Taking AI pilot programs into production
From this successful example, several lessons can be learned. First, unstructured data must be prepared for AI models through intuitive forms of collection, and the right data pipelines and management records. “You can only utilize unstructured data once your structured data is consumable and ready for AI,” says Cealey. “You cannot just throw AI at a problem without doing the prep work.”
For many organizations, this might mean they need to find partners that offer the technical support to fine-tune models to the context of the business. The traditional technology consulting approach, in which an external vendor leads a digital transformation plan over a lengthy timeframe, is not fit for purpose here as AI is moving too fast and solutions need to be configured to a company’s current business reality.
Forward-deployed engineers (FDEs) are an emerging partnership model better suited to the AI era. Initially popularized by Palantir, the FDE model connects product and engineering capabilities directly to the customer’s operational environment. FDEs work closely with customers on-site to understand the context behind a technology initiative before a solution is built.
“We couldn’t do what we do without our FDEs,” says Cealey. “They go out and fine-tune the models, working with our human annotation team to generate a ground truth dataset that can be used to validate or improve the performance of the model in production.”
Second, data needs to be understood within its own context, which requires models to be carefully calibrated to the use case. “You can’t assume that an out-of-the-box computer vision model is going to give you better inventory management, for example, by taking that open source model and applying it to whatever your unstructured data feeds are,” says Cealey. “You need to fine-tune it so it gives you the data exports in the format you want and helps your aims. That’s where you start to see high-performative models that can then actually generate useful data insights.”
For the Hornets, Invisible used five foundation models, which the team fine-tuned to context-specific data. This included teaching the models to understand that they were “looking at” a basketball court as opposed to, say, a football field; to understand how a game of basketball works differently from any other sport the model might have knowledge of (including how many players are on each team); and to understand how to spot rules like “out of bounds.” Once fine-tuned, the models were able to capture subtle and complex visual scenarios, including highly accurate object detection, tracking, postures, and spatial mapping.
Lastly, while the AI technology mix available to companies changes by the day, they cannot eschew old-fashioned commercial metrics: clear goals. Without clarity on the business purpose, AI pilot programs can easily turn into open-ended, meandering research projects that prove expensive in terms of compute, data costs, and staffing.
“The best engagements we have seen are when people know what they want,” Cealey observes. “The worst is when people say ‘we want AI’ but have no direction. In these situations, they are on an endless pursuit without a map.”
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
#unstructured #data #fuel #enterprise #success
