Bringing AI pilot programs into production
Many lessons can be learned from this successful example. First, unstructured data should be prepared for AI models through intuitive forms of collection and the right data pipelines and management records. “You can only use unstructured data if your structured data is consumable and ready for AI,” says Seely. “You can’t put AI into a problem without doing the prep work.”
For many organizations, this may mean that they have to find partners who can provide technical support to improve the model in a business context. The traditional technology consulting approach, in which an external vendor leads the digital transformation plan over a long time frame, is not fit for purpose here as AI is advancing very rapidly and solutions need to be configured according to the company’s current business reality.
The forward-deployed engineer (FDE) is an emerging partnership model that is better suited for 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 the technology initiative before building a solution.
“We couldn’t do what we do without our FDEs,” says Seely. “They go out and fine-tune the models, working with our human annotation team to produce a ground truth dataset that can be used to validate or improve model performance in production.”
Second, the data needs to be understood in its own context, which requires the model 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 your unstructured data feed,” Seely says. “You need to fine-tune it so that it can give you data exports in the format you want and help you achieve your goals. This is where you start to see high-performance models that can generate really useful data insights.”
For the hornets, Invisible used five foundation models, which the team tailored to context-specific data. This included teaching the models to understand that they were “looking” at a basketball court rather than a football field; To understand how the game of basketball works differently from any other game that the model may have knowledge of (how many players are on each team); And to understand how to detect 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, poses, and spatial mapping.
Finally, while the AI technology mix available to companies varies by the day, they can’t escape the old-fashioned commercial metrics: clear goals. Without clarity on the business objective, AI pilot programs can easily turn into open-ended, meandering research projects that prove costly in terms of computation, data costs, and staffing.
“We’ve seen the best engagement when people know what they want,” says Seely. “The worst is when people say ‘we want AI’ but they have no direction. In these situations, they are on an endless quest without a map.”
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by the editorial staff of MIT Technology Review. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes writing surveys and collecting data for the surveys. The AI tools that may have been used were limited to secondary production processes that underwent thorough human review.
