Deploy the AI pilot program into production
Several lessons can be learned from this success story. First, unstructured data must be prepared for AI models through intuitive collection formats, appropriate data pipelines and administrative records. “Unstructured data can only be leveraged once structured data is available and AI-enabled,” Cealey says. “You can’t just throw AI at a problem without doing some prep work.”
For many organizations, this may mean finding a partner to provide technical support to fine-tune the model to fit the business context. Traditional technology consulting approaches, where an external vendor leads a digital transformation plan over a long period of time, are not fit for purpose here, as AI advances too quickly and solutions need to be configured to fit the company's current business realities.
Forward Deployment Engineer (FDE) is a new partnership model better suited for the AI era. First popularized by Palantir, the FDE model connects product and engineering capabilities directly to the customer's operational environment. FDE works closely with customers in the field to understand the context behind technology initiatives before building solutions.
“We couldn't do what we do without FDE,” Cealey says. “They go out into the field to fine-tune their models and work with human annotation teams to generate ground truth datasets that can be used to validate or improve model performance in production.”
Next, the data needs to be understood within its own context, which requires careful tuning of the model to the use case. “For example, you can't expect an out-of-the-box computer vision model to improve inventory management by taking that open source model and applying it to whatever your unstructured data feed is,” Cealey says. “You have to export the data in the format you want and fine-tune it to achieve your goals. From there, you start to see high-performance models that can actually generate useful data insights.”
For the Hornets, Invisible used five foundational models that the team fine-tuned to context-specific data. This includes teaching the model to understand that it's “seeing” a basketball court rather than, say, a soccer field. To understand how the game of basketball differs from other sports that the model has knowledge about (such as how many players are on each team). You can also understand how to identify rules such as “out of scope”. After fine-tuning the model, we can now capture subtle and complex visual scenarios, including high-precision object detection, tracking, pose, and spatial mapping.
Finally, while the mix of AI technologies available to companies is changing every day, companies cannot avoid old-fashioned commercial metrics: clear goals. Without a clear business objective, an AI pilot program can easily become a never-ending, meandering research project that proves expensive in terms of compute, data costs, and staffing.
“The best engagement we've seen is when people know what they want,” Cealey said. “The worst thing is when people say, 'I want AI,' but have no direction. In these situations, they end up in an endless pursuit without a map.”
This content was created by Insights, the custom content division of MIT Technology Review. It was not written by the editorial staff of MIT Technology Review. Researched, designed, and written by human writers, editors, analysts, and illustrators. This includes survey creation and survey data collection. Any AI tools that may have been used were limited to secondary production processes that had passed thorough human review.
