Building a data science team is as tough as assembling any collection of people who are meant to reach a common business goal. However, it has some unique challenges, too, especially with data science a newly established function inside many businesses.
In NewVantage Partners’ latest survey on artificial intelligence and big data, only 38% of respondents from the Fortune 1000 said they have managed to create a data-driven organization. Even fewer — 27% — reported success at building a data culture within their firms.
Becoming data-driven requires effort, and one key piece is how the data science organization is structured — hived off from the rest of the company, a central node that touches many points of a network, or something in-between.
“If data science is just off in a silo doing data science and nobody wants it, you’re not going to be effective,” said Dan Gremmell, the vice president of data science at Plated, an American ingredient-and-recipe meal kit service owned by Albertson’s. “You’re going to fail.”
Fortunately, there are some proven approaches, and Gremmell addressed them at the DATAx New York event last November.
How best to organize a data science team, especially starting from scratch? One approach is functional, said Gremmell. “I see a lot of really large companies use this functional data-science layout – every group [like marketing] has its own hierarchy of data scientists and data leaders. What happens, of course, is that all those groups become isolated,” and a lot of waste is built across the organization. While very large organizations may be able to work this way, this setup won’t work for a majority of companies.
Option two is to “centralize everybody,” said Gremmell. “You take all your data scientists, data analysts, and data engineers and you hoard them, you keep them isolated.” If an organization is trying to conduct a lot of research, advance a technology, or do other things that are highly experimental, this structure can work well, he said.
However, “If you’re trying to solve everyday business problems and you want people to get behind you, this is not the best structure — the [data science] team never really learns the business. … They never really understand the perspective of a business partner.”
Option three is called “embedding.” In this structure, data science personnel are allocated to cross-functional teams that work in a specific focus or vector – like growth, for example, or product. “This is a really nice strategy designed to help people understand the organization and get involved with the business and understand the problems of the business,” Gremmell said.
The only catch is that sometimes an employee can get isolated – an individual data analyst in a cluster, for example, may begin to feel “cut off from the mother ship,” said Gremmell. “They’re really focused on the business, almost too much." As a result, embedding works best in really small startups that have a collaborative culture and don’t have a lot of staff, Gremmell said.
As a company starts to grow, there is another option — “structured embedding.” In this structure, the company builds clusters of focus throughout the enterprise – a growth cluster, a product cluster, an operations cluster.
Each cluster has an end-to-end data science process that works directly with the business but, Gremmell explained, has “a support system built throughout it.” A cluster includes data analysts, data scientists, and data engineers, and the data analysts and scientists know who their data engineer is. “They build structures and they’re able to self-allocate work,” he said. At the same time, they are directly connected to the business through cross-functional teams. “They work understanding business problems and they are able to collaborate effectively together to help solve problems. It’s kind of an ‘in-it-together’ mentality.”
Structured embedding is used by Airbnb, Gremmell noted. A data scientist there has described it this way: “We still follow the centralized model, in that we have a singular data science team where our careers unfold, but we have broken this into sub-teams that partner more directly with engineers, designers, product managers, marketers, and others.”
At the top of any data science structure, of course, there needs to be a leader, and that leader needs to be on the executive team, if the management expects the culture to become data-driven, said Gremmell.
“You’re never going to get someone who isn’t fluent [in data] and [in] talking data to help push data-related absorption through the organization,” he said. At Plated, Gremmell is in charge of data science and leads the strategy planning process. “I like leading processes and I’m good at building processes,” he said. “But it’s also because we can help make sure we infuse data science into the strategy and cascade it through the organization.”