Build an Automation-First Data Science Team
“Those who rule data will rule the world,” says Softbank CEO, Masayoshi Son. Forrester predicts that AI-driven companies will take $1.2 trillion from competitors by 2020. Everyone seems to think that data is going to be a vital part of the business world, but Gartner also estimates that 85% of AI projects won’t deliver for CIOs.
Dr. Greg Michaelson of DataRobot is brutally familiar with the world of AI-driven solutions. In his recent ODSC East talk, “Building an Automation First Data Science Team,” Greg delivered some tough love for organizations looking to utilize AI solutions long-term.
What’s the Struggle?
According to Dr. Michaelson, there are four main reasons for the failure of most AI initiatives:
- Organizations are still learning — The definition of AI isn’t consistent across the board. Many organizations may need help to identify realistic use cases.
- Organizations are more likely to experiment — Instead of committing to full deployment, organizations prefer to launch localized, temporary initiatives. It’s easier to find one individual excited about AI than a whole organization.
- Organizations become entrenched — Change is difficult for those without a roadmap. These organizations may miss the boat in favor of what Michaelson calls “nibbling at change.”
- Predictive models are intense to deploy — Most organizations aren’t equipped to do so.
Automation is threatening across the board for a variety of reasons. Teams without data scientists may not have the infrastructure or the understanding of use cases in the first place. Organizations with data science teams may see automation as introducing unnecessary risk.
That risk comes from this idea: if “anyone can build models,” who is responsible for the quality and risk management of those models? Dr. Michaelson points out that he could have even your grandmother building predictive models through DataRobot’s software within a week, but that doesn’t account for the knowledge required to mitigate the risk of real-life deployment.
Michaelson believes that the risks and issues are far outweighed by the benefit of the democratization of data. Building an automation first data science team is a smart move for the future of AI-driven organizations.
How to Create an Automation First Team
Navigation in the new space amid smarter tools is a critical “why” in creating automation first. Michaelson believes automating certain aspects of the pipeline is essential to reducing the backlog. While we may disagree on what can be safely automated, there’s no argument against the concept of automation.
Step 1 — Teach Everyone To Spot Opportunity
Data scientists make up too small of a pocket of your organization to be the only ones responsible for spotting opportunities. This isn’t a technical skill. It requires merely a small amount of training and knowledge of your organization inside and out.
Train everyone in your organization to do this. Not every idea will be a winner, but some of them will, and those potential winners are too valuable to pass up. Democratize this process and let your people do what they do best.
Step 2 — Create Your Roadmap and Prioritize Project Management
Build a list of things you could use machine learning. A roadmap — i.e., a list of AI solutions you want to build in the next two to three years — allows you to see where you’re leaving money on the table in potential automation missed cases. So how do you create this roadmap?
- You need an AI champion — i.e., the most excited individual in the organization.
- That champion educates teams to spot use cases (democratizing the process).
- Teams create big lists where every idea is entertained.
- Select the winners from the big list
- Set expected start and end dates
- Monitor progress closely
Step 3 — Automate What You Can
Save your hardest problems for your smartest people. Put your best data scientists on your most complex and delicate problems and automate everything else. These opportunities provide the foundation for your roadmap from the previous step.
The organization will always have problems that require bespoke, hand-coded solutions. However, Michaelson believes that the potential for automation is much higher than most organizations believe.
Step 4 — Over Communicate Everything
You cannot communicate enough about projects that go on in the space. There’s nothing more disheartening than working on a project and not getting any visibility. Everyone must be aware of the potential value of the automation projects because small wins create the momentum to keep building.
Automating Vanilla Solutions
Those “vanilla” projects are only vanilla in the building aspect. Many of them have the potential for huge wins on the business impact side, allowing data science teams to work on more complex tasks and freeing up the pipeline from common bottlenecks.
Building an automation first data science team requires some restructuring and a lot of communication. Still, as companies tackle the mountain of data and stem the tide of lost dollars, the process of automation could launch a brand new, AI-first mentality.
Read more data science articles on OpenDataScience.com, including tutorials and guides from beginner to advanced levels! Subscribe to our weekly newsletter here and receive the latest news every Thursday.
