5 ways to improve AI/ML deployments

Before any AI/ML deployments, organisations need to marry their data science research efforts with their project management best practices.

In January 2019, Gartner released a survey where 37% of respondents said they were already using artificial intelligence (AI) in some capacity, but 54% of respondents reported skills shortages in their organisations that prevented them from moving forward with AI more aggressively.

This is not referring to data scientists, who continue to be in short demand and are aggressively being hired, rather to the fact that many organisations do not operational their AI efforts with IT project methodologies to ensure that projects meet their business goals.

"What we are seeing is a lot of data science teams that are working on many concurrent ML and AI initiatives, but fewer that have deployed the models into actual production applications," said Nathaniel Gates, CEO of Alegion, which specializes in training machine learning (ML) data.

Gates added that highly skilled data scientists may lack practical business experience in data preparation and project management. "These people are skilled at conceptualizing, building out, and testing AI and ML algorithms," he continued. "But we don't typically find much AI project expertise surrounding these data science teams. They often lack practical experience in data preparation for AI and machine learning."

Before any AI/ML deployments organizations need to marry their data science research efforts to their project management best practices.

How to improve deployments

Below are five ways for organizations to improve their AI deployments.

1. Develop an internal handoff process that transitions the initial data science algorithm and early data work into IT project management

The transition will ensure data quality and volume preparation and also places the project under a skilled project manager.

2. Use a combination of human data evaluation and machine learning automation with your data

Skilled individuals who know the data are invaluable in data evaluation for quality, but they may lack the bandwidth to review all of the data algorithms processes. Consequently, it is vital to employ data evaluation automation in the form of machine learning, which can be trained by human experts to assess data for quality.

3. Use an agile development methodology for your ML

AI projects should be conducted in manageable sprints that allow parts of the AI application to be planned, built, and tested quickly and iteratively.

"An agile development methodology with a focus on continuous and iterative improvement is imperative to successful ML development," said Gates.

4. Centralise your AI data and ML

"The most mature companies we have worked with have consolidated their ML training data requirements for AI into a centralised shared service that can be utilised across the multitude of data science projects within the enterprise," said Gates.

5. Use skilled project managers

AI and ML teams should be augmented with project managers who can enforce project management methodologies and best practices.

"AI and ML teams too often have no members who understand how to navigate the organisational waters outside of the team," said Gates. "We talk to data scientists all the time who know that they need lots of ML training data, fully understand that they can't produce it with the team they've got, and yet don't know anything about the organisation's budgeting, procurement, and project management processes."