreds — nearing a thousand — individuals directly contributing to that process,” he said.

What comes after scale? Muehmel said it’s about embedding. “Ultimately, the goal is to get everything embedded — to embed the analytics, to embed the AI processes directly into the applications, to the dashboards, throughout the organization.” When that happens, he said, all those people using the data are “shielded” from all the operational parts and pieces, like cloud environments — they can access and work with the data they need without having to worry about where it’s running.