For all of the hype that machine learning has generated over the past few years, there is one question on every budget approver’s mind: When will we see our return-on-investment?

Organizations will not see an immediate return, according to Valentino Zocca, vice president, data science at Citi and who moderated a panel at the AI in Finance Summit in Midtown Manhattan. “It’s a journey, and it takes time.”

A significant governing factor on the pace of ROI is how the firm organizes its machine-learning resources, added Kamalesh Rao, a senior data scientist at Société Générale and who participated on the panel.

Centralized and decentralized approaches each have their strengths and weaknesses.

The quickest way to see an ROI would be to embed one or two data scientists into existing data practices, according to Rao.

“It might not be the entire organization, but an individual siloed practice that can deliver results on a platform, which can scale quickly,” he said. “Maybe they already have some sort of delivery platform that does not use artificial intelligence or machine learning, but at least they have a data engineer, a data engineering practice, and a data engineering system.”

UBS Asset Management realized that increasing the proximity between its IT and data science teams was beneficial, agreed fellow panelist Zachary Glassman, a data scientist at UBS Asset Management.

“The more that your data science team can be thinking about how this will look in production and how they can engineer it,” he said. “It might not have been their traditional role, but it will enable things to get into production faster.”

However, large organization, such as banks, tend to take a “cathedral” approach, which uses pre-made tools, a central platform, and a consulting model that farms out talent.

“I think that is a harder way to go, but in the long run, it may be more efficient,” said Rao. “The first approach will get you a faster return, but in the long run, it may be more costly.”

Eventually, organizations should reach a point where machine-learning engineering becomes yet another box on the developer’s checklist like information security, according to Marsal Galvalda, head of machine learning, commerce platform at Square and panelist.

“Ideally, it is more a continuum and not a separate and siloed team,” he said. “It is all a matter of how you embed this organization with a machine-learning mindset. In practice, you will start with some sort of center-of-excellence or something, but the ideal endpoint for machine learning is to be just another tool that engineers use to develop services and parts.”